Log — LLM Providers Wiki

Append-only history. Each entry starts with ## [YYYY-MM-DD] <op> | <title> where <op> is ingest, query, lint, or split, so grep "^## \[" log.md | tail -5 works.

[2026-09-04] ingest | The Uncensored AI (Medium/Rekhi, URL-only, T3)

Model: claude-opus-5 · effort: high. Routed here by the hub from Telegram — one of three sources arriving together, and the one that landed on this spoke's existing abliteration cluster. Medium was a hard 403 for WebFetch; fetched with the firecrawl skill per ../HUB.md → Edge handling.

uncensored-ai-rekhi is the first source here that looks downstream of the abliteration technique rather than at it. The spoke held the mechanism (arditi-refusal-direction), the tool (heretic), the volume and a publisher (readyart); nothing on who runs the results. This essay supplies two tiers — rented (3+ vendors, 80B+ models, $30–$200/month, ~1,000 customers, per Group-IB) and squatted (hundreds of exposed deployments with guardrails removed, per a 293-day SentinelOne/Censys sweep) — now paged as dark-llm-services. Both are providers by every functional test and train nothing, which is the same slot readyart opened from the free end.

Tiered T3 and the reasons are on the page. Pseudonymous author, no primary evidence beyond one anecdote, every figure quoted from a report this wiki has not read, and one attribution too loose to chase ("a study published in Nature Communications a few weeks ago"). Recorded as leads to primary reports, not as findings.

Two corrections written into the corpus rather than inherited. The essay uses "abliteration" for safety removal in general and quotes it as "targeted fine-tuning" — flatly contra heretic, which does the edit with no training at all. abliteration gains a three-way table separating weight editing / safety-breaking fine-tuning / prompt jailbreaking, because only the first two need open weights and the essay's most alarming evidence (97% jailbreak rate; GPT-5 broken in 24h) is from the third, against hosted models. And the popular account drops the cost — abliteration as a clean subtraction — where this corpus's only measured source says the trade is the whole story.

open-weight-models gains "Publication is irreversible, and that is a separate axis": every licence row governs serving by an identifiable party, none reaches a downloaded copy. New entity maxime-labonne (mlabonne) — two sources now, and he is both the popularizer and Heretic's weakest measured baseline. Deferred: Rekhi, Eric Hartford/Dolphin, DavidAU, Group-IB, Trend Micro, SentinelOne, Censys, Tenable, HiddenLayer, FAR.AI — one source each. Synthesis gains a downstream-half section under "Safety alignment is a build-time setting" and a new ../defensive-security-wiki adjacency. Pages: 1 source + 1 concept + 1 person + 3 refreshed = 6. Runner-up spokes: ../ai-governance-wiki (irreversibility + jailbreak research as policy objects) and ../defensive-security-wiki (the criminal-tooling half); noted as cross-spoke context per ../HUB.md, not split.

[2026-08-30] ingest | Claude Opus 5 System Card (research pass, URL-only, T1)

Model: claude-opus-5 · effort: high. Hunted, not pushed — the research pass for ../research-wiki's model-substrate edge needed a measurement of Opus 5's honesty, and the only source at the bar is Anthropic's own card. Landed here rather than in research-wiki on the Sonnet 5 card precedent: a system card's dominant substance is a model's capability and safety evaluation, which is this spoke's domain. New source summary claude-opus-5-system-card (Report, T1, url-only — the PDF is 194 pages / 16 MB and is not held in raw/, same as the Sonnet 5 card). Updated claude-opus-5 with the card's three load-bearing facts: not more capable overall than Fable 5 (primary-source confirmation of the near-frontier positioning the launch post only implied), the +11% accuracy / +6% hallucination split against Opus 4.8 on AA-Omniscience, and the same ASL-3 / CB-1 treatment as Opus 4.8. Also retired that page's "unknown context window" item, which had been contradicting the 1M figure recorded two paragraphs above it from anthropic-model-docs. The substrate reading — what the numbers mean for a wiki maintained by this model — is folded into ../research-wiki's synthesis, not duplicated here.

[2026-07-19] ingest | Cerebras pricing (hub-routed, URL-only, T2)

Second Cerebras source from Telegram, hub-routed here (runner-up last time; the pricing/market angle is this spoke's turf — "API access & pricing" — while the mechanism/company sit in ../llm-inference-wiki). Cerebras is not a model maker but a speed-differentiated inference provider reselling open models on wafer-scale HW; its pricing page publishes no per-token rate card and instead sells daily-token subscription buckets (Cerebras Code Pro $50/mo→24M tok/day, Max $200/mo→120M tok/day, both sold out; $5 free credit; Developer from $10; Enterprise = contact sales). Created source summary cerebras-pricing (WebPage, url-only). Folded a new "speed-priced, no rate card" axis into llm-api-pricing and a "second pricing shape — the speed provider" paragraph into synthesis. No duplicate org node — the provider/company cerebras-systems and the service cerebras-inference already live cross-wiki in ../llm-inference-wiki; linked as bridge nodes (added to the index footer), not duplicated. Cross-spoke note: the Cerebras Code subscription is a coding-agent product (../agentic-tooling-wiki) — recorded as context, not routed away. Volatile T2 snapshot (tiers/prices/"sold out" will drift). Runner-up: ../llm-inference-wiki.

[2026-06-05] ingest | Gemma 4 with Quantization-Aware Training (hub-routed)

Hub-routed Telegram drop. Clear match (existing gemma-4); runner-up ../llm-inference-wiki (quantization-as-mechanism) declined — that spoke owns sampling/KV-cache/batching, not quantization, and the source's angle is deployable footprint in the open-weight market. Ingested:

• New source gemma-4-qat (source, url-only, BlogPosting): QAT checkpoints for E2B/E4B/26B-MoE; Q4_0 4-bit + a targeted 2-bit mobile scheme (mixed precision, channel-wise, static activation scaling); E2B text < 1 GB; QAT claimed above PTQ quality; GGUF/compressed-tensor on HF; broad local runtime list (llama.cpp/Ollama/LM Studio/MLX/vLLM/SGLang/Transformers.js/LiteRT-LM/Unsloth).
• New concept quantization (DefinedTerm) — int4/2-bit footprint lever, QAT vs PTQ; scoped to the market/deployability angle, with the deeper inference-mechanics depth flagged as ../llm-inference-wiki's (llm-inference) territory (bridge, no dup).
• Updated gemma-4 (QAT bullet + expanded deployment list; also tidied a stray </content> tag).
• Updated synthesis point 2: quantization is the second footprint lever after sparse MoE — footprint now contested by training technique, not just architecture.
• Updated index (added quantization + the source). Volatile caveat: the "<1GB" / "higher-than-PTQ quality" figures are vendor claims, undated vs neutral benchmarks — snapshots. Site rebuilt + verified.

[2026-06-01] split | llm-providers-wiki created from _inbox cluster llm-providers (4 sources)

The cluster was seeded by the parked deepseek-api-docs stub; on user instruction ("seek for content about llm-providers" → option (a)), the router web-searched the model/provider landscape and curated 3 strong sources to ingest, taking the cluster to 4 and triggering this spin-out.

Scaffolded from CLAUDE.template.md; domain = the LLM model & provider landscape. Ingested 4 sources (all URL-only): deepseek-api-docs (TechArticle), open-source-llms-2026 (HF, BlogPosting), llm-api-pricing-comparison (CloudZero, Article), llm-leaderboard-stats (llm-stats.com, Dataset). Created 4 concept pages (llm-provider, open-weight-models, llm-api-pricing, llm-benchmarks) + 1 provider page (deepseek). Synthesis thesis: collapsing price floor (open weights + DeepSeek) vs. still-premium proprietary frontier; "best model" is multi-axis. Cross-wiki bridges: anthropic/claude-opus-4-8 (../research-wiki), llm-inference/kv-cache (../llm-inference-wiki). Deleted the parked _inbox deepseek record. Caveat baked into the schema: pricing/rankings are volatile — pages are dated snapshots. Spoke count 8 → 9.

Note on sourcing: 3 of the 4 founding sources were router-found (web search), not human-dropped — a departure from the usual human-curated flow, done at explicit user request.

[2026-06-04] ingest | Introducing Gemma 4 12B (Google blog)

Hub-routed (Telegram). New source summary gemma-4-12b-announcement (BlogPosting, url-only). Gemma 4 12B: a dense, mid-size, encoder-free multimodal open-weight model — vision via a single matrix-mul embedding, native audio projected straight into the token space, no separate encoders. Benchmarks "nearing" the 26B-A4B at <½ the memory; runs on 16GB VRAM/unified memory; Apache-2.0; HF/Kaggle; llama.cpp/MLX/vLLM/Transformers. New Thing pages: gemma-4 (SoftwareApplication — family: E4B / 12B / 26B-A4B) and google (Organization — first provider page beyond deepseek; dual-track Gemini + Gemma). Updated open-weight-models (added a multimodality at the local tier trend + footprint-as-battleground) and synthesis (modality/memory as a new axis; "dual-track labs" recurring read). Index updated with a new SoftwareApplication group. Gemma 4 was previously only name-dropped (the 26B-A4B) in open-source-llms-2026; this gives the family a real page + the new 12B variant. Snapshot caveat: benchmark/footprint claims are the vendor's own.

[2026-06-04] lint | full pass (13 pages)

Triggered after the Gemma 4 ingest. Structure green: no dangling wikilinks (all 12 local slugs resolve; the 4 non-local targets — anthropic, claude-opus-4-8, llm-inference, kv-cache — are the declared cross-wiki bridges); no orphans (every page linked ≥3×); index catalogs all pages + synthesis; @types correct (providers=Organization, models=SoftwareApplication, concepts=DefinedTerm, sources=CreativeWork subtypes) — none need narrowing.

Fixed — missing cross-links (created by the new google/gemma-4 pages): linked previously plain-text mentions → google in llm-provider, llm-api-pricing, llm-api-pricing-comparison; gemma-4+google in open-source-llms-2026. Bumped those 4 pages' updated: to 2026-06-04.

Flagged, not auto-fixed (need a source / judgment):

• Opus version drift: pricing pages list Claude Opus 4.7 ($5/$25) while benchmarks/leaderboard/ provider pages reference claude-opus-4-8 as current #2. Both are dated mid-2026 snapshots from different sources (not a contradiction), but Opus 4.8 pricing is uncaptured — refresh when a pricing source for 4.8 arrives. Don't invent a number.
• Gemini snapshot lag: llm-api-pricing-comparison quotes CloudZero's Gemini 2.5 Flash $0.15/$0.60 while llm-api-pricing cites Gemini 3.1 Pro — faithful to each source, but the 2.5 figure is stale; flag on next pricing refresh.
• Thin anchor: google (1.1KB) is the lightest page — a deliberate brief org anchor; optional enrichment (name the Gemini model line + its llm-benchmarks placement). Not an error.

[2026-06-05] ingest | Google AI updates — May 2026 (hub-routed)

Hub-routed Telegram drop (broad Google "what we shipped in May" roundup). Routed here over runners-up ../search-marketing-wiki (Universal Cart / agentic-commerce, AI-search) and ../agentic-tooling-wiki (Android Halo / agent surfaces) because the source's dominant in-scope substance is model releases. Ingested the model/provider signal; explicitly scoped out the consumer-product/hardware/search items (noted as cross-spoke context in the source page, not parked separately — facets of one roundup).

• New source summary google-ai-updates-may-2026 (source, url-only, BlogPosting).
• New Thing page gemini (SoftwareApplication) — the closed-weight frontier family that the wiki kept referencing but had no page for (filled the gap the 2026-06-05-earlier lint flagged on google). Anchors Gemini 3.5 ("frontier intelligence for agents and coding"), Gemini Omni (multimodal video gen), Gemini for Science.
• Updated google (Gemini section + dual-track-converges-on-agents/coding point; linked gemini).
• Updated synthesis "Dual-track labs" recurring read: open/closed split is licensing+footprint, not target workload — both Gemini and Gemma now pitched at agents-and-coding.
• Updated index (added gemini model + the source). Volatile caveat: Gemini 3.5/Omni facts are marketing-roundup positioning, undated benchmarks — treated as snapshots. No pricing given. Site rebuilt + verified.

[2026-06-09] ingest | +3 major providers (OpenAI, Mistral AI, Llama) — all-spokes cron test

Filled the most glaring landscape gaps — the frontier incumbent and the non-Chinese open-weight pole: openai (Organization, src — proprietary GPT/o-series; the de-facto OpenAI-compatible API standard; MS partnership), mistral-ai (Organization, src — Europe's leading lab; Apache-2.0 Mixtral MoE + proprietary API), llama (SoftwareApplication, src — Meta's open-weight family that catalyzed the wave; open-weight but non-OSI custom license). Sharpens the licensing-spectrum thread (Llama vs Apache-2.0 gemma-4/mistral-ai). Wikipedia-sourced background; pricing/rankings remain dated snapshots. 16 → 19 pages.

[2026-06-10] ingest | Qwen + xAI/Grok + Amazon Bedrock — all-spokes pass

Three new pages filling named-but-unpaged gaps. qwen (Org/SoftwareApplication, url, Wikipedia) — Alibaba's leading open-first family the thesis kept citing: 0.6–32B dense + MoE (30B-A3B), reasoning/VL/audio/Coder/omni, mostly Apache-2.0; 200k+ HF derivatives, 234M app users; the strongest open-weight challenger to deepseek. xai-grok (Org/SoftwareApplication, url, Wikipedia) — xAI's proprietary frontier model; 2M-token context (the "xAI leads context" claim), and an open→closed arc (Grok-1 Apache-2.0 → all later closed), the inverse of llama/qwen → sharpens the "open is a spectrum/strategy-over-time" thread. amazon-bedrock (Org/SoftwareApplication, url, AWS) — the wiki's first cloud reseller: one-API aggregator of Claude/Llama/Mistral/Cohere/Nova (not a maker); the demand-side mirror where model choice is a config param and differentiation moves to routing/caching/data. Folded into synthesis (new 2026-06-10 section) + index (Organization rows). All vendor/encyclopedic snapshots — standing volatility caveat. No contradictions. 19 → 22 pages.

[2026-06-11] ingest | Claude API — Refusals and Fallback (hub-routed)

Hub-routed Telegram drop (platform.claude.com). Classified against wikis.md: dominant substance is the Anthropic provider API surface + model routing + billing, so routed here over the runner-up ../agentic-tooling-wiki (the SDK-middleware / sub-agent-fallback angle — noted as cross-spoke adjacency, not split). Not ../research-wiki (that's the model-substrate capability/cost bridge, not API mechanics) and not ../ai-governance-wiki (this is API-level classifier behavior, not policy/GRC). Ingested:

• New source claude-refusals-and-fallback (source, url-only, TechArticle): stop_reason: "refusal" as an HTTP 200; stop_details.category (cyber/bio/frontier_llm/reasoning_extraction); three fallback paths (server-side fallbacks beta, SDK BetaRefusalFallbackMiddleware, manual); fable-5 → opus-4.8 chain; per-attempt billing via usage.iterations[]; fallback-credit beta; sticky routing; batch behavior; "refusals are invisible to error-rate monitoring" pitfall.
• Updated synthesis recurring-read "engineering sets real cost": model routing is now a first-class API primitive (server-side fallback + fallback-credit + sticky routing), not just app glue — hardening the cost-is-engineering thread into the provider's own surface.
• Updated index (added the source under TechArticle/sources).
• Linked bridge nodes anthropic & claude-opus-4-8 cross-wiki (no dup).

[2026-06-12] ingest | Artificial Analysis — independent LLM benchmark platform

All-spokes daily expansion. Added artificial-analysis (@type WebPage) — the neutral, reproducible, methodology-disclosed benchmark the open question explicitly wanted (a rigor step up from the flagged-as- biased llm-leaderboard-stats/llm-api-pricing-comparison). Captured the four-axis frame (quality/ speed/price/context), the Intelligence Index v4.0 composite (GPQA Diamond, HLE, τ²-Bench, Terminal-Bench, SciCode, IFBench, …), the blended price at 7:2:1 cache:input:output, and live TTFT — incl. that it benchmarks API providers (the host/reseller layer, amazon-bedrock) not just labs. Resolves the "vendor/SEO bias — want a neutral benchmark" open question (residual caveat: disclosed editorial weighting + volatile snapshot). Wired to llm-benchmarks / llm-api-pricing; synthesis + index updated. 1 new page.

[2026-06-13] quality | maintenance cycle (tier + freshness backfill)

Quality Cycle (not expansion — spoke grown 06-12, never pad). Backfilled tier: + freshness: volatile on all 15 source pages: T1×2 (official Anthropic/DeepSeek docs), T2×8 (Artificial Analysis + Wikipedia entities + HF/llm-stats), T3×5 (Google/AWS/CloudZero vendor pages). Operationalizes the volatile-snapshot staleness tracking the spoke most needs (pricing/leaderboards churn weekly). Scorecard → hub quality-log.md. 0 new pages. No synthesis/index change (frontmatter-only).

[2026-06-15] ingest | Cohere / North Mini Code — thenewstack.io (honest stub)

Article body JS-gated; fetch failed. From headline + URL: Cohere is pivoting from enterprise sovereign-AI to developer-targeted coding models with "North Mini Code." T4 (The New Stack trade press). Created cohere (Organization, first Cohere page) + cohere-north-mini-code (TechArticle stub). Entity discovery: Cohere is a new org node (no match in entity-index), but entity pages for a new org from a T4 trade-press stub without confirmed founding facts felt over-reaching — noted as a gap; a T1 primary source would warrant a full entity page. Synthesis updated with sovereign-AI niche. Runner-up: none.

[2026-06-14] dedup | Google — made canonical here (merged the duplicate agentic-tooling node)

The entity-index fuzzy audit flagged Google paged twice (agentic-tooling + here). Per ENTITIES.md's canonical-node rule, made this the one Google node (provider landscape is Google's core identity; richer + url-sourced), folding in the agent-platform-vendor facet (ADK, agentskills adoption) with cross-links to adk/agentskills-spec, + aka: aliases. The agentic-tooling copy was deleted; its google links now resolve cross-wiki here.

[2026-06-15] lint | North Mini Code — T4 stub upgraded to T3 primary source

Quality cycle: retried the JS-gated The New Stack stub, then sourced the primary instead — cohere's own blog (cohere.com/blog/north-mini-code) + the Cohere Labs Hugging Face model card (found via WebSearch). Rewrote cohere-north-mini-code from a headline-only T4 stub into a full T3 model page (retyped TechArticle → SoftwareApplication): 30B MoE / 3B active, 128 experts (8/tok), 256K ctx / 64K gen, Apache-2.0, single-H100 FP8; SWE-Bench Verified 83.2% pass@1, Terminal-Bench v2 62.8%, Artificial Analysis Coding Index 33.4. Correction: the earlier stub framing (and the synthesis provider-map) called Cohere a non-open-weight sovereign player — North Mini Code is in fact Apache-2.0 open-weight, so cohere, the synthesis "enterprise/sovereign" bullet, and the index were corrected to put Cohere on the open-weight axis (sovereignty via open weights, not opposite them). url swapped thenewstack.io → cohere.com.

[2026-06-18] ingest | GLM-5.2 (Z.ai) — Simon Willison review (hub-routed)

Hub-routed Telegram drop (simon-willison's post; T2, independent practitioner). Clear match — the open-weight text-LLM market; no runner-up (text-only model, so not speech-audio; not inference mechanics). GLM/Z.ai were plain-text mentions before this; now paged. New:

• glm-52 (source, url-only, SoftwareApplication): MIT, 753B-A40B MoE, 1.51TB, 1M ctx, text-only. #1 open-weight on artificial-analysis Intelligence Index v4.1 (51, ahead of MiniMax-M3 44 / DeepSeek V4 Pro 44); #2 Code Arena WebDev behind Claude Fable 5. ~$1.40/$4.40 per 1M via OpenRouter. Caveat recorded: token-hungry (~43K output tok/task), so the cheap rate overstates the real-cost gap.
• New entities: z-ai (Organization, fmr Zhipu AI — the GLM maker, open-weight challenger slot) and simon-willison (Person, recurring cross-wiki practitioner voice; the SVG-pelican tester).
• Folded into synthesis (open-weight-challenger bullet + the "does the frontier premium survive?" open question — strongest open-vs-closed data point yet, an open model #2 on a coding board at ~1/5 the price). Updated open-weight-models (GLM-5.1→5.2) and artificial-analysis (noted v4.1). No contradiction with existing pages. Volatile snapshot dated 2026-06-18.

[2026-06-18] ingest | DefinedTerm enrichment pass (subagent)

Deepened the spoke's thinnest DefinedTerm pages with sourced mechanism detail (all five were ~32–40 lines). Two new source pages + one refreshed in place:

• hf-quantization-concepts (TechArticle, T1, huggingface.co) — int8 ~4× smaller than FP32 / int4 halves again + weight-packing & memory-bandwidth payoff; FP8 E4M3/E5M2; affine scale+zero-point; per-tensor vs per-channel; PTQ vs QAT (QAT better at low bit-widths). Drove quantization: added "size math", "why int4 is the workhorse + FP8 alternative", and grounded the QAT-vs-PTQ claim.
• moe-architecture (Article, T2, en.wikipedia.org) — gating/router picks top-k experts (k=1/2), conditional computation, total-vs-active split (less compute despite 30× more params), shared experts (DeepSeek). Drove open-weight-models "MoE dominance" bullet (now explains the mechanism behind large-total/small-active serving).
• artificial-analysis refreshed in place to v4.1 composition: 9 evals, four weighted categories (Agents 34% / Coding 24% / SciReasoning 24% / General 18%; agents+coding = 58%), ±1% 95% CI. Drove llm-benchmarks: new "Anatomy of a composite" section.
• llm-provider and llm-api-pricing already developed with Related present — left untouched (no padding). Index updated with the two new source rows. One blocked fetch: en.wikipedia.org/wiki/Quantization_(machine_learning) 404'd; substituted the HF concepts doc. All volatile figures dated 2026-06-18; cross-wiki bridges (llm-inference, kv-cache, claude-opus-4-8) left as-is.

[2026-06-28] ingest | Claude API — Rate limits

New source page claude-api-rate-limits (TechArticle, T1, platform.claude.com), routed from the hub. Anthropic's rate-limit reference: two limit types (monthly spend caps — Start $500 / Build $1,000 / Scale $200,000 / Custom none — and per-model RPM/ITPM/OTPM), token-bucket pacing (continuous replenish, bursts trip 429s), auto tier promotion, separate pools for Batches / Managed Agents / Fast mode, and the full anthropic-ratelimit-* header set + per-workspace limits. Dominant hook for this spoke = cache-aware ITPM: cached input tokens (cache_read_input_tokens) don't count toward the rate limit (except Haiku 3.5 †), so prompt caching is a throughput lever, not only a price cut (2M ITPM @ 80% hit ≈ 10M tok/min). Folded into llm-api-pricing (caching bullet) and the synthesis "engineering sets real cost" recurring read; sits beside claude-refusals-and-fallback as the provider-API-access pair. No new entities (publisher Anthropic already a cross-wiki bridge node). All tier numbers dated 2026-06-28 (volatile snapshot).

[2026-06-30] ingest | Claude Sonnet 5 (Anthropic model release)

Routed by the hub router from Telegram. New model claude-sonnet-5 (claude-sonnet-5), released 2026-06-30 — Anthropic's "most agentic Sonnet yet," pitched as performing close to Opus 4.8 at the Sonnet price. Created a single SoftwareApplication + source:true page (the model- announcement precedent set by cohere-north-mini-code / glm-52). T1 (primary vendor announcement), freshness volatile. Facts: pricing $2/$10 intro through 2026-08-31, $3/$15 standard; HLE 34.6% (no tools) / 46.8% (w-tools); OSWorld-Verified 78.5%; default for Free/Pro, on Claude Code/Platform/API; safety reported lower undesirable-behavior rate than Sonnet 4.6, weak on cyber exploit dev vs Opus; context window not stated. Folded into synthesis open question "does the frontier premium survive?" as a new angle — the squeeze coming from inside a frontier lab, not only open weights. Index updated. No new entities (Anthropic + Opus 4.8 already cross-wiki bridge nodes in ../research-wiki). Vendor-reported numbers await an independent artificial-analysis re-run.

[2026-06-30] ingest | Claude Sonnet 5 System Card (PDF, 145pp)

Follow-up to the same-day claude-sonnet-5 announcement — the primary pre-deployment report (routed from Telegram). Created source page claude-sonnet-5-system-card (@type Report, T1) with the full capability table (Sonnet 5 vs Sonnet 4.6 / GPT-5.5 / Gemini 3.5 Flash), RSP determination, and alignment/welfare assessment. Correction: the announcement-page figures (HLE 34.6/46.8, OSWorld 78.5) were Sonnet 4.6's column in the card — a lossy-fetch misattribution; corrected the claude-sonnet-5 model page to the card's Sonnet 5 numbers (SWE-bench Verified 85.2, HLE 43.2/57.4, OSWorld 81.2, FrontierCode v1 38.8) and flagged the correction in place per record-don't-overwrite. Filled the open context-window field: 1M standard, 10M via context compaction (trigger 200k). Two notable new facts folded into synthesis: (1) the card admits Sonnet 5 "trails our Opus and Mythos-class models" on almost all benchmarks — so "close to Opus" is price-adjusted, not parity; (2) it discloses a Mythos class above Opus (Claude Mythos 5) as Anthropic's current frontier. Cross-spoke: the RSP/threshold/safeguard machinery is ../ai-governance-wiki substance (noted in the source page; runner-up that spoke). No new entities (Anthropic + Opus 4.8 already cross-wiki bridge nodes; Mythos 5 not yet paged — left as a named future node). All numbers dated 2026-06-30 (volatile).

[2026-06-30] ingest | Agents-A1 (InternScience) — open-weight 35B agentic MoE

T3 first-party (GitHub repo + model card; self-reported benchmarks). Routed here by the hub router: it's a model (weights + training method), not a harness, so it's the model-market's subject — the agentic capability angle is cross-spoke context with ../agentic-tooling-wiki, noted, not split. Apache-2.0, 35B MoE (active params undisclosed), 262K context, built on Qwen3 lineage (card ships a qwen3 reasoning parser). Pitch: "trillion-parameter performance with a 35B agent" — close the gap by scaling the agent horizon, not parameters, via a 3-stage supervised method (full-domain SFT → per-domain teachers → multi-teacher domain-routed on-policy distillation), not RL. Six unified domains (long-horizon search / engineering / scientific research / instruction-following / tool-calling + general). Self-reported: GAIA 96.04, IFEval 94.82, Seal-0 56.36, SciCode 44.33, FrontierScience-Olympiad 79.0. New pages agents-a1 (model+source), internscience (Organization). Touched synthesis — added a third mechanism to "does the frontier premium survive?" (training technique, after cheaper tokens and flagship-self-squeeze) and listed A1 among the open-weight challengers. arXiv tech report 2606.30616 (primary, but benchmark claims still self-run); awaits an independent artificial-analysis re-run. Volatile snapshot.

[2026-06-30] ingest | Prompting Claude Sonnet 5 (Anthropic docs) — refresh of an existing model

T1 primary (platform.claude.com). Dedup: subject claude-sonnet-5 already paged → enriched in place (new "API behavior & migration" section) + new source summary prompting-claude-sonnet-5. Key facts: sampling params (temperature/top_p/top_k) now rejected with 400; manual extended thinking (budget_tokens) removed; adaptive thinking on by default; effort param (low→max, default high, xhigh for hardest) with the cross-gen mapping medium S5 ≈ high S4.6 / high S5 ≈ max S4.6; new tokenizer ~30% more tokens (truncates 4.6-tuned max_tokens; raises effective cost); more agentic + more literal instruction-following; code-review recall drop = harness effect not regression; computer use computer_20251124. Touched synthesis — the tokenizer adds a Sonnet-5 instance of the "token accounting, not rate, sets the bill" thread (alongside GLM-5.2 token-hunger), with the effort-tier uplift as the offsetting lever. Cross-spoke: harness-facing advice → ../agentic-tooling-wiki; adaptive-thinking mechanics → ../llm-inference-wiki (noted, not split). Volatile snapshot.

[2026-06-30] ingest | Nano Banana 2 Lite & Gemini Omni Flash (Google) — Gemini generative-media tiers

T3 vendor blog (blog.google). Two gemini-family generative-media models, now with concrete IDs + pricing (the May roundup only had positioning). Nano Banana 2 Lite (gemini-3.1-flash-lite-image): text-to-image, ~4s, $0.034/1k images, replaces gen-1 Nano Banana; claims prompt adherence, character consistency, legible in-image text (Elo-vs-latency-vs-cost chart). Gemini Omni Flash (gemini-omni-flash-preview): video gen + conversational editing, $0.10/sec = Veo 3.1 Fast, 10s max, preview (no audio refs, no API scene-extension, character-consistency wobble). Pitched as an image→video pipeline (Nano Banana still → Omni Flash animation; Interactions API ≤3 sequential edits). On AI Studio / Gemini API / Enterprise Agent Platform + consumer surfaces. New source gemini-omni-flash-nano-banana-2-lite; enriched gemini (two release entries). Touched synthesis — added a generative-media sub-theme note: new pricing units ($/image, $/sec) and a spin-out watch (a dedicated image/video-gen spoke if ≥3 non-Google generative-media sources land; parallels the audio-gen→speech-audio carve-out). Volatile.

[2026-07-05] ingest | Fable 5 vs Sonnet 5 — token cost (Njenga, Medium) → llm-providers-wiki

Routed from hub. Member-only Medium blog (T4, paywalled past the baseline) comparing Fable 5 and Sonnet 5 cost inside Claude Code. Only the baseline test is readable: a single "Hello" cost $0.4745 on Fable 5 vs ~3.2× less on Sonnet 5, with Sonnet reading more from cache (24.8k vs 15.6k) and so billing below fresh input. New source page fable-5-vs-sonnet-5-token-cost; new Thing page claude-fable-5 (first page for it — folds in the existing Code Arena WebDev coding-lead datapoint, so "fast, coding-strong", which makes the cost anomaly notable); new joe-njenga Person node (entity discovery). Synthesis: added a caching-efficiency-as-intra-lab-cost-axis open question + logged the unverified anomaly under Contradictions (held pending a T1 Anthropic price sheet). Cross-spoke context noted on the source page (Claude Code → agentic-tooling; prompt caching → llm-inference kv-cache) — not re-paged. Numbers are one unrepeated, harness-inflated run; recorded as claim, not fact.

[2026-07-08] ingest | Grok 4.5 cheap vs Fable 5 / GPT 5.5 (The Decoder) → llm-providers-wiki

Routed from hub (Telegram). New source grok-4-5-price-vs-benchmarks (Article, T3, the-decoder.com): Grok 4.5 at $2/$6 per 1M undercuts Opus 4.8 ($5/$25), GPT-5.5 ($5/$30), and Fable 5 ($10/$50); it trails on coding benchmarks (DeepSWE 53% vs Fable 5 70%, Terminal Bench 83.3%, SWE Bench Pro 64.7%) but xAI argues cost-per-task wins (claims 4.2× fewer tokens than Opus 4.8, ~80 tok/s) — the "good-enough + cheap" China/deepseek play from a closed US lab. Touched 4 pages: xai-grok (repositioned from premium to price-leader; added Grok 4.5 pricing/benchmarks/strategy), claude-fable-5 ($10/$50 rate corroborates Njenga's ~3.2× anomaly — Fable is genuinely dear, likely a premium tier not a budget one), llm-api-pricing (frontier snapshot + the cost-per-task = price × tokens-per-task lever), and synthesis (xAI blurs the premium-vs-cheap split; cost-per-task axis; Fable-expensive now on two reads → moved from unverified anomaly to two-source corroboration in Contradictions). Still no T1 Anthropic price sheet for Fable 5 — reads remain third-party. Entity (The Decoder outlet) deferred. T3 (AI-news, specific numbers but not primary rate cards; token-efficiency figures are xAI's). Ran avoid-ai-writing. +1 page.

[2026-07-08] ingest | OpenAI GPT-5.6 (Sol/Terra/Luna) vs Grok 4.5 clash (Business Insider) → llm-providers-wiki

Routed from hub (Telegram). Fetched via firecrawl (WebFetch hard-blocked by BI; paywalled subscriber page). New source openai-gpt56-grok45-clash (NewsArticle, T3, Hugh Langley, 2026-07-08). Substance: OpenAI GPT-5.6 rolls out wider Thursday (staggered at the Trump administration's request) in three flavors — Sol (flagship; agentic coding/biology/cybersecurity), Terra (everyday), Luna (speed+affordability); Musk's Grok 4.5 launches "in coming days," billed "Opus-class but faster, more token-efficient and lower cost" (his own words = 2nd source on grok-4-5-price-vs-benchmarks's cost-per-task framing); Gemini 3.5 Pro later this month. Governance backdrop: Trump admin export-banned Anthropic's Fable & Mythos (cyberattack-misuse risk) → Anthropic suspended → Fable 5 redeployed last week. Musk–Altman rivalry (Musk lost his 2024 suit). Touched 4 pages: openai (added GPT-5.6 Sol/Terra/Luna

• Luna token-economy), xai-grok (Musk's framing + launch timing), claude-fable-5 (export-ban/suspend/ redeploy availability shock + Mythos), synthesis (efficiency now the shared closed-lab pitch; tiering goes intra-family; new export-control availability axis + ai-governance cross-spoke adjacency). T3 (reputable journalism; launch specs are pre-release vendor claims). Ran avoid-ai-writing. +1 page.

[2026-07-09] ingest | "Don't use Claude Fable 5" (Ruben Hassid, Substack) → llm-providers-wiki

Hub-routed (Telegram). New source dont-use-fable-5-hassid (BlogPosting, T3, opinion; non-technical AI consultant translating docs). Value: a third independent read of $10/$50 (with grok-4-5-price-vs-benchmarks

• fable-5-vs-sonnet-5-token-cost), so the sticker is now well-corroborated (still no T1 Anthropic sheet). Adds two new facts: pay-per-use credits after 2026-07-12, and usage economics — cost scales with conversation length (whole thread re-read each turn: ≈$0.15 short / ≈$6 at 19 turns / ≈$14 at 40), reframing the open "which lever dominates net cost" toward turns/context, not just per-token rate. Buyer's ladder: Fable 5 for hard goals at high/max effort (1–2 turns), Opus 4.8 as cheaper+smarter workhorse, Sonnet 5 not justified over Opus — a practitioner inversion of Anthropic's mid-tier push. Touched claude-fable-5 (3rd read + usage economics + positioning; refreshed index one-liner), synthesis (Fable pricing bullet). Paged author ruben-hassid (Person). Cross-spoke: the tasks→goals + effort framing echoes ../agentic-tooling-wiki's effort-level (linked). Ran avoid-ai-writing. +2 pages.

[2026-07-10] ingest | Mistral homepage — model lab → sovereign full-stack platform (hub-routed, Telegram)

Ingested mistral.ai (first-party site). New source summary mistral-platform (WebSite, source, T2 — first-party marketing, volatile) and refreshed the mistral-ai org page in place (its first update since the 2026-06-09 Wikipedia ingest). The story: Mistral now presents as a sovereign, EU-hosted full-stack platform, not just a model lab — product suite Vibe (work agent; the renamed Le Chat) / Vibe for Code / Studio / Forge / Compute over the model line; deployment self-hosted / EU-cloud / partner (AWS/GCP/Azure/IBM/SAP/Snowflake/NVIDIA/ Outscale); customers HSBC, ASML, Stellantis, EPO. New models named (no specs): Medium 3.5, Small 4, OCR 4, Voxtral TTS — named inline, not stubbed (proportionate to a spec-less marketing page). Synthesis: expanded the "European open-weight" position to "European open-weight → sovereign platform," collapsing the gap to the cohere enterprise/sovereign row (Mistral reaches data-residency from the open-weight side). Cross-wiki adjacencies flagged: agentic-tooling (Vibe for Code / Studio agents), speech-audio (Voxtral TTS). No structural move → no verify. +1 page, 1 refreshed.

[2026-07-16] ingest | Inkling — Thinking Machines Lab's open-weights model (hub-routed, Telegram)

Routed source: inkling-willison-review (simonwillison.net, 2026-07-16, T2). Pulled the first-party announcement + model card it links as a second source, inkling-announcement (T1), because the practitioner post carries the framing but not the architecture, benchmark table, or hardware numbers.

New pages (5): inkling (SoftwareApplication — Apache-2.0, 975B-A41B MoE, 45T tokens of text/image/audio/ video, 1M ctx, 66 layers, 256 routed + 2 shared experts, relative pos-embeddings not RoPE; Inkling-Small 276B-A12B in testing), thinking-machines-lab (Organization), tinker (SoftwareApplication — the fine-tuning platform), plus the two source summaries. Refreshed open-weight-models and simon-willison in place. No mira-murati page: leadership didn't survive the evidence-only check against the fetched text — left for a later source.

Synthesis: new position on the provider map — US open-weight, customization-as-product. Thinking Machines ships Apache-2.0 weights while stating outright the model is "not the strongest overall model available today, open or closed," and monetizes fine-tuning on Tinker. No frontier tier to seed, unlike google's Gemma/Gemini dual track; nearest neighbour is Mistral's Forge, reached from the opposite direction. Two new open questions: (1) is "open weight" still one category — Inkling wants 2TB+ VRAM (600GB at NVFP4) against Gemma 4 12B's 16GB, same licence, ~125× the hardware, so only the local tier really attacks the price floor; (2) does free-weights-paid-tuning work as a business. Also fed the compatibility-as-strategy thread: a brand-new lab still ships an OpenAI-compatible API.

Caveats recorded: benchmarks self-reported at effort=0.99, no artificial-analysis re-run; training data documented only as "publicly accessible data repositories"; model card admits role-play/indirect-framing safety leakage. No structural move → no verify. +5 pages, 2 refreshed.

[2026-07-22] ingest | Mage-Flow-Turbo (Microsoft, Hugging Face)

Routed from the hub (route in ../log.md). URL-only, T3 — first-party model card, self-reported benchmarks against a self-chosen (all-open) comparison set, no independent re-run. New page: mage-flow-turbo (SoftwareApplication) — MIT-licensed 4B NR-MMDiT text-to-image, rectified flow matching in a custom Mage-VAE latent space, 4-step distilled Turbo variant, native-resolution packing 512–2048px at any aspect ratio, 0.59 s per 1024² on a single A100, GenEval 0.88, claimed parity-or-better vs Qwen-Image 20B / Z-Image 6B / FLUX.2 32B. arXiv 2607.19064; diffusers + mage-flow CLI + Gradio app. Dedup: no mage/microsoft/flux page in wiki/. No microsoft Organization page created — a canonical microsoft node already exists in ../agentic-tooling-wiki, so it is linked cross-wiki per the bridge-node convention in CLAUDE.md rather than duplicated. No other entities paged. Gap-relevance: feeds two live threads. (1) Generative-media sub-theme — this is source 2 of 3 and the first non-Google one, the exact trigger condition gemini-omni-flash-nano-banana-2-lite wrote down for spinning out an image/video-gen spoke; synthesis section updated, cluster tally recorded, not yet spun out. (2) "Parameter count is the wrong axis" — 4B vs 20–32B rivals is agents-a1's claim in a different modality, reached by co-design + step distillation instead of agent horizon. Also noted: it adds no pricing unit (open weights, self-hosted), which sharpens the open-vs-hosted split inside the generative-media cluster — Google's tier has $/image, this has none. And the corpus has no artificial-analysis-style independent referee for image models, logged as the sub-theme's gap. Verify deferred per hub policy (content-only). avoid-ai-writing run. 52 → 53 pages.

[2026-07-23] ingest | Fable 5 vs GPT-5.6 vs Kimi K3 for creators (Medium)

Routed from the hub (route in ../log.md). URL-only, T4 — Medium creator-opinion essay, benchmark positions asserted with no scores/leaderboard/methodology. New pages: fable5-gpt56-kimi-k3-creators (OpinionNewsArticle source) and kimi-k3 (thin anchor). Kept the K3 claim recorded but did not mint a real model page off one weak source: kimi-k3 is an explicit thin, unverified anchor (2.8T open-weight, weights due 2026-07-27, claimed #1 editorial writing

• edging Fable 5 on frontend coding) flagged for upgrade on the 27th. No Moonshot-AI org page created (delinked — no distinct evidence; Moonshot already appears inline across open-weight-models/ open-source-llms-2026/llm-provider). Dedup: Kimi was a recurring inline mention (K2.6, ~1.1T) with no page; linked the K2.6 mention in open-source-llms-2026 forward to kimi-k3 as the lineage successor. Gap-relevance: feeds the "does the frontier premium survive?" open question (an open model claimed at/above a closed frontier model on writing+coding) and restates the benchmark-vs-real-utility gap from a creator's seat (cousin of dont-use-fable-5-hassid/fable-5-vs-sonnet-5-token-cost). Synthesis: added a flagged weak watch item under the frontier-premium question, not a finding. Entities: none paged (author a Medium byline; Moonshot AI evidence-thin here). Verify deferred per hub policy (content-only). avoid-ai-writing run.

[2026-07-23] ingest | Alphabet Q2 2026 earnings — Pichai's remarks (Google)

Routed from the hub (route in ../log.md). URL-only, T1 for provenance (first-party CEO message); all figures company-reported and unaudited here — dated snapshot. Broad multi-spoke source handled per HUB "three+ spokes" rule: routed whole to llm-providers-wiki as the dominant in-scope substance (the Gemini model lineup + Google's model-scale metrics), with other facets recorded as cross-spoke context inside the source page, not fragmented into other spokes. New page: alphabet-q2-2026-earnings (BlogPosting source). In-scope core: Gemini 3.6 Flash, 3.5 Flash-Lite, 3.5 Flash Cyber (first domain-specialized Gemini logged here), 3.5 Pro testing, Gemini 4 pre-training; consumer Omni video-in-app (+40% DAU). Scale: ~22B tokens/min (↑~38% Q/Q), 950M Gemini-app MAU, 9M+ developers, ~900M Gemma downloads (Gemma 4 ~300M), enterprise token-concentration (500 customers >1T tokens/yr). Updated gemini (Q2 lineup bullet incl. the Cyber domain-variant + Gemini 4 pretraining) and google (new "Scale (Q2 2026)" section — the usage-volume view of the dual track). Synthesis: extended the dual-track "Recurring reads" bullet with the usage-scale dimension — the market's growth axis is now consumption volume (the demand-side complement to the output-token cost thread), plus the tier/vertical differentiation (cost ladder + Cyber). Dedup: no Gemini-3.6/3.5-tier page existed; folded into the existing gemini anchor rather than a page per model (dated-snapshot convention). Cross-spoke facets logged (see below), not paged. Entities: none new — google canonical node updated; Pichai not paged (spoke convention: page the org/model, not the exec byline). Verify deferred per hub policy (content-only). avoid-ai-writing run.

[2026-07-24] ingest | Introducing Claude Opus 5 (Anthropic)

Routed from the hub (route in ../log.md). URL-only, T1 for provenance (first-party launch post); the benchmark claims inside are vendor-self-reported and stated as ratios with no absolute scores, so they're treated as T3-grade evidence on a T1 document. New pages: claude-opus-5-announcement (BlogPosting source) + claude-opus-5 (SoftwareApplication). Core facts: claude-opus-5, $5/$25 per 1M — identical to Opus 4.8; fast mode at 2× price for ~2.5× speed; effort levels written out (low/high/xhigh/max); pitched as "close to the frontier intelligence of Fable 5 at half the price"; Frontier-Bench v0.1 >2× Opus 4.8, CursorBench 3.2 within 0.5% of Fable 5, ARC-AGI 3 3× next-best, Zapier AutomationBench ~1.5×, OSWorld 2.0 past Fable 5 at ~1/3 the cost, organic chemistry +10.2 pts. Available on API / Claude.ai / Claude Code / Claude Cowork; AWS, Vertex AI, Microsoft Foundry named as partners. Opus 4.8 not deprecated — it's the fallback for cyber-flagged Cowork requests. Context window is never stated — logged as an open gap on both pages. Updated claude-fable-5: the post is the first T1 source placing Fable 5 above Opus 5 on intelligence, and "half of $5/$25" implies the $10/$50 rate the three third-party reads gave; tier field split (T1 positioning / T3 rate). Synthesis: three folds — (1) the frontier-premium question gains its strongest data point, a flat generation-over-generation price and the lab discounting its own flagship from inside; Mythos 5 downgraded from "capability frontier" (the Sonnet-5 system-card reading) to a cyber/biology specialist; (2) the Fable-5-pricing question gains a T1-implied sticker; (3) a third shape of selling speed — a paid runtime mode on identical weights, next to Cerebras' silicon and OpenAI's Luna tier. Dedup: no Opus-5 page existed; claude-opus-4-8 stays a ../research-wiki bridge node, not duplicated here. Entities: none new — anthropic is the canonical cross-wiki node; no byline on the post. Verify deferred per hub policy (content-only). avoid-ai-writing run.

[2026-07-26] ingest | gemma-4-31B-it-scotoma (ReadyArt) — abliteration of Gemma 4 31B-it

Routed from the hub (Telegram). Hugging Face model card; T3 — it is the official artifact page, so the facts about what it is are primary, but every behavioural claim is the author's unmeasured self-report on their own model. No benchmarks, no evals, no dataset. New pages: gemma-4-31b-it-scotoma (SoftwareApplication source), abliteration (DefinedTerm, mechanism), readyart (Organization). Core facts: Apache-2.0, 33B params BF16, base gemma-4-31B-it; method = locate refusal directions with heretic abliteration → Jacobian-lens projection keeping ~22% of abliteration magnitude → merge at 1.5× scaling into BF16 across layers 7–41; 367 downloads/month; "loosened, not lobotomized"; self-described research artifact. Recorded contradiction: the card claims it "loosens gemma-4-31B-it's cautious reflex" and that "scotoma refuses basically as much as its base model." Both kept, flagged on the page — a refusal-removal technique disclaiming that refusal rates moved. Updated gemma-4: a 31B-it size appears that no Google source in the corpus mentions — added to the family table marked reported-not-confirmed (the 31B name vs 33B param count is the card's own, too). Also a new bullet: derivatives edit the safety layer, not just precision. Updated open-weight-models: fourth structural consequence — a permissive license on downloadable weights is permission to rewrite and redistribute, safety posture included; Apache-2.0 doesn't distinguish running from editing. Synthesis: new thesis paragraph ("open weights have a third face: the behaviour is editable too"), pairing this against claude-refusals-and-fallback — a closed refusal can be routed around, an open one deleted. New open question: what is a safety claim worth on a downloadable model? Dedup: no abliteration/refusal-editing page existed; gemma-4 refreshed in place, not duplicated. Entities: readyart (publisher — in-domain in its own right as a model publisher; the llm-provider typology has no slot for a train-nothing derivative shop, which is itself the finding). No author node — the card carries no byline. Cross-spoke: the Jacobian lens is the parked hub _inbox ai-interpretability stray (tally 1, 2026-07-10). This is sighting 2 and the first applied use — noted on the page, not parked; cluster still short of the ≥3 spin-out trigger. Verify deferred per hub policy (content-only). avoid-ai-writing run.

[2026-07-26] ingest | OmniRoute — the free MIT AI gateway (github.com/diegosouzapw/OmniRoute)

Routed here by the hub (runner-up: ../agentic-tooling-wiki, which owns the 33+ coding agents and the 104 MCP tools it exposes — but the substance is provider aggregation, quota and cost, i.e. this spoke's API-access-and-pricing clause). T1 by origin (project's own repo), with the standing caveat written into the page: every figure is self-reported — 290+ providers, ~1.53B free tokens/month, 89% average compression, and a self-scored comparison table against OpenRouter/LiteLLM. New: omniroute (source), ai-gateway (concept), diego-souza (entity). Updated: llm-api-pricing — the "model routing" cost lever now names the gateway layer, plus two new levers the corpus hadn't recorded: client-side prompt compression and free-tier aggregation. amazon-bedrock reframed as the managed shape of the same layer. Synthesis: the reseller paragraph gains its self-hosted twin — the reseller commoditized model choice, the local router commoditizes the provider's generosity. Two new open questions: how long free tiers survive being pooled, and whether compression becomes a provider feature or stays adversarial. Dedup: no gateway/router/OpenRouter page existed; the concept was only implicit in amazon-bedrock. The nearest prior art was the one-line "model routing" lever in llm-api-pricing, now expanded. Gap-relevance: gives the demand-side thread its second data point and its first non-cloud one. Verify deferred per hub policy (content-only). avoid-ai-writing run.

[2026-07-27] ingest | Opus 5 blows past Fable 5 and GPT-5.6 Sol on ARC-AGI-3 (The Decoder, 2026-07-26)

Routed here by the hub (runner-up: none — a benchmark result on the model market is this spoke's "capability/cost benchmarks" axis; ../research-wiki owns Anthropic-as-substrate, not leaderboards). T3 tech-press reporting on ARC Prize's analysis; the underlying scores would be T2 read direct. New: opus-5-arc-agi-3 (source), arc-agi (benchmark concept), arc-prize (entity). Updated: claude-opus-5 (first third-party numbers + the unreconciled 3× claim), claude-fable-5 (first non-Anthropic capability read on it — every earlier one was price), llm-benchmarks (novelty benchmarks as a second kind), synthesis (new tension section). What it adds:

1. An independent referee. Nearly every capability figure in this spoke is the lab's own. ARC Prize scores the model with no harness, publishes results/replays/code, and hedges its own finding (Kamradt: Opus 4.8 won some environments outright).
2. A vendor ratio that doesn't reconcile. Anthropic's launch post says "3× the next-best model on ARC-AGI 3." Against the 7.8% record that's ~3.9×; against ARC Prize's ~20% Fable-class figure it's ~1.5×. Recorded as a tension, both numbers kept.
3. The decay of a novelty benchmark. Opus 5 postdates ARC-AGI-3's public format; the ~4× jump doesn't reproduce on Ning's private Witness (43.4, tied with Fable 5 and Kimi K3, and below Opus 4.8 on the one unfamiliar rule combination). The generalizable read is that a novelty test's shelf life starts the day its format goes public. Entities deferred: Matthias Bastian (author), Greg Kamradt, Guanghan Ning, The Decoder (publisher — 2nd source from it after grok-4-5-price-vs-benchmarks; page it on a 3rd). Witness benchmark noted on arc-agi rather than paged; page it if it recurs. Verify deferred per hub policy (content-only). avoid-ai-writing run.

[2026-07-28] ingest | Sakana Fugu — Get Started (fugu.sakana.ai)

Routed here by the hub (runner-up: ../agentic-tooling-wiki — Fugu is a multi-agent system and installs into Codex/Claude Code, but the document is API onboarding: keys, endpoints, model IDs, effort levels, provider pools, pricing pointer. Same shape as deepseek-api-docs / claude-api-rate-limits, which live here. The agent-system angle is cross-linked, not split). T1 first-party docs for the interface; states no price, no benchmark, no context window. Link arrived as a share.google redirect (second time) — resolved out of the share page's HTML, then probed for the canonical fugu.sakana.ai/get-started. New: fugu-get-started (source), sakana-fugu (model), sakana-ai (provider — first Japanese lab in the spoke). Updated: ai-gateway (a third shape: the gateway that calls itself a model), synthesis (a new section on "model" as a product boundary), index. What it changes:

1. A multi-agent system sold as a model. The docs concede it in their first sentence. fugu routes across all supported providers by default, with the pool narrowed per API key rather than per request — ai-gateway behaviour where the routing is the product and the model string is the label. Three of this spoke's assumptions break: you can't say which model answered (so a benchmark measures a routing policy), you can't quote a per-token price for a named model, and the frontier-lab/cloud-reseller distinction in llm-provider collapses into one endpoint.
2. The SKU has stopped tracking the weights. Read beside claude-opus-5's fast mode (a runtime toggle on identical weights), the trend runs both ways — one product over many models, two products over one model. What the buyer purchases is a configuration, which is why capability and price claims now have to name what they measured.
3. Cyber is a product axis across three labs — Mythos (claude-opus-5), Gemini 3.5 Flash Cyber (alphabet-q2-2026-earnings), fugu-cyber — none defining the category, and Anthropic's cyber-capable line already drew an export ban (claude-fable-5). Two operational tells recorded: the effort dial starts at high (no low/medium; max only on fugu-ultra-v1.1), and the Codex provider block raises stream_idle_timeout_ms to 2 hours against Codex's ~5-min default, which is the vendor telling you its turns run at agent scale. Also noted for ../agentic-tooling-wiki: the shipped base_instructions carry agent-conduct guards (don't run commands that would kill your own session; never force-kill by raw PID) — a provider shipping guardrails inside the model catalog, i.e. the instruction end of the axis nono argues must be enforced. Verify deferred per hub policy (content-only). avoid-ai-writing run.

[2026-07-28] ingest | Kimi K3 — Open Frontier Intelligence (MoonshotAI/Kimi-K3)

Refresh in place, exactly as the placeholder page asked for: kimi-k3 was a T4 anchor built on one Medium essay (fable5-gpt56-kimi-k3-creators) with the note "revisit and upgrade on the 27th." The weights landed; the page is now T1 from the repo and technical report. No duplicate created. Updated: kimi-k3 (rewritten), open-weight-models (a third licence shape), fable5-gpt56-kimi-k3-creators (claims scored against the release), synthesis (the frontier-premium question), index. Three things worth keeping:

1. A split decision against current flagships. 2.8T total / 104B active (16 of 896 experts), 93 layers (69 Kimi Delta Attention + 24 Gated MLA), 1M ctx, native MXFP4 weights via QAT, MoonViT-V2 vision. Ahead on GPQA Diamond (93.5 vs Opus 4.8's 91.0), SWE-Marathon (42.0 vs Fable 5's 35.0), BrowseComp and Video-MME; behind on DeepSWE (67.5 vs 70.0). Trading wins with the closed frontier is a harder data point than glm-52 or agents-a1, which approached it.
2. A vendor table that documents its own confounds — different harnesses per model (K3 in Kimi Code, rivals in Claude Code/Codex/Terminus; the harness alone moves K3's DeepSWE 67.5→67.3), a modified H20-calibrated SWE-Marathon branch, BrowseComp needing context compaction (91.2 with, 90.4 on the raw 1M window), and an explicit note that Fable 5 hit fallbacks on 35% of tasks in their run, i.e. caveating their own win. Above the disclosure standard of claude-opus-5-announcement's ratio-only claims, still first-party.
3. The licence tightened as the capability closed. K2.6 was Modified MIT; K3 ships a bespoke licence, MIT-shaped except for a Model-as-a-Service revenue threshold (US$20M over any 12 months) requiring a separate agreement — Llama's move, aimed at clouds and API resellers rather than users. open-weight-models's licensing axis now has three shapes, and the synthesis records the two-sided answer: capability is competed away from below faster than openness is. Calibration recorded on the T4 source: it got the specs right (2.8T, open weights, the date) and the rankings wrong ("#1 editorial writing" is absent from the release; "edged Fable 5 on coding" is half true). Specs propagate; comparative judgments don't. Verify deferred per hub policy (content-only). avoid-ai-writing run.

[2026-07-28] ingest | DeepSeek V4 vs GLM-5.2 vs Qwen: 10x Price Gap (tech-insider.org)

Routed here by the hub. T4 — and the tier is the story. Affiliate leaderboard ad (trovedrops with sub_id tracking), Google Preferred Sources widget, and a sibling article using the identical X vs Y vs Z: $N Price Gap [2026] template on a different model trio. A monetised content template, not an analysis anyone needed to write. No methodology on any figure, no benchmark linked to its scoreboard. WebFetch 403'd; retrieved via Firecrawl per HUB edge handling. New: deepseek-glm-qwen-price-gap (source). Updated: glm-52 (contested-pricing section), synthesis (new "Whose price is on the sticker?" section), index. No new entities — deepseek, qwen, z-ai all already paged.

The value of ingesting a T4 source into a spoke that already holds T1 and T2 on the same subjects is that it can be graded, so that's what the page does. Corroborated: GLM-5.2 at 753B MoE / ~40B active, 1M context, MIT — matches glm-52 exactly (T2, Willison on first-party facts). deepseek-v4-flash / deepseek-v4-pro and context caching match deepseek-api-docs (T1, vendor's own reference). The spec sheet holds up. Contradicted, minor: GLM-5.2 release date given as 2026-06-13; glm-52 has 2026-06-16. Contradicted, and it's the headline: the article calls $1.40 / $4.40 GLM-5.2's vendor list price and reports third-party resale separately at ~$0.55 / $1.85. glm-52 attributes $1.40 / $4.40 to OpenRouter hosts — i.e. resale. The "10x price gap" to DeepSeek V4 Flash ($0.14 input) is computed off $1.40 being Z.ai's own rate; if Willison's attribution is right, the comparison sets a vendor price against a reseller price and calls the difference a vendor gap. Recorded as an open conflict — neither source is Z.ai's pricing page, and nothing here resolves it. Unverifiable: 93.5% LiveCodeBench, +6.7 SWE-bench Pro, +17 long-horizon agentic — no scoreboard, no date. The one independent ranking held here points the other way on overall capability (artificial-analysis Intelligence Index v4.1: GLM-5.2 51, DeepSeek V4 Pro 44). Different benchmarks, so not a flat contradiction, but a reader taking the article's framing as a general capability claim would be misled.

What survives the tier: it's a snapshot of how the open-weight price floor is marketed (three Chinese labs, two MIT + one Apache-2.0, all pitched on undercutting closed APIs — the open-weight wedge showing up in SEO copy), and its resale-layer observation is real and thin in this corpus. Folded into synthesis as the structural point: open weights make a model's price a market, not a number, so "what does GLM-5.2 cost" has no single answer and comparisons against single-vendor closed models aren't like-for-like. That sits next to the tokens-per-task multiplier already in llm-api-pricing — two separate ways a headline per-token rate fails to predict a bill. Verify deferred per hub policy (content-only). avoid-ai-writing run.

[2026-08-02] ingest | Introducing the Neutrino-1 models (Fermion Research)

Routed here by the hub from Telegram. T1 — first-party primary release post, dated 2026-07-27. Runner-up spokes: ../llm-inference-wiki (real claim, see below) and ../machine-learning-wiki.

Pages. neutrino-1 as the source summary (SoftwareApplication, the announcement-as-model-page pattern this spoke already uses for cohere-north-mini-code and agents-a1), plus a new mechanism page ternary-weights and the lab entity fermion-research. Updated quantization (a third category beside PTQ and QAT), qwen, index and synthesis. No authors are named on the release, so the lab is the only entity; entity-index had no match for it.

Why it needed a new mechanism page. quantization framed low precision as PTQ-or-QAT, and both start from a higher-precision model. Fermion's claim is that there was never one — "no full-precision product model that was rounded afterward" — with rounding a trained model to the same depth said to land near chance. That is a training result wearing a compression result's clothes, and it does not fit inside the existing page. Ternary also now has three named models in the corpus (BitNet b1.58-2B4T, Ternary-Bonsai-8B, Neutrino-1), two of them known only through a competitor's table.

Two things recorded rather than smoothed over.

Every number is Fermion's. The comparison table re-runs each rival on Fermion's own stack. The disclosure is unusually good — it says when its number differs from a rival's card, prints its own losses on IFEval (77.2 vs Gemma-4-E4B's 88.26) and BFCL (68.9 vs Llama-3.1-8B's 76.1), and labels the GSM8K cell where its zero-shot no-CoT run sits beside Llama's 8-shot-with-reasoning 84.5. It is still a table one party controls end to end, with no independent evaluation anywhere. The single externally-anchored measurement is the engine test on BitNet's public weights (102.4 tok/s vs bitnet.cpp's 89.0, same machine, same session) — the only claim where Fermion does not own both sides.

The architecture is Qwen3's, and the post does not say so. Checked against Qwen3-8B/config.json during the ingest: 36 layers, 4,096 hidden, 32 query / 8 KV heads, 128-dim heads, 12,288 FFN, 151,936 vocab, 40,960 positions, untied embeddings, matching parameter count — every stated field. The 0.6B matches Qwen3-0.6B including tied embeddings. Recorded on neutrino-1 with both readings and no verdict: reusing a published Apache-2.0 geometry is normal, legal, and arguably the correct experimental control when the weight format is the variable under test; and a post detailed enough to publish per-layer zero density never naming it is a real omission. synthesis carries it as a new tension — the architecture layer of the same erosion already logged for model strings and SKUs.

Cross-spoke, noted not split. ../llm-inference-wiki owns the KV-cache arithmetic, the bandwidth-bound decode argument and the drafted-decode mechanism (27,648 tokens verified with zero divergences); this spoke owns the format as a market lever — the standing quantization dual-lens split. ../machine-learning-wiki owns §4, which is a genuine post-training finding: behavioural training pressure degrades specific other axes rather than uniformly, protection works only for axes explicitly represented in a stage's batches, and dose curves differ per axis (tool calling saturates ~25 steps, instruction following ~50).

Content-only, no page moves — verify deferred per the hub's standing policy. avoid-ai-writing run over the new prose.

[2026-08-03] ingest | Astra — canonical model node from the ten-proofs announcement

Follow-up to the hub's ten-proofs route earlier today. That source went to ../research-wiki (theorem proving is its cluster E), and the router deliberately did not page Astra there to avoid duplicating a model node outside the spoke that owns the model market. The human asked for the page here, which is where it belongs.

New page: astra (SoftwareApplication, T1 — OpenAI's own publication, read via firecrawl after a 403). Updated: openai (model lineup), index, synthesis (new open question).

What's actually known is thin, and the page says so. OpenAI named Astra on 2026-08-01 in a mathematics paper, not a model announcement: an internal version produced ten results in maths/TCS, humans plus the model wrote the manuscripts, the model wrote the Lean. No params, architecture, context, modality, date, availability, pricing, safety documentation, or benchmark table. Whether the internal version is the one that ships is unstated, as is its relation to GPT-5.6.

The one number needed unpacking. "~$2,000 at Sol API rates" prices an unreleased model's run in the shipping flagship's currency (openai-gpt56-grok45-clash — Sol is the GPT-5.6 flagship). It is a token-volume proxy wearing a dollar sign, not a price; Astra has no rate card. Recorded that way, because the obvious misreading ("ten open problems for $2,000") is a price for a product that can't be bought. Still the most useful cost datum the spoke holds on a frontier result — an order of magnitude for the inference budget behind one, which llm-api-pricing's cost-per-task thread had no figure for.

New open question opened. Astra's public evidence is machine-checkable proofs rather than a vendor eval — the first claim shape here that a stranger can falsify. But only the output is falsifiable; the capability isn't (attempts, steering, and shipped-version parity are all unverifiable). Watch whether other labs copy the form or whether it stays confined to domains with formal checkers.

avoid-ai-writing run. Verify deferred per hub policy (content-only, no page moves).

[2026-08-03] lint | close the Fable 5 price-sheet question — it was answered in a sibling spoke

This spoke's claude-fable-5 page ended with "What's still missing: an actual Anthropic price sheet… The $10/$50 figure remains inferred or third-party." That was false, and had been since before the page existed. Anthropic's 9 June 2026 launch announcement states "Pricing (both): $10 / $50 per Mtok" outright; the hub holds it as research-wiki/claude-fable-5-mythos-5-announcement, created 2026-06-10 — a month before this page (2026-07-05). Corrected the page, its tier: line (the rate is T1, not T3-inferred), the synthesis passage and the index entry. Nothing was retracted: the four-source triangulation (Njenga → The Decoder → Hassid → the Opus 5 arithmetic) stands as independent corroboration that landed on the published number exactly, which is a stronger result than either the sheet or the reads alone. Only the "no first-party rate exists" claim was wrong. The 2026-07-24 paragraph is marked superseded in place rather than rewritten. Still genuinely undisclosed by Anthropic: context window, max output, intended tier. The caching half of the cost question is untouched and still rests on Njenga's single unrepeated run. Systemic finding, logged to the hub: the answer sat in a sibling spoke for two months and this spoke never looked. Entity dedup has npm run entity-index to prevent exactly this for entities; claims have no equivalent.

[2026-08-04] ingest | Anthropic platform model docs — and a correction to our Opus 5 pages

The Claude Opus 5 launch post arrived again via Telegram. Dedup first: already held, ingested 2026-07-24 the day it published, freshness: volatile and only 11 days old — inside the 60d window, so not stale and not due for refresh. Re-ingesting it verbatim would have produced nothing.

So the useful move was the cross-check, not the re-read. Pulling Anthropic's platform documentation — the models overview and the effort reference — against what the announcement says turned up one error and several gaps. New page: anthropic-model-docs (T1).

The correction. The launch post lists the effort levels as "low, high, xhigh, and max." Both our Opus 5 pages copied that. The effort docs say verbatim: "Claude Opus 5 supports all five effort levels." The missing one is medium — and the same page tells developers to use low and medium "liberally" as the primary control for token cost and response time. Corrected on claude-opus-5; flagged in place on claude-opus-5-announcement rather than rewritten, per record-don't-overwrite, since the quote is what the post actually says.

Worth naming the direction: the marketing copy dropped a cost-saving option, not a capability claim. That is the opposite of the failure mode this spoke usually watches for.

Gaps the announcement left, now recorded: 1M context (default and maximum), 128k max output synchronously and 300k on the Batch API, effort defaulting to high, and availability on Bedrock, Claude Platform on AWS, Google Cloud and Microsoft Foundry — four channels the post never mentions.

The find worth carrying into synthesis: Opus 5's reliable knowledge cutoff is May 2026, against January 2026 for both claude-fable-5 and claude-sonnet-5. The tier Anthropic sells as below the frontier has four months more world knowledge than the tier above it. Every comparison this spoke holds is capability-per-dollar; none of them price recency.

Two API facts with product consequences: thinking is on by default and cannot be disabled at xhigh/max (400) — a change from claude-opus-4-8; and effort moves thinking volume, not visible response length, so it is not the verbosity lever it looks like.

Standing lesson: this spoke's Anthropic coverage is built almost entirely from launch posts and press. Those are right for positioning and wrong for facts a developer acts on. The docs are one page deeper on the same domain and were never read.

[2026-08-05] lint | freshness regrade (quality cycle)

gemma-4-qat and google-ai-updates-may-2026 volatile → stable, both 61 days past the 60-day window. Both are dated Google announcements: a BlogPosting records what the vendor said on that date, and re-reading the URL returns the same text, so the flag names re-verification work that can never be done. The second is explicit about it — it is a May 2026 roundup. Whether the lineup they describe still holds is a synthesis.md question, not a freshness one. Rule already in ../QUALITY.md.

[2026-08-20] ingest | awesome-gemma → the variant family, and the 31B confirmed

Routed from the hub (Telegram). github.com/google-gemma/awesome-gemma — a curated Gemma directory, T3: real structure and inclusion criteria, but self-published and mostly relaying Google's own collection pages. Its footnote disclaims official Google support, so the google-gemma org name is recorded as unresolved rather than treated as a Google byline.

Three things came out of it.

The variant family got a page. gemma-variants — 16 first-party domain forks under the Gemma name, more entries than the core models get. The finding is not the list but its shape: they run on Gemma 2, 3 and 4 in parallel, three architectures depart from decoder-only, and two (Gemma Scope 2, DolphinGemma) are research instruments rather than products. "Gemma" names a re-training practice, not a lineup.

The 31B stopped being a rumour. gemma-4 had carried the 31B-it as reported, not confirmed since the gemma-4-31b-it-scotoma ingest, because no Google source in the corpus named the size. The catalog pointed at the arXiv technical report; fetching the primary settled it — abstract gives the suite as 2.3B to 31B. Paged as gemma-4-technical-report (T1, and the spoke's first arXiv source for Gemma). It also showed that encoder-free is a 12B-only choice rather than a family-wide one, which no source here could previously distinguish, and that thinking mode ships in the open track.

A gap the catalog exposed by accident. 14 hosted providers sell one open model per-token, and six of them — Cerebras, Together AI, Fireworks, Baseten, Runpod, Cloudflare Workers AI — have no page anywhere in the corpus. Filed as growth edge 7.

Report benchmark claims ("leap in performance", "rivals larger frontier open models") are the maker grading itself and stay flagged; they now sit against this page's own efficiency-first positioning line as a recorded tension, not a rewrite.

[2026-08-22] ingest | Heretic — automatic censorship removal (github.com/p-e-w/heretic)

Routed from the hub (Telegram). Runner-up spoke: ../machine-learning-wiki (post-training model adaptation toolchain); landed here because the abliteration cluster and the open-weight argument it belongs to are already in this spoke.

New pages: heretic (T1 for method/licence/repo metadata, T3 for its own benchmark table), arditi-refusal-direction (T1, arXiv 2406.11717, read from the abstract), philipp-emanuel-weidmann (Person, from the GitHub profile + repo list).

The gap this closed. abliteration shipped in July with an explicit Gaps line — "the corpus has no primary source on the technique itself," everything traced to one derivative model card. Both ends are now sourced: the paper that found refusal lives in a one-dimensional subspace, and the tool that implements the edit. The page was rewritten around the mechanism (difference-of-means residual direction per layer; orthogonalize the attention out-projection and MLP down-projection against it) instead of around the sketch.

The trade got a number. On gemma-3-12b-it (97/100 refusals unedited), three abliterations all hit 3/100 at KL divergence 1.04 / 0.45 / 0.16, the last from Heretic's automated TPE-on-Optuna search. Refusal suppression is not where the competition is; the collateral damage is. Recorded with the caveat that KL divergence is a proxy for capability and the table is the author's own run — though his evaluator ships with the tool, which is how the rival rows exist.

Two existing pages got sharper, one got smaller. gemma-4-31b-it-scotoma's "no evals" gap is worse than recorded: ReadyArt used Heretic for step 1 and Heretic has a built-in --evaluate-model that would have produced exactly the comparison the card's pitch rests on. readyart is a downstream user of a commodity tool, not an independent weight surgeon — 5,000+ HF models carry the Heretic tag. open-weight-models gains the line that editing safety posture is a build step now.

Synthesis: new subsection Safety alignment is a build-time setting on open weights — for an open-weight model, the refusal behaviour in a launch post describes the published checkpoint, not the artifact users run, and no licence axis the spoke tracks says anything about who may change it. Cross-spoke adjacency extended to ../machine-learning-wiki and sharpened for ../ai-governance-wiki.

[2026-08-28] ingest | Qwen3.8-Flash-Next (Qwen launch post, T1 + Willison link post, T2)

Model: claude-opus-5 · effort: high. Routed here by the hub router from Telegram; the arriving source was Simon Willison's link post, and the primary Qwen release post was fetched and ingested with it (CNN-style blocks did not apply; firecrawl was used because WebFetch returned a thin, partly garbled summary of the link post).

Two new pages. qwen38-flash-next (T1, the launch post): 125B main model + 51B N-gram embedding params, 6B active per token; GDN in three of four layers with Qwen Sparse Attention on the fourth (micro-block indexer, 7.6×/4.9× prefill/decode kernel speedup at 1M ctx); Gated Residual with FP8 residual state; Muon optimizer; ultra-sparse MoE + multi-token prediction. 262,144 native context, 1M with YaRN. Served as qwen3.8-flash at $0.15/$0.47 per 1M. Framed by Qwen as an early preview of the Qwen4 architecture, the same move Qwen3-Next made for Qwen3.5. qwen38-flash-next-willison (T2): Unsloth 72.5 GB / 78.9 GB quants on a DGX Spark within a day, SVG-pelican test at xhigh effort, and a slip ("125B tokens") flagged on the page.

Three caveats recorded rather than smoothed: every capability and throughput number is first-party, two headline benchmarks (CoWorkBench, RecreationBench) are Qwen's own in-house suites, and the launch post states no licence — Apache-2.0 would be an inference from the family, not a fact about this checkpoint.

Updated qwen (new "Shipping the next architecture before the next family" section), moe-architecture (ultra-sparse MoE + capacity that is not an expert), llm-api-pricing (budget rate card on a frontier comparison set), simon-willison, index.md. Synthesis: thesis §2 gains a fourth footprint lever (capacity addressed by lookup, offloaded to host memory), plus a new subsection "The open tier is now previewing architectures, not just chasing them."

[2026-08-29] ingest | Anthropic — Expanding our support for scientists

Routed here by the hub router from Telegram (runner-up: ../ai-governance-wiki). Model: claude-opus-5 · effort: high.

anthropic-scientist-access-expansion (T1, first-party announcement, published 2026-08-27): 10,000 Claude seats for scientists for one year — standard free, premium $15/month at 5× the usage limits — gated on being a principal investigator or equivalent at an academic or nonprofit institution, who then adds their own lab. The AI for Science credits program widens from biology to other fields including compute-heavy work, at up to $50,000 in credits per project, open to any researcher. Claude Science (launched June 2026) is named as the product this feeds, and is described in a single sentence — the wiki has nothing else on it.

The substantive part for this spoke is the last paragraph. Anthropic states that biology and chemistry researchers are limited to Opus-class models, that Fable models continue to block professional biology and drug-development queries on dual-use grounds, and that a US-government-partnered access program for life-sciences professionals to use Mythos-class models has enrolled its first participants. That last item moves the biology trusted-access program recorded on ../research-wiki's claude-mythos-5 from planned (June) to live.

New pages: anthropic-scientist-access-expansion, research-access-programs (concept — access sold to a class of user rather than by the token; the identity gate and the domain gate). Updated: llm-api-pricing (new section, "A third axis: priced by who you are"), claude-fable-5 (a customer segment barred outright, not just priced out), claude-opus-5 (Opus-class is a ceiling for one class of buyer, not the middle rung), index.md. Synthesis: "The tier ladder doesn't order everything" gains the horizontal cut by research field; growth edge 3 partly advanced; two coverage edges added (9 — no other provider's research program is paged, so the concept has no market; 10 — Claude Science has no page); the ../research-wiki adjacency now names diffusion of innovation as the frame for subsidized access.

Recorded as missing: what "5×" multiplies, year-two terms, the program's total value, how "professional biology" is judged per query, and any independent account of uptake.

[2026-09-01] ingest | CodeRabbit's Fable 5.1 code-review benchmark

Routed from the hub (runner-up: ../agentic-tooling-wiki, which owns CodeRabbit-as-a-harness; the dominant substance here is a model comparison, so it landed on this side and the harness angle is a cross-spoke note).

coderabbit-fable-5-1-review (T2) — 45 code-review tasks against 105 catalogued defects, run over Fable 5.1, Fable 5, Opus 5 and GPT-5.6 Sol.

Three things worth keeping. No model wins on both axes: ranked by recall the order is Sol 69.7% → Fable 5.1 61.0% → Opus 5 55.2%, and by precision it inverts (39.3% → 37.3% → 31.6%). Precision is 31.6–39.3% for everyone, so most of what any of them says on this task is not marked valid — a number no capability index in llm-benchmarks reports. And high reasoning scored worse than low on Fable 5.1: 21:36 against 18:38, 57.1% recall against 61.0%. That is the first counter-instance in this spoke to the assumption that effort buys quality monotonically, and it is one unpublished harness, so it is recorded to watch rather than to conclude from.

The 5.1-over-5 upgrade, at the same $10/$50: −34% comments, −70% nitpicks, +49% latency, recall flat. Restraint, not reach.

New: claude-fable-5-1 (model page — 1M context, 128K output, cache reads $0.25/MTok, thinking always on with depth via output_config.effort, forced tool_choice removed, no Priority Tier, ZDR refused), coderabbit-fable-5-1-review, coderabbit. Updated: llm-benchmarks (new section), claude-fable-5 (superseded pointer), index.md, synthesis.md (new section; growth edge 1 partly advanced and re-specified — third-party yes, independent and reproducible no).

GPT-5.6 Sol is named throughout and still has no page here; it appears only inside source pages. Recorded as missing: the task set, the rubric, "marked valid", repeat runs, and Opus 5 / Sol latency.

[2026-09-04] ingest | Message Batches API (Anthropic platform docs)

Routed from the hub (Telegram). Third page in the provider-API-access thread beside claude-api-rate-limits and claude-refusals-and-fallback. T1 vendor doc, no stated publication date, so published: is left off and every number is an ingest-date snapshot.

The market fact is the 50% discount on all batch usage, and the two seams around it: batch traffic has its own rate limits and doesn't spend the Messages-API budget, while prompt caching inside a batch is best-effort only because requests run in any order. So one lever buys throughput as well as price, and two levers that both appear on the llm-api-pricing list interfere rather than add.

Also recorded: 100k requests / 256 MB per batch, a 24-hour expiry (unbilled), 29-day result retention, workspace scoping, spend limits that batches can overshoot, stream/speed/max_tokens: 0 rejected, and the retired-Opus rows that price Opus 4.1 batched above Fable 5.1 batched.

New: claude-message-batches-api. Updated: llm-api-pricing (batch lever rewritten), synthesis.md (new section under the rationing thread), index.md.

[2026-09-05] ingest | Willison's Astra pelican grid — the first outside contact, and the price

Telegram arrival, as a share.google wrapper; resolved to the canonical simonwillison.net/2026/Sep/4/astra-pelicans/ before recording, per the strip-the-wrapper rule that fbclid links taught us.

Dedup first, and it changed the shape of the ingest: astra already exists, written 2026-08-03 off the Ten advances in mathematics announcement. So this is a new source page (astra-pelican-grid, T2 — the tier the spoke gives Willison's other link posts) plus a refresh in place of the Thing page, not a second Astra page.

Two of that page's three open questions are answered. Astra is on the API as of 2026-09-04 at $10 per million input, $50 output, against Sol's $5/$30. The page had said in plain words that "Astra's own pricing does not exist publicly" — corrected in place with the date, original struck, because the sentence was true when written.

That also re-prices the announcement's one quantity. OpenAI's ~$2,000 was token volume expressed at Sol rates; at Astra's own card the same volume is 1.67× to 2× more depending on the input/output split, so $3,300–$4,000 (our arithmetic, bounds given because the split is unpublished). The "denominated in another model's currency" tell this spoke recorded is now retired.

Artifact facts worth having: reasoning levels low / medium / high / xhigh / max, and no reasoning=none, where the GPT-5.6 line supports it. The floor of the model is a reasoning model.

The cost-per-result line is the one to keep. "Astra low produces a better pelican than ANY of the GPT-5.6 Sol models at any level, for 9.55 cents." Per-task cost printed beside per-task quality is what llm-api-pricing keeps asking for and almost never gets.

Recorded as a question, because he asks it as one: Astra and Luna both consumed 16 input tokens on the same prompt where Sol and Terra used 26, and Willison wonders whether the two are more related than OpenAI let on. One prompt, two numbers, consistent with a shared tokenizer and with several other stories. It goes to astra's "where it sits against GPT-5.6" question as a weak signal, not as an answer.

What kind of source this is, stated on the page. One prompt, one sample per cell, no repetition, no rubric, no blinding, judged by eye by the person who ran it — not a benchmark, and he does not present it as one. T2 rather than T4 for three properties the vendor tables lack: four models the author did not build, the cost of each cell published, and the weak results shown. Also flagged: the pelican tag now holds 141 posts, so the prompt's value as an unrehearsed probe is not permanent, and nothing from outside can check whether that matters.

One loose end left unresolved on purpose: the post links "the transcript that created the GPT-6 Nova pelicans" while the title, body and tag all say Astra. Most likely a slip; recorded as an inconsistency in the source rather than tidied away, because a second GPT-6 variant name would matter.

Entity discovery: none new — simon-willison and openai both already canonical here.

Updated: astra (shipped/priced, the struck pricing sentence, three Open items rewritten and a fourth added), synthesis.md (the checkable-artifact open question — the model is now purchasable and still unmeasured), index.md.

[2026-09-06] ingest | uncensored-ai-rekhi — re-seen source, two citations chased

The curator re-sent the Medium essay two days after its first ingest. No new page: the source summary was refreshed in place, per the idempotent-ingest rule. Medium returned 403 to both WebFetch and a browser-UA curl; the text came through the r.jina.ai reader. Re-read in full against the existing page — nothing in it was wrong, and nothing new in the essay itself needed recording.

What changed is the sourcing. The page's stated weakness was that every figure in it is quoted from a report this wiki had not read, one of them attributed only as "a study published in Nature Communications a few weeks ago." Two of those are now identified:

• Hagendorff et al., Large reasoning models are autonomous jailbreak agents, Nature Communications (s41467-026-69010-1, preprint arXiv:2508.04039). The 97.14% is the aggregate across four attacker models × nine targets; underneath it the spread runs 90% (DeepSeek-R1) to 12.86% (Qwen3), which the essay does not mention.
• Microsoft Security Blog, A one-prompt attack that breaks LLM safety alignment, 2026-02-09. GRP-Obliteration weaponizes GRPO — the alignment training method — in reverse. Also carries the comparison the essay drops: 81% GRP-Obliteration vs 69% abliteration vs 58% TwinBreak, utility within a few percent of the aligned base.

That last figure went onto abliteration as the first number in this corpus grading the technique against an alternative from outside the abliteration scene — with the caveat that it measures attack success, was run by the party proposing the winner, and still leaves the independent-capability-eval gap open.

Both papers remain unread here (abstract and secondary reporting only) and are the obvious research-pass targets for this spoke.

Pages: refreshed uncensored-ai-rekhi (new section: two citations chased), abliteration (the 81/69/58 comparison). No new pages, no index change.

[2026-09-10] ingest | Inception Platform docs — Mercury diffusion LLMs

Routed in from the hub (Telegram). The docs site is Mintlify and publishes an llms.txt index plus a .md twin of every page, so the models, streaming, reasoning-efforts, rate-limit and FAQ pages were read as markdown rather than scraped HTML.

Source: inception-platform-docs, T1 — first-party product documentation, the primary record of what is sold at what price. freshness: volatile (a live price list under a promotion); no published:, since the docs carry no dates anywhere.

New: inception-labs, the provider node — thin by necessity, because the docs name no founders, funding, headcount or history, and nothing else was verified at ingest (no Wikipedia or Wikidata identifier claimed; only the company site).

What is recorded: three models with full price/context/endpoint rows; the 80%-off promotion kept separate from the list price ($0.20/$0.75 list, $0.04/$0.15 promotional); cached input at a flat 10× discount; block-wise streaming and the "diffusing" mode that streams intermediate noisy states; four reasoning_effort levels ordered by latency only; FIM and Next-Edit endpoints; rate-limit tiers; the one-time 100M free tokens; text-only, no vision.

What is not recorded, and it is the important half: no benchmark, no tokens-per-second figure, no latency measurement, no comparison against any named model. The product's entire claim is speed and the documentation contains no evidence for it. Added as a growth edge — a third-party latency or quality evaluation of Mercury against a named autoregressive model.

Synthesis: new section. First commercial non-autoregressive text provider in the corpus, entering at the two workloads where autoregressive decode is felt by a waiting human (inline completion, voice). Also the mirror of the generative-media sub-theme: diffusion moving out of media into text, from a company whose models cannot see an image.

Cross-spoke seam noted, not split: the mechanism (denoising a span instead of an autoregressive loop, block-wise decode) belongs to ../llm-inference-wiki, which has no page on diffusion decoding at all. Flagged on the source page; the runner-up is logged at the hub.

[2026-09-10] ingest | STUDIO 1939 — style LoRA for MiniMax H3

Routed in from the hub (Telegram). Read the raw README.md and the Hugging Face model API rather than the rendered page, which gives the frontmatter, the download/like counts and a real modification date.

Source: studio-1939-lora, T3 — a creator's model card. Downloads (4,320) and likes (69) are the platform's and reliable; every quality and throughput claim is the author's, with no comparison and nothing independently run. published: 2026-08-25 from the API's lastModified, which is the only firm date on it. The card also promises faster endpoints "this Wednesday" with no date attached, and never names the 1939 feature it was trained on, so the public-domain claim is unverifiable here.

New: minimax-h3, thin by construction — reached only through an adapter's card, so the page records that the model serves LoRAs and ships under a community licence, and states that its capability, price and architecture are unread.

Why it earned a synthesis entry: the generative-media sub-theme had a hosted tier (Google) and open weights (Microsoft). This is the third layer — a third party's adapter over a third party's model, served by a fourth. That is the open-weight argument appearing in video, and it moves style from a provider's property to a thing anyone can train over one film's frames.

It also brings the sub-theme's first workflow numbers instead of model numbers (261 shots in 28 minutes; 108 in 21) and one checkable craft claim: character identity across scenes from a frozen, verbatim-repeated description, with no reference image.

The spin-out trigger has fired and was not acted on. This section has said since 2026-07-22 that ≥3 image/video-gen sources make a dedicated generative-media spoke the candidate. It is now at 3, and the three cover three different layers. Creating that spoke moves pages, so it is recorded for the curator rather than done here.

[2026-09-16] ingest | Introducing System One Models & Jev (TypeSafe AI)

Source: system-one-models-and-jev, T3, published: 2026-09-15, via Telegram (msg 2123), routed by the hub. New entity typesafe-ai (Organization, thin; no verified same-as found). Cross-wiki links to psychology-wiki's daniel-kahneman / heuristics-and-biases for the name. Synthesis section A second non-autoregressive pitch, with a self-graded number. Runner-up spoke: llm-inference (parallel sampler).

[2026-09-16] ingest | Meridian (Viggle model card)

Source: meridian, T3, via Telegram (msg 2129), routed by the hub. New entities viggle-ai (thin) and, from the chess route, liquid-ai (thin). Generative-media cluster now 4 non-Google sources; spin-out trigger stands, flagged to the curator again. Synthesis section The generative-media cluster reaches four.

[2026-09-19] ingest | Jev on Vercel AI Gateway — the typed-decision model gets a distributor

Routed from the hub: https://vercel.com/kb/guide/typesafe-jev-and-ai-sdk (Ben Sabic, Vercel Knowledge Base, 2026-09-19). Arrived by Telegram the same day it was published. Second source on Jev in four days, and the first not written by typesafe-ai.

New page: jev-ai-sdk-guide, T2 — first-party for the integration, second-party for the model. Vercel resells Jev and repeats the vendor's price, so it is not an independent check and growth edge 1 is untouched; what earns T2 over the launch post's T3 is a documented shipping API with limits, plus two stated limitations the vendor's own post omitted.

Refreshed in place: system-one-models-and-jev — one open item closed (availability and price after early access: GA on AI Gateway as typesafe-ai/jev, $0.042/1M input with free output held through the transition) and one advanced (what "calibrated" means: agreement between predicted probabilities and observed outcomes across many examples, not a guarantee about any one answer — still no curve, still no Brier score). typesafe-ai expanded from a thin node to two sources, while staying thin about the company — no funding, headcount, location or second model.

What the guide documents that nobody had: 64,000-token request context with 32,000 for state; choice questions up to 255 options; score rubrics of 2–10 levels returning a probability-weighted mean; Zero Data Retention and No Training per request; and a confidence statistic in providerMetadata.typesafe.confidence — how concentrated the distribution is, 0 to 1, distinct from the selected option's probability and not returned for booleans.

The finding. Vercel writes the sentence TypeSafe didn't: "Your schema constrains those answers, but doesn't guarantee they're correct." This spoke had read "can't hallucinate" as a true claim about a narrow object. It is — and the complement is that a typed decision model trades invented text for a confidently wrong typed value, which is harder to catch because it arrives already parsed. Folded into synthesis as "The typed-decision model gets a distributor, and a caveat."

Also worth watching: the AI SDK treats probability distributions as optional across providers, so experimental_evaluate is a category interface with one member in it.

No entity pages. Ben Sabic is a byline with no evidence attached; Vercel is already paged in ../cloud-wiki as vercel and is linked cross-wiki rather than duplicated (../ENTITIES.md canonicalisation). The AI SDK is deliberately not paged here — a framework for building LLM applications is ../agentic-tooling-wiki's subject; this page keeps only the model-access substance.

[2026-09-23] ingest | Claude Opus 5.5, GPT-6 Sol, GPT-6 Luna, and a new price war (Willison)

Routed from the hub (Telegram). opus-and-sol-and-luna-willison (T2). Every price checked against the vendors' pages the same day: anthropic-model-docs re-read (Opus 5.5 row), new openai-api-pricing-docs (T1). Two of Willison's remarks do not match OpenAI's page (Terra output $12 vs Sol $10; "scheduled November increase" is a promotional floor) — recorded, not repeated. New Thing pages claude-opus-5-5, gpt-6. astra's "Sol $5/$30" annotated; llm-api-pricing gains a September repricing section; thesis point 1 updated in synthesis.

[2026-09-23] ingest | Introducing Claude Opus 5.5 (Anthropic)

Routed from the hub (Telegram). claude-opus-5-5-announcement (T1 for Anthropic's own decisions; all benchmarks first-party or vendor-chosen). Read from the full page text, not a summary. Prices agree with anthropic-model-docs. claude-opus-5-5 refreshed: the "nothing known beyond price" section is replaced by the launch claims, and freshness set to volatile to match the other model pages. New synthesis section "A vendor concedes the margin, and sells the middle setting": Anthropic says its own table overstates the gap to Fable 5.1, and medium effort matches max on FrontierCode, a weak second instance of the effort result in coderabbit-fable-5-1-review (annotated there). Growth edge 1 is not advanced: nothing here is independent. No entity pages: the 21 quoted companies are testimonials, not evidence about the companies.

[2026-09-27] ingest | AxiomicLabs/Tiny_Theory_of_Mind (Hugging Face dataset)

Routed from the hub (Telegram); runner-up ../machine-learning-wiki. New source page tiny-theory-of-mind (T1 — primary data about the org's own release, read from the dataset card, JSON schema and API metadata). New thin Organization page axiomic-labs (the publisher; no Wikipedia/Wikidata entry found, so no same-as:). No Person pages: the three credited contributors are bylines with no independent evidence. published: 2026-09-23 set from the dataset's own createdAt. Synthesis gains "A benchmark for models nobody else here tracks": the benchmark scores sub-150M-parameter base models by log-likelihood, not instruction-following, and none of its 36 baselines is a model this wiki otherwise pages, so it doesn't advance an open question — noted, not forced into one. index.md updated (Organization + Dataset sections).

[2026-09-27] quality | quality cycle

Re-checked kimi-k3. glm-52 regraded stable. Drain: astra holds. Scorecard in the hub quality-log.md.

[2026-09-27] ingest | Yes, Claude can do Nine Loops (Anthropic, guest post by von Hippel)

Routed from the hub. New source page claude-nine-loop-amplitude-von-hippel (T3: a paid guest post on Anthropic's own site, with an addendum by the outside physicist who validated the result and who was paid in Claude usage credits for doing it — both facts recorded on the page rather than smoothed over). published: 2026-09-25 from the page's own date. Fable 5.1 (Claude Science harness) computed the nine-loop planar N=4 super-Yang-Mills hexagon amplitude two independent ways (direct bootstrap; indirect form-factor), one loop past the 2023 human record (Dixon & Liu). Closes part of claude-fable-5-1's flagged gap — no prior source here covered agentic/long-horizon behaviour — with a caveat that the evidence is vendor-shaped, not independent.

New entity pages: matt-von-hippel (Person, author; same-as 4gravitons.com, verified as his own blog from the source's own text) and lance-dixon (Person, addendum author/validator; same-as Wikipedia + Stanford profile, both verified). No pages for Liam Fitzpatrick / Siddharth Mishra-Sharma (the Anthropic researchers who ran it) or Song He's CAS group (concurrent human-led match using GPT-6): named in passing, no further connective evidence. Entity budget: 2 of 8 used.

Updated claude-fable-5-1 (new paragraph + updated: bump) and anthropic-scientist-access-expansion (the Claude Science mention now has a worked example of the credit mechanism; updated: bump) rather than leaving the new source dangling. synthesis.md gains "The first long-horizon evidence, and it arrives vendor-shaped" plus two cross-spoke-adjacency notes: ../agentic-tooling-wiki for the Claude Science harness itself (runner-up, not paged here — this spoke keeps only what the source says about the model), and ../mathematics-wiki as the routing runner-up for the physics object itself, since no physics spoke exists in this hub. index.md updated (source + two Person entries). avoid-ai-writing pass run in detect mode: fixed a bold-label/period bullet shape in the source page and a doubled "genuinely" + missing article in the synthesis addition.

[2026-10-03] lint | quality cycle refresh of amazon-bedrock

amazon-bedrock re-checked against the homepage: the 100,000+ figure holds; "Managed Agents, powered by OpenAI" and AgentCore added as new, vendor percentages recorded as claims.

[2026-10-06] ingest | Interfaze 1 Lite (Hugging Face model card, URL-only, T3)

Model: claude-sonnet-5-5. Routed here by the hub; no runner-up spokes.

interfaze-1-lite is an Apache-2.0 "mixture-of-architectures" model: a Qwen-based reasoning core that calls Chandra OCR, Whisper+pyannote, SAM 2.1, TimesFM, Llama Guard and PaddleOCR as tools. Scores (GPQA Diamond 85.9, olmOCR-Bench 83.8) are self-reported, so T3. One 80GB GPU, offline; Transformers or a hosted OpenAI-compatible API. No published: date on the card, so none recorded. No org page: nothing known about interfaze-ai beyond the card. Folded into synthesis under "The architecture has stopped corresponding to a lab" as a fifth point; index updated.

[2026-10-07] ingest | LFM2 Technical Report (arXiv 2511.23404, T1) — via research pass

Hunted via the Parallel-Search-MCP web_search/web_fetch tools (new MCP, evaluated this cycle) against hub quality-log.md edge #51, llm-providers-wiki, promoted from the dream ledger same day. Closes the liquid-ai half of synthesis growth edge 11.

lfm2-technical-report: Liquid AI's own 31-author writeup of LFM2 — hybrid gated-short-convolution/grouped-query-attention backbone found by hardware-in-the-loop search, 350M–8.3B params (dense + one MoE variant), 10–12T pretraining tokens, open weights. Self-reported benchmarks (IFEval 79.56%, GSM8K 82.41% on the 2.6B) carry the usual first-party caveat. liquid-ai updated from thin-node to a real org page; two new Person pages, ramin-hasani (CEO) and daniela-rus (co-founder, MIT CSAIL director), both same-as verified against Wikipedia/MIT CSAIL/liquid.ai. maxime-labonne gains a third appearance (named author, first stated affiliation). index.md updated.

[2026-10-07] lint | heretic.md freshness re-check (Parallel-Search-MCP eval, second test case)

Re-checked via Parallel-Search-MCP web_search (user asked specifically: "Parallel-Search-MCP heretic models"). Stars 27,968→32,482, forks 3,032→3,652, latest release v1.4.0 (14 Jun 2026, Gemma 4 support). HF ?other=heretic tag count now 5,914 (was "well over 5,000"). No method change in the newer release notes. updated: bumped; no tier change (still T1 for the tool, T3 for its self-run benchmark table).