Spokes.wiki Search About

Synthesis — LLM Providers

The evolving thesis. Spun out of the hub _inbox llm-providers cluster on 2026-06-01, seeded by deepseek-api-docs and grown with three landscape sources the router went and found (open-weight roundup, API-pricing comparison, the llm-stats leaderboard).

Current thesis

The 2026 model market is defined by a collapsing price floor meeting a still-premium frontier, with open weights as the force pushing the two apart.

  1. A ~600× cost spread (llm-api-pricing): budget APIs at ~$0.10/1M input vs frontier reasoning at $30/$180. deepseek disrupted the bottom — capable open weights plus an OpenAI-compatible ultra-cheap API — dragging the floor down and forcing proprietary labs to justify their premium. On 2026-09-22 the proprietary labs cut too (opus-and-sol-and-luna-willison): GPT-6 Sol $2/$10 and Luna $0.10/$0.50 halve their GPT-5.6 predecessors, and Opus 5.5 takes 20% off Opus 5. The lists now match across labs at three points ($2/$10, $4/$20, $10/$50), so the premium the labs defend is no longer visible in price.
  2. Open-weight has caught up enough for production (open-weight-models): Llama 4, Qwen3, gemma-4, DeepSeek, Kimi, GLM. Two structural enablers — sparse MoE (frontier capability at practical inference cost) and Apache-2.0/MIT licensing clarity — plus a context explosion (256K → 1M → 10M tokens), and now multimodality at the local tier: gemma-4 12B runs encoder-free vision + native audio on 16GB (gemma-4-12b-announcement), making memory footprint a competitive axis alongside capability/cost/context. quantization is the second footprint lever (after sparse MoE): Gemma 4's QAT checkpoints push the floor to a sub-1GB E2B and a 2-bit mobile scheme, claimed above PTQ quality gemma-4-qat — footprint is now contested by training technique, not just architecture. ternary-weights takes that one step further (2026-08-02): neutrino-1 claims an 8B-class model at 2.56 GB download / 3.88 GB on disk with weights trained in a ternary format rather than quantized into one, so there is no float model anywhere in its lineage. If it holds, footprint stops being a treatment applied to a model and becomes a property of it — but every number supporting it is first-party and the method is unpublished. A fourth lever appeared on 2026-08-26: Qwen3.8-Flash-Next carries 51B N-gram embedding parameters that are addressed deterministically and can live in host memory, on top of a 125B-total / 6B-active MoE backbone. Sparse MoE separated total from active parameters; this separates stored capacity from the per-token matrix-multiply entirely, and Qwen prices the result at $0.15/$0.47 per 1M.
  3. "Best model" is multi-axis (llm-benchmarks): composite scores over capability × speed × price × context. Proprietary labs (anthropic, OpenAI) still lead reasoning; open/Chinese labs (Qwen, DeepSeek) lead cost-per-quality; xAI leads context — and with Grok 4.5 now also contests cost-per-quality from inside the closed frontier (grok-4-5-price-vs-benchmarks). A sharper cost axis emerges from that source: cost-per-task = per-token price × tokens-per-task, so a token-efficient model can be cheaper in practice even at a higher sticker rate (and vice-versa — Fable 5 is both pricey per token, $10/$50, and no cheaper per task).

The unifying tension: capability still concentrates at a few proprietary frontier labs, while cost and access are being democratized from below by open weights. Whether the frontier premium holds depends on whether open-weight reasoning closes the gap — the live question of the wiki.

Open weights have a third face (2026-07-26): the behaviour is editable too. The wiki has read open weights as licensing (who may run it) and footprint (who can afford to). abliteration adds a lever the license grants without naming: gemma-4-31b-it-scotoma is Gemma 4 31B-it with its refusal behaviour surgically weakened in layers 7–41 and republished under Google's own Apache-2.0. Alignment shipped as weights is alignment shipped as something a third party can rewrite in an afternoon and redistribute on the same terms. The contrast with the closed tier is exact: a Claude refusal arrives as an HTTP 200 you can route around but never delete (claude-refusals-and-fallback). And the instrument doing the cutting is Anthropic's own Jacobian lens — an interpretability tool built for auditing models, pointed at modifying one.

The provider map now sorts into five positions plus a reseller layer. The market is no longer a simple open-vs-closed binary; the players cluster by strategy:

Llama also sharpens the licensing thread: open-weight but not OSI-open (custom community license, OSI-disputed), against the Apache-2.0 cleanliness of gemma-4/mistral-ai/qwen — "open" is a spectrum, not a binary. Across the top sits a cloud-reseller layer: amazon-bedrock is not a model maker but a one-API aggregator of Claude/Llama/Mistral/Cohere/Nova — the demand-side mirror of the map, where model choice becomes a config parameter and differentiation moves up to routing/caching/data. Through a reseller, open and closed models are the same API call; the buyer meets the whole price/openness spectrum through one pane.

That layer has a second, self-hosted shape: the ai-gateway you run yourself. omniroute (MIT, ~30K stars) fronts a claimed 290+ providers from a process on your laptop, cascading from subscription seats down through cheap providers to free ones as each quota dies, and it pushes the demand-side logic past routing into two levers the reseller never offered — client-side prompt compression (12 stacked engines, 15–95% claimed) and free-tier aggregation (~1.53B tokens/month pooled across 43 provider free tiers). Both are unilateral: neither needs the provider's agreement, and the second arguably runs against its interest, which is why the project itself flags 15 of those pools as ToS-questionable. The reseller commoditized model choice; the local router commoditizes the provider's generosity.

Recurring reads

Generative-media sub-theme (watch for spin-out)

The gemini line now has sourced generative-media tiers with their own pricing units, not just $/token: Nano Banana 2 Lite (text-to-image, $0.034/1k images, ~4 s) and Gemini Omni Flash (video, $0.10/sec, Veo-3.1-Fast parity, 10 s max). This is a distinct modality from the text/multimodal LLMs at the spoke's core — and a parallel to how audio-gen was carved out to ../speech-audio-wiki. For now it stays here as part of the Gemini family; if non-Google image/video-gen sources accumulate (≥3), a dedicated generative-media spoke is the spin-out candidate. Logged so the cluster is visible rather than buried.

Source 2 — and the first non-Google one (2026-07-22). mage-flow-turbo is Microsoft's MIT-licensed 4B text-to-image model (rectified flow matching, 4-step distilled, native 512–2048px), claiming 0.59 s per 1024² image on an A100 and parity-or-better against Qwen-Image 20B / Z-Image 6B / FLUX.2 32B. It moves the sub-theme in two ways. It splits the modality along the same open-vs-closed axis the text market runs on — Google's image tier is priced per image, this one is weights you download, so the generative-media cluster now has both a hosted-API instance and an open-weight one, and only the first has a price. And its argument is the spoke's own recurring one, one modality over: 4B beating 20–32B is agents-a1's "not by scale" claim in image form, bought with tokenizer–backbone co-design and step distillation instead of agent horizon. Same T3 caveat as A1: first-party card, first-party comparison set, no independent re-run — and the corpus still has no artificial-analysis equivalent for image models, which is the missing referee for this whole sub-theme. Cluster now at 2 of 3.

Source 3, and the trigger is met (2026-09-10). studio-1939-lora is a third-party style LoRA for a video model — two ranks over minimax-h3, trained on frames from one public-domain 1939 animated feature, served through fal's standard LoRA endpoints. It is the first generative-media source here that is neither a provider's hosted tier nor a lab's open weights, but somebody else's modification of somebody else's model, which is the open-weight argument finally showing up one modality over. It also brings the sub-theme's first workflow numbers rather than model numbers: 261 shots in 28 minutes for one 8-minute film, 108 in 21 for another, and character identity held across scenes by repeating a frozen description verbatim rather than by conditioning on a reference image. T3 — a creator's card, no comparison, nothing independently run.

The cluster is now at 3 non-Google-plus-Google sources across image and video, which is the spin-out condition this section wrote for itself. Recording it rather than acting on it: creating a generative-media spoke moves pages and is a structural change, and the last several spin-outs were the curator's call. What is now true is that the trigger fired, and that the three sources cover three different layers (hosted tier, open weights, third-party adapter), which is a healthier founding corpus than three of the same kind would be. Still missing, and it was the gap named at source 2: no artificial-analysis equivalent for image or video models — the referee this whole sub-theme lacks.

Open questions

Growth edges

Ranked; each names the kind of source that would close it (see ../QUALITY.md → Growth edges).

  1. An independent re-run of any vendor benchmark claim. Every model page here rests on numbers the vendor produced and nobody outside can reproduce; it is the standing complaint behind the spoke's benchmark tensions. — needs: a T1/T2 third-party evaluation of a model this wiki already pages. Partly advanced 2026-09-01 and re-specified, not closed. coderabbit-fable-5-1-review is a T2 evaluation by someone other than the labs, covering four paged models on one task — so the third-party half is met. The independent and reproducible halves are not: the harness belongs to a company selling the filter its own precision figures argue for, and neither the 45 tasks nor the scoring rubric is published, so nobody can re-run it either. It also does not re-run a vendor claim; it measures a different job. — needs now: an evaluation whose task set and rubric are published, by a party with nothing to sell downstream of the result.
  2. Whether the verifiable-artifact form spreads. astra shipped machine-checkable Lean proofs with a capability claim — falsifiable output, unfalsifiable capability. — needs: a second lab doing the same, or an analysis of the form.
  3. Free-tier economics. How long a pooled free tier survives is asked and unsourced. — needs: a first-party pricing/policy change with stated reasoning (T3 acceptable — it is a first-party fact). Partly advanced 2026-08-29: anthropic-scientist-access-expansion is a first-party access policy with stated reasoning, but it gives away seats to a screened class, not an open pool — the opposite design to the one being asked about here. It bounds the giveaway by identity instead of by quota, which is one answer to the pooling problem and worth watching for imitators.

Coverage edges (added 2026-08-08, at the curator's request for a wider backlog). These widen what the spoke covers instead of answering an open question above; one ordinary solid source closes any.

  1. The terms nobody reads. Data retention, whether prompts train the model, and zero-retention tiers decide which provider an enterprise may legally use, and no page scores that dimension — llm-api-pricing prices tokens only. — needs: the providers' own terms and enterprise addenda (T3, first-party facts), read side by side.
  2. Batch and off-peak pricing. Most providers sell asynchronous batch work at roughly half price, which changes the cost comparisons this spoke keeps making. — needs: first-party batch pricing and its stated turnaround guarantees.
  3. Embeddings and rerankers. A whole product line with its own pricing and leaderboards, absent from a spoke that otherwise tracks every model release. — needs: provider docs plus MTEB.
  4. The GPU-cloud hosts serving open weights have no nodes. The concept is covered — ai-gateway splits managed resellers from self-hosted routers, and amazon-bedrock and omniroute are paged. What is missing is the class beneath it: awesome-gemma listed 14 hosted providers for one open model, and of Cerebras, Together AI, Fireworks, Baseten, Runpod and Cloudflare Workers AI, none has a page anywhere in the corpus (Modal is paged, but in ../platform-ops-wiki as compute, not as a model host). These are where an open-weight model actually gets bought per-token, so the spoke's pricing comparisons currently miss the market that open-weight-models creates. OpenRouter is named in ai-gateway but has no node of its own. — needs: their own per-token pricing pages (T3, first-party facts), or a third-party comparison of open-model hosting rates against direct frontier APIs.
  5. Availability, not just capability. No page holds an outage record, an SLA, or a status-page history, so a model can be scored here without anyone asking whether it stays up. — needs: published status-page history or a postmortem.
  6. Does any other provider run a research access program? research-access-programs rests on a single provider's announcement, so the spoke has a concept and no market. OpenAI, Google, Mistral and the open-weight labs all court academics and none is paged. — needs: two or three providers' own academic/nonprofit program pages, read side by side.
  7. Claude Science has no page. A named Anthropic research product, launched June 2026, appears in this wiki only as one sentence quoted inside the scientist-access post. — needs: its own product page or launch post (T1, first-party).
  1. Entities known only through the page that found them. New 2026-10-07, promoted from the hub dream ledger. HALF-CLOSED the same day (research pass) by lfm2-technical-report (T1, arXiv 2511.23404, 31 Liquid AI authors): liquid-ai now has a first-party account of the LFM line itself — architecture, parameter range, training data volume, self-reported benchmarks — not just the one chess fine-tune in machine-learning-wiki that found it. Game-engines' Michal Valient is untouched; he is a different spoke's entity and still known only through his company's own biography. — needs, for the Valient half: any independent source describing him directly.

The tier ladder doesn't order everything

The spoke's working model of Anthropic's line is a price/capability ladder: claude-sonnet-5 mid, claude-opus-5 flagship-by-price, claude-fable-5 premium, Mythos a domain specialist. Every comparison held here — the benchmark ratios, the cost-per-task framing, the $10/$50-vs-$5/$25 split — sorts on that one axis.

Anthropic's own documentation shows one dimension running the other way. Per anthropic-model-docs, Opus 5's reliable knowledge cutoff is May 2026; Fable 5's and Sonnet 5's are January 2026. The tier sold as below the frontier carries four months more world knowledge than the tier above it.

That is not a contradiction — nothing claims Fable 5 is newer — but it is a hole in how this spoke has been reasoning. Recency is a real axis of model quality for anything touching current events, current prices, or recent releases, and no source in the corpus prices it. A buyer choosing Fable 5 for "the best model" gets a better reasoner with an older world. Watch whether the next Anthropic release keeps the flagship-lags-by-recency shape or closes it.

It also sharpens the standing caution about where these facts come from. The knowledge cutoff appears in the docs and in no launch post — same place the Opus 5 effort ladder turned out to be mis-stated (see anthropic-model-docs). Launch posts are the right source for positioning and the wrong one for anything a buyer or developer would act on; this spoke had been building almost entirely from them.

The premium tier is also the rationed one (2026-08-30, quality cycle). Re-reading claude-api-rate-limits against Anthropic's docs — the page had gone 63 days stale — turned up a fact the spoke's price comparisons cannot see: Fable 5 gets a quarter of Opus 5's input-tokens-per-minute at every tier (500K vs 2M on Start, 1.5M vs 5M on Build, 4M vs 10M on Scale), and half the RPM above Start. Opus 5 and Sonnet 5 also now carry buckets separate from the 4.x pools.

So the ladder has a third dimension beside price and capability: throughput you are permitted. The model Anthropic places above its own flagship on intelligence, and charges double for, is the one an organization can push least traffic through. That is the same shape as Cerebras's sold-out coding tiers — where the spoke recorded that fast inference is rationed, not just priced — arriving now from a frontier lab and about its top model rather than a speed reseller. Two independent instances make it a pattern worth naming: at the frontier, capacity is the constraint the rate card does not show.

And for some buyers the ladder has no upper half at all (2026-08-27). In anthropic-scientist-access-expansion Anthropic states that researchers in biology and chemistry "will still be limited to our Opus-class models," that Fable blocks professional biology and drug-development queries on dual-use grounds, and that Mythos-class access for life-sciences professionals runs through a program built in partnership with the US government, whose first participants are now enrolled. The ladder is therefore cut horizontally by field: the rungs a buyer can reach depend on what they work on, and the buyers with the strongest case for frontier capability in biology are the ones held furthest from it. Nothing in this spoke's price comparisons can see that, because a per-token table has no column for who is allowed to pay. See research-access-programs.

The third dimension has a discount attached (2026-09-04). claude-message-batches-api completes the throughput picture the two sections above sketch. Batch traffic is billed at half list and runs on its own rate limits — the docs say using it "does not affect rate limits in the Messages API" — so a buyer who can wait gets both the cheaper token and capacity that was never in the real-time bucket. Set against the rationing finding, that reads as a price on immediacy: the same model, the same tokens, at half the money, provided you surrender the guarantee of when. What Anthropic actually sells at the top of the ladder is not the model, it is synchronous access to it.

The levers also interfere. Prompt caching — this spoke's other double-acting lever — works inside a batch only best-effort, because batched requests run concurrently and in any order; Anthropic's own suggested fix is to buy the 1-hour cache duration, a paid workaround for a discount. So the "stack the levers to ~25% of list" arithmetic on llm-api-pricing is optimistic where two levers touch, and the spoke should stop treating the discounts as additive.

A provider that is not autoregressive (2026-09-10)

inception-platform-docs puts inception-labs's Mercury models in the corpus: diffusion LLMs sold for chat, reasoning, tool calling and code, behind an OpenAI-compatible API. Every other text provider in this spoke ships a transformer decoder that emits one token at a time. This one denoises a span in parallel, and the difference reaches the API surface — streaming arrives block-wise, in "chunks of refined text," and an optional mode streams the intermediate noisy states so a client can watch the answer resolve.

The competitive claim is latency, and the spoke cannot check it. The docs say "fastest" and "frontier-quality" and publish no benchmark, no tokens-per-second figure and no comparison against any named model. llm-benchmarks holds no Mercury result; neither does anything else here. So the spoke's multi-axis "best model" frame has a provider on it whose entire pitch sits on the one axis with no evidence attached. That is a growth edge, not a finding.

Where the price lands depends on which price you read. List is $0.20 input / $0.75 output per 1M — budget tier, beside Mistral Small and GPT-4.1 Nano. An 80%-off banner runs across the docs, taking Mercury 2.5 to $0.04/$0.15, below anything in llm-api-pricing. A promotional floor is not a floor, and the page records both numbers for that reason. Cached input is a flat 10× discount on all three models, which is more legible than most published caching rules.

The chosen market is narrow and well-picked. Not capability rankings: inline code completion (a FIM endpoint and a Next-Edit endpoint, wired into Cursor, Zed and VS Code) and realtime voice (ElevenLabs and Vapi integrations, an instant reasoning level). Both are workloads where autoregressive decode is felt directly by a human waiting. A new architecture entering at the two points where the incumbent architecture's cost is most visible is a coherent strategy even without a published number behind it.

And the architecture cuts the other way too. Mercury is text-only — no vision, no image input, stated in the FAQ — from a company whose method comes from image generation. The generative-media sub-theme this spoke watches (mage-flow-turbo) is diffusion moving into media; this is diffusion moving out of it, into the workload transformers own.

A second non-autoregressive pitch, with a self-graded number (2026-09-16)

system-one-models-and-jev puts typesafe-ai beside inception-labs as the spoke's second provider selling not a transformer decoder on the latency axis. The difference is the shape of the output: Mercury still writes text, Jev returns a typed value with a probability — a yes/no, one of up to 255 choices, a score — and so cannot type-error or hallucinate in the only sense that applies to it, because it cannot emit prose. That makes "can't hallucinate" a true statement about a narrow object, and the post's own comparisons (70–500 ms against 3–329 s; $0.042 input, free output) are measured on the four closed-set classification workflows where such a model is at home.

Unlike Mercury, there is a number: 67.8% aggregate accuracy on evals.typesafe.ai. It should be read with the grading rule in view. The reference label is an average of GPT-6 Astra and Fable 5.1 at high thinking, so 67.8% is agreement with two other vendors' models rather than with anything known to be correct, and the reference models cannot lose on their own labels. The eval was written by the vendor's capabilities team, which the post says. So the spoke's standing edge 1 — nobody re-runs anybody's numbers — applies here in the cleanest form it has taken: a public harness, no human labels, and a headline ratio the vendor concedes is the high end. What is new is not the speed claim but the category: a model whose API surface is a typed function signature, which is the structured-outputs argument carried to its end.

The generative-media cluster reaches four (2026-09-16)

meridian is a fourth non-Google image/video source and a third page on the minimax-h3 base. It adds a layer the cluster did not have: not a model, not a style adapter, but a pipeline that puts an external geometry estimator in front of the diffusion transformer so the camera can move through a clip that was filmed from one place. The licensing is the finding for this spoke. The weights are downloadable and the code is Apache, yet the base licence excludes the EU, UK, South Korea and the US, and the geometry model is non-commercial — so "open-weight" here describes a file, not a right to use it, in most of the markets this spoke's readers sit in. The spin-out trigger (≥3 non-Google gen-media sources) fired on 2026-09-10 and stands; the count is now four, across four layers (model, adapter, base, pipeline). Moving pages is the curator's call and is flagged again.

The typed-decision model gets a distributor, and a caveat (2026-09-19)

jev-ai-sdk-guide is the second source on Jev in four days, and the first not written by typesafe-ai. Vercel now resells it through AI Gateway as typesafe-ai/jev, which ends the waitlist and closes one of the three items system-one-models-and-jev left open: the $0.042/1M input with free output survived the move out of early access. A launch price that holds through general availability is worth one line in a spoke where most of them don't.

A reseller is not an independent check, and this one still adds something. Vercel repeats TypeSafe's price and runs no benchmark, so edge 1 is untouched — nobody has re-run the 67.8%. What it supplies instead is the product's actual surface: a 64,000-token request with 32,000 for the state, choice questions up to 255 options, score rubrics of 2 to 10 levels, Zero Data Retention per request, and a confidence statistic — distribution concentration, 0 to 1, absent for booleans — that the launch post never mentioned.

The sentence the vendor didn't write. This spoke read "can't hallucinate" as true about a narrow object: no prose, so no fabricated text. Vercel states the complement:

"Your schema constrains those answers, but doesn't guarantee they're correct."

and, on the calibration promise, that calibration describes agreement between predicted probabilities and observed outcomes across many examples and "doesn't guarantee an individual answer is correct." Neither contradicts anything TypeSafe published. Both name the trade the category makes: a typed decision model swaps invented text for a confidently wrong typed value, and the second failure is harder to see because it arrives already parsed. The distributor said it, four days after the lab didn't.

A category interface, not a driver. The guide treats the probability distribution as optional — other evaluation models "may omit them" — so experimental_evaluate is written for a class of models with one member. Whether a second System One model appears to fill it is the thing to watch, and it is the same question inception-labs poses from the other non-autoregressive direction: one provider, one architecture, no field yet.

Where the routing decision sits. The source is a Vercel how-to, and the SDK is ../agentic-tooling-wiki's subject while ../cloud-wiki owns vercel as a host. The substance this spoke took is the model, its access path and its price; the integration technique is noted on the page and not paged here. Runner-up spoke logged at the hub.

A vendor concedes the margin, and sells the middle setting (2026-09-23)

claude-opus-5-5-announcement is Anthropic's own launch post for Opus 5.5. Two things in it matter more than its table.

The vendor discounts its own benchmarks. Opus 5.5 leads Fable 5.1 on every row where both appear, by up to 10.6 points on Terminal-Bench 4.0. The post then says the real-use gap "is narrower than these scores suggest" and that benchmark margins have become "a less reliable guide to real-world differences". This spoke's standing complaint about vendor numbers (llm-benchmarks, growth edge 1) now comes from a vendor, in its own launch post. It does not close growth edge 1. Every number in the post is still Anthropic's or chosen by Anthropic.

The pitch moved from the top setting to the default. The argument is cost per task at medium effort, not the max-effort score. On FrontierCode, medium scores 54.6% and max 54.4%, so max buys nothing there. Quoted customers say the same with other workloads: Deloitte's low-effort bug catch (72%) beats Opus 5 at high (56%), and Rogo's low beats Opus 5's high with about 60% fewer output tokens. That is a second source where more effort does not buy more quality, after coderabbit-fable-5-1-review. It is weaker than the first. The FrontierCode gap is inside any plausible error bar, and the customer figures compare against the older model, not against Opus 5.5 at a higher setting. Still, the vendor now sells the middle of its own dial, and that fits the CodeRabbit result better than it fits the tier ladder's assumption that effort buys quality.

The 40% saving is mostly tokens, not price. The list price fell 20%; the other half of the claimed 40% comes from using fewer tokens per task. This matches the token-efficiency read in the thesis: a sticker price alone no longer tells you what a model costs to run.

A benchmark for models nobody else here tracks (2026-09-27)

tiny-theory-of-mind, via Telegram: axiomic-labs's synthetic theory-of-mind benchmark, scored by base-model log-likelihood rather than instruction-following. It sits outside every axis llm-benchmarks tracks — its 36 baselines are all sub-150M-parameter checkpoints, and no model this wiki otherwise pages (Claude, GPT, Gemini, Llama, Qwen) appears on its table. So it does not advance any open question here; it is filed because the routing runner-up was ../machine-learning-wiki and the dominant substance is a model+provider landscape entry (a lab, its checkpoints, its own benchmark), not a training-method question. Worth a second look only if a later ingest runs a chat-tuned model against it, which would connect it to the tier ladder for the first time.

The first long-horizon evidence, and it arrives vendor-shaped (2026-09-27)

claude-nine-loop-amplitude-von-hippel closes part of the gap claude-fable-5-1 flagged on 2026-09-01 — nothing here covered the Fable tier's agentic/long-horizon behaviour, only single-turn benchmark scores. Inside Anthropic's Claude Science harness, Fable 5.1 ran largely unsupervised for hours-to-days, writing its own Python/SymPy from scratch, and reached a nine-loop planar N=4 super-Yang-Mills amplitude by two independent methods — one loop past the best published human result (Dixon & Liu, 2023).

The evidence is a new kind for this spoke (sustained autonomy on a checkable, non-fuzzy problem, not a leaderboard score) and weak provenance by this spoke's own tier rules: Anthropic chose the problem, ran the harness, paid the writer, and paid the one outside validator in its own product's credits. Dixon's two-week independent re-derivation stands on his own methods, not Claude's — closer to a real check than a self-report, though still not a disinterested benchmark. A concurrent human-led effort (Song He's group, CAS, with GPT-6 assistance) reaching most of the same answer around the same time is the closest thing to an outside check this source has. File it as one vendor-adjacent data point on autonomy; watch for a second, non-Anthropic-adjacent instance before treating long-horizon agentic capability as measured here at all.

Contradictions / tensions

None internal yet. Cross-source tension to watch: leaderboards crown proprietary reasoning while the open-weight roundup argues practical fit beats rank — a framing disagreement, not a fact conflict. Fable-5-is-expensive, now on two reads (2026-07-08): Njenga's T4 test put Fable 5 at ~3.2× Sonnet 5 on a trivial prompt; The Decoder's rate card (grok-4-5-price-vs-benchmarks, T3) independently lists Fable 5 at $10/$50 ≈ 3.3× Sonnet's $3/$15. Two non-Anthropic sources agree it's the dear model — corroboration, not proof. Odd for a model pitched as the fast pick; likely a premium tier. Still pending a T1 Anthropic price sheet before recording as fact.

Efficiency is now the shared pitch — and tiering goes intra-family

openai-gpt56-grok45-clash shows OpenAI and xAI timing rival launches hours apart, both leading with token efficiency. Two structural reads land on the thesis. (1) Tiering moved inside the family: OpenAI's GPT-5.6 ships as Sol / Terra / Luna (flagship / everyday / speed-&-affordability) — the capability×price split the market already had across labs is now explicit within one lineup, with Luna as the cheap-fast tier. (2) Cost-per-task is the industry frame, not one vendor's spin: Musk pitches Grok 4.5 as "Opus-class but faster, more token-efficient and lower cost" — the same cost-per-task argument, in the maker's own words, as buyers watch token spend. Gemini 3.5 Pro is expected the same month. So the price-floor pressure the thesis pins on open weights now has a second driver: efficiency competition among the closed frontier labs themselves.

A new availability axis — export controls. The same source reports the Trump administration export-banned Anthropic's Fable and Mythos models (cyberattack-misuse risk); Anthropic suspended them and redeployed Fable 5 last week. For the first time in this spoke a model's availability turned on government policy, not price or licensing — a governance variable sitting above the market map (cross-spoke: ../ai-governance-wiki).

A second pricing shape — the speed provider. The market's price story so far is a per-token rate card (the ~600× spread, output 2–6× input, cost-per-task as the hidden multiplier). Cerebras breaks the shape: a wafer-scale inference provider (cerebras-inference) reselling open models on speed, it publishes no per-token price (corrected 2026-09-22 — it now publishes a two-model rate card, GPT OSS 120B $0.35/$0.75 and Qwen 3.8 27B $0.99/$1.49 per M tokens) and sells daily-token subscription buckets as well (Cerebras Code Pro $50/mo→24M tok/day, Max $200/mo→120M tok/day; $5 free credit; Developer from $10). Both coding tiers were sold out (mid-2026, still so on 2026-09-22) — capacity is rationed, so what's scarce and priced here is throughput, not model quality. This adds an axis the rate-card view misses: a provider can compete on how fast and how much per day rather than how much per token. It also sharpens the boundary with ../llm-inference-wiki — Cerebras the mechanism/company lives there (cerebras-inference, cerebras-systems); what lives here is its place in the provider market and its unusual price structure.

A third shape — speed as a paid mode on one model (2026-07-24). Opus 5 sells a fast mode at 2× the base rate for ~2.5× the default speed (claude-opus-5-announcement). Cerebras prices speed by building different hardware; OpenAI's Luna tier prices it by shipping a different model. Anthropic prices it as a runtime toggle on identical weights, which puts latency alongside effort as a per-request dial the buyer pays for directly. Three ways to sell speed now coexist — separate silicon, separate model, same model different meter — and only the third leaves the capability question untouched. Whether the 2× premium prices real serving cost or just willingness to pay is not answerable from the announcement.

The first independent capability read, and why it doesn't settle much

Almost every capability claim in this spoke is the lab's own. opus-5-arc-agi-3 is the exception: ARC Prize scores Opus 5 at 30.2% on ARC-AGI-3 against a 7.8% prior record (GPT-5.6 Sol Max), with "Fable-class" at ~20%, plus 90.4% / 97.5% on ARC-AGI-2 and -1. Three things it changes and one it doesn't.

A benchmark nobody has saturated ranks differently from the composites. The indices in llm-benchmarks score what the vendor ships and crowd into a narrow band at the top; ARC-AGI-3 scores the model without a harness on tasks it has never seen, and the field is spread from 7.8% to 30.2%. It also inverts the family ordering Anthropic sells: Fable 5 is the premium tier at 2× the price, and here the cheaper flagship beats it.

A vendor ratio that doesn't reconcile. The launch post claims "3× the next-best model on ARC-AGI 3" (claude-opus-5-announcement). Against the 7.8% record that's ~3.9×; against ARC Prize's ~20% Fable-class figure it's ~1.5×. Both claims can't describe the same comparison. Recorded as a tension, not resolved — this is the same shape as the Fable-5 pricing case, where third-party numbers arrived before the vendor's own.

Novelty benchmarks decay once labs aim at them. Opus 5 postdates ARC-AGI-3's public format, and the jump does not reproduce on Guanghan Ning's private Witness benchmark, where Opus 5 (43.4) ties Fable 5 and kimi-k3 and falls below Opus 4.8 on the one unfamiliar rule combination. ARC Prize's Greg Kamradt argues that isn't disqualifying. Either way the useful read is structural: an evaluation whose value is novelty has a shelf life measured from the day its format goes public, so private and rotating benchmarks buy resistance to targeting at the cost of reproducibility — the axis the artificial-analysis disclosure story sits at the other end of.

What it doesn't change: the open question below about the frontier premium. A 30.2% that may be partly benchmark-targeted, on a test the vendor's own harness would score higher on, is not evidence about capability-per-dollar in production.

"Model" is becoming a product boundary, not a technical one

Sakana Fugu is a multi-agent system sold as a model: OpenAI-compatible endpoint, a model string, a reasoning-effort dial, and — in the docs' own first sentence — an agent system underneath that routes across all supported providers by default, with the pool narrowed per API key rather than per request fugu-get-started. Three of this spoke's working assumptions break at once.

You can't say which model answered, so a capability benchmark measures a routing policy on the day it ran. You can't state a per-token price for a named model, because the composition is the product. And the reseller/lab distinction collapses: Sakana presents as a lab with its own model family while being, underneath, a composer of other labs' models — llm-provider's "frontier lab" and "cloud reseller" categories arriving at one endpoint.

Read it beside claude-opus-5's fast mode (a runtime toggle on identical weights, 2× price for ~2.5× speed) and the trend is the same from both directions: the SKU has stopped corresponding to a set of weights. One vendor sells two products off one model; another sells one product over many models. In both cases what the buyer purchases is a configuration, and the rate-card view this spoke was built on describes less of the market each month. The ai-gateway page carries the mechanism; the honest consequence for the map is that capability and price claims now need to name what they measured, not just which model string was passed.

Two smaller markers from the same source. Cyber is now its own SKU across three labs — Anthropic's Mythos (claude-opus-5), Google's Gemini 3.5 Flash Cyber (alphabet-q2-2026-earnings) and now fugu-cyber — with none of them publishing what makes a model a cyber model, against a backdrop where Anthropic's cyber-capable line drew an export ban (claude-fable-5). Domain specialization has quietly become a product axis beside size, speed and price, and the security one is the axis regulators already watch. And the effort dial starts at high: no low, no medium, with a Codex integration that raises the stream idle timeout to two hours. A model whose cheapest setting is "high" and whose turns run for hours is not competing in the same market as a chat completion.

The architecture has stopped corresponding to a lab

neutrino-1 states an architecture matching qwen's published Qwen3-8B/config.json on every field — 36 layers, 4,096 hidden, 32 query over 8 KV heads, 12,288 feed-forward, 151,936 vocabulary, 40,960 positions, untied embeddings, and a matching parameter count — and its release post never mentions Qwen. The 0.6B matches Qwen3-0.6B the same way.

Nothing improper follows. Configurations are not proprietary, Qwen3 is Apache-2.0, and holding the geometry fixed while changing the weight format is the right experiment when the format is what you are testing. The point for this map is different: Qwen's contribution to the open ecosystem is now a default shape as much as a set of checkpoints, and a shape leaves no attribution trail. The 200,000+ Hugging Face derivatives are countable; adopted geometries are not.

This is the erosion logged in "Model is becoming a product boundary", one layer down. There, a model string stopped identifying a set of weights. Here, a set of weights stops identifying a design lineage. The spoke holds four positions on that gradient now: agents-a1 discloses its Qwen3 lineage, readyart publishes openly-labelled derivatives of someone else's checkpoints, sakana-fugu presents other labs' models as a family of its own, and neutrino-1 sits at a fourth point — plausibly original weights on a borrowed skeleton, with the borrowing unstated.

interfaze-1-lite (2026-10-06) is a fifth point, and a literal one: an Apache-2.0 release whose reasoning core is Qwen-based and whose "model" is a bundle of other open projects (Chandra OCR, Whisper, SAM 2.1, TimesFM, Llama Guard) called as tools. Its GPQA Diamond 85.9 and olmOCR-Bench 83.8 are self-reported and describe the bundle, not any one weight set.

The practical consequence is a caveat on footprint claims, which are always comparative. "An 8B at 2.56 GB" means something different if the 8B in question is a known-good geometry someone else validated. This wiki records the match, not a verdict, and needs a second Fermion release or an independent evaluation to say more.

Whose price is on the sticker?

An open-weight model doesn't have a price — it has a vendor rate and a set of independent hosts serving the same weights, and the corpus now holds two sources that disagree about which is which. glm-52 (T2, simon-willison) puts $1.40 / $4.40 per 1M at OpenRouter hosts; deepseek-glm-qwen-price-gap (T4) calls those figures Z.ai's list price and reports resale separately at ~$0.55 / $1.85. Neither is first-party, and the T4 article's headline claim — a 10x gap to DeepSeek V4 Flash at $0.14 input — only holds on its own reading.

Left open, because the underlying point is more interesting than the arithmetic: open weights turn a model's price into a market rather than a number. Anyone can serve them, so "what does GLM-5.2 cost" has no single answer, and comparisons against closed models (one vendor, one rate) are not like-for-like. That belongs alongside the tokens-per-task multiplier in llm-api-pricing — two different ways a headline per-token rate fails to predict a bill.

Method note worth keeping: this conflict surfaced because a T4 aggregator was ingested into a spoke already holding T1 and T2 on the same subjects. Its spec sheet checked out against both, its GLM release date was three days off, and its pricing framing broke exactly where the headline lived. Weak sources are cheap to ingest and worth grading against strong ones already in the corpus.

Safety alignment is a build-time setting on open weights

The spoke has argued that a permissive licence is permission to rewrite a model, not just to run it (open-weight-models, gemma-4-31b-it-scotoma). The 2026-08-22 ingest of heretic moves that from an observation about one derivative to a fact about the tier.

Three things now sit on the table together. The mechanism is published and simple: Arditi et al. 2024 find refusal in a one-dimensional subspace across 13 open chat models, removable by orthogonalizing two matrices per layer against a direction you get from a difference of means. The tooling is finished: pip install heretic-llm, one command, 20–30 minutes on a 4B model on a consumer GPU, explicitly pitched at users with no knowledge of transformer internals. The volume is there: well over 5,000 tagged models on Hugging Face, and readyart — this spoke's first derivative publisher — turns out to be a downstream user of that tool rather than an independent practitioner.

The consequence for how this wiki reads a model release: for an open-weight model, the refusal behaviour reported in a launch post or a system card describes the checkpoint the lab published, not the artifact users run. A closed model's safety posture is a property of the product; an open model's is a default that ships with an uninstaller. That is a real difference between the two tiers, and none of the licence axes on open-weight-models captures it — Apache-2.0, MIT, the Llama MAU clause and the Kimi K3 revenue threshold all say who may serve the weights, and nothing about who may change what they refuse.

What is still missing is the cost side. heretic's own table quantifies the trade for the first time in this corpus — same 3/100 refusals as hand-made abliterations at 0.16 KL divergence against 1.04 and 0.45 — but KL divergence is a distance between output distributions, not a capability score, and the wiki holds no independent MMLU/GSM8K read on any abliterated model. So the honest statement is: removing refusals is cheap and easy, and how much intelligence it costs is still self-reported by the people doing the removing.

The downstream half arrives, and it is weakly sourced (2026-09-04). Everything above is about supply: the mechanism, the tool, the volume, the publisher. uncensored-ai-rekhi is the first source here that asks who runs the output — a T3 Medium essay that relays security-industry reporting the wiki has not read. It describes a demand side with two shapes. A rental tier: three-plus vendors selling hosted uncensored inference at $30–$200/month on models over 80B, with maybe a thousand customers, now paged as dark-llm-services. And a squatted tier: a 293-day SentinelOne/Censys sweep finding hundreds of exposed open-source deployments with guardrails explicitly removed. The llm-provider typology has no slot for either; both are providers by every functional test and train nothing.

That completes an argument the spoke had been making one-sided. The licence axes on open-weight-models all constrain serving by an identifiable party; the essay's structural point is that publication is irreversible, so none of them reaches a copy already downloaded, and safety work that binds is safety work done before release. Note what the price says: $30/month is not competing with llm-api-pricing on capability, because the product is the absence of a refusal — supply-side free, and priced on convenience.

Two cautions, and they pull opposite ways. The numbers are third-hand from vendors whose business is the size of the threat, and none is verified here. But the essay also conflates weight editing, safety-breaking fine-tuning and prompt jailbreaking under the word "abliteration" — and only the first two need open weights at all. Its most alarming evidence (a 97% jailbreak rate, GPT-5 broken in a day) comes from the third category, against hosted frontier models. So the popular argument that open weights are the problem draws much of its force from evidence about closed ones, which is a reason to be more careful with the framing rather than less, in both directions. The correction is recorded on abliteration with the three-way table.

And the popular account has dropped the cost entirely — abliteration as a clean subtraction, "a model that simply doesn't have the machinery for refusal anymore." The corpus's own position is the paragraph above this one: the trade is real, it is measured only by the people making the edit, and nobody neutral has run a capability eval. The version of this story circulating outside the field asserts an answer the field has not produced.

The open tier is now previewing architectures, not just chasing them

The story this wiki has told about open weights is a chase: capable models arriving a few months behind the frontier at a fraction of the price. qwen38-flash-next does not fit that shape. Qwen released it as an explicit early preview of the architecture behind Qwen4, exactly as Qwen3-Next previewed the design that then ran through Qwen3.5, 3.6, 3.7 and 3.8 — the next generation's geometry, open, a family early.

Two consequences for the map. Openness is being used as a design-review channel: the weights go out so the community can examine an architecture before the flagship is built on it, a different motive from distribution or ecosystem share, and no closed lab in this corpus does it or can. And it sharpens "the architecture has stopped corresponding to a lab" above — if the shape travels faster than the checkpoints do, then publishing the shape first is not a side effect of open weights, it is the point. The 200,000+ derivatives count checkpoints; this is Qwen putting the skeleton on the table a generation ahead.

The counterweight is that nothing here is independently checked. Every capability, throughput and cost figure is Qwen's own, two headline benchmarks are its in-house suites, the launch post states no licence at all, and the only outside contact the corpus holds is simon-willison running 1–2 bit community quantizations and liking a pelican (qwen38-flash-next-willison). A lab that publishes its next architecture early also gets to frame, first, what that architecture is good at.

What this spoke's T4 count means, and what it doesn't (2026-08-30)

The 2026-08-30 tier decision requires a spoke to say beside its histogram what a low floor is measuring. This spoke's floor is not low — T4 3 of 51 tiered pages — and the reason is worth stating, because the contrast with search-marketing-wiki is the useful part.

Search marketing's T4 share is the highest in the corpus and it is a fact about the field: the industry publishes argument from practice and there is no body of evidence being ignored. Here the opposite holds. Model providers publish constantly — rate cards, system cards, model documentation, benchmark tables — so a primary source exists for almost every claim this spoke needs, and the T1 count (18) reflects that. The three T4 pages are therefore a fact about those three sources, not about the industry:

The discipline that follows is the reverse of search-marketing's. There, T4 is unavoidable and the work is to say so. Here a T4 ingest should be rare and should carry a reason for existing — usually that a primary source is missing on that specific point, as with sakana-fugu's undisclosed routing pool, or that marking a weak claim against held primaries is itself informative.

Nobody wins on both axes, and the effort dial went backwards (2026-09-01)

coderabbit-fable-5-1-review is the first source here that scores a job rather than a capability: 45 code-review tasks against 105 catalogued defects, with recall, precision, comment volume and latency all reported. It changes two things.

A capability index has to choose an error to care about, and usually hides the choice. Ranked by recall the order is GPT-5.6 Sol (69.7%) → Fable 5.1 (61.0%) → Opus 5 (55.2%); ranked by precision it inverts almost exactly (39.3% → 37.3% → 31.6%). No model is best at both, and every composite in llm-benchmarks would have to pick a weighting to produce a winner — the same editorial choice that page already flags in the Artificial Analysis Index, now visible as a straight reversal rather than a rounding difference. The floor is also lower than any index suggests: precision runs 31.6%–39.3% across all four models, and no capability benchmark in this spoke reports what share of a model's output was wrong.

And raising effort made the model worse. Fable 5.1 at high reasoning took 21:36 against 18:38 and scored 57.1% recall against 61.0%. Everything this spoke records about effort — the tier ladder, the price-for-capability argument on claude-fable-5, Anthropic's own positioning of Opus 5 as "close to Fable 5 at half the price" — assumes more compute buys more quality, monotonically. One vendor harness on one task family is not a refutation, and it is the first counter-instance in the corpus. Watch for a second before treating it as anything more. 2026-09-23: a weak second arrived. claude-opus-5-5-announcement reports medium effort at or above max on FrontierCode, and sells medium as the default. See "A vendor concedes the margin" above.

The 5.1-over-5 upgrade fits the same shape: 34% fewer comments, 70% fewer nitpicks, 49% more latency, recall unchanged. Priced identically to its predecessor, the newer model in this test bought restraint rather than reach — which is a product decision, not a capability gain, and it is invisible to any score that counts only what was found.

Cross-spoke adjacency

Index — LLM Providers Wiki

Catalog of every page, grouped by schema.org @type. Spine: synthesis (thesis), log.md (history), this file (catalog). Some wiki-links resolve cross-wiki (../research-wiki, ../llm-inference-wiki) — intentional bridge links. Pricing/ranking facts are dated snapshots.

DefinedTerm (concepts)

Organization (providers)

SoftwareApplication (models)

TechArticle / BlogPosting / Article / Dataset (sources)

Person

Synthesis

Bridge nodes (live in sibling wikis, linked cross-wiki)

anthropic · claude-opus-4-8 (../research-wiki) · llm-inference · kv-cache · cerebras-inference · cerebras-systems (../llm-inference-wiki)

Related pages

Links to

Linked from