Claude Fable 5

Which lens this is. The market treatment: price discovery, tier positioning, and third-party capability reads. The model as substrate — what it costs the wiki-maintenance thesis, the guardrail/Mythos split, the export suspension — is research-wiki/claude-fable-5, which also holds the primary Anthropic announcement. Same slug in two spokes is deliberate (HUB.md → dual-lens); link the lens you mean explicitly.

Anthropic‘s claude-fable-5 — a member of the Claude 5 model family alongside Sonnet 5 and Opus 4.8, available as a model choice inside Claude Code. It is the family’s fast option rather than the reasoning flagship, but “fast” here is not “weak”: on Code Arena WebDev it holds the closed-model coding lead, sitting just above GLM-5.2 (see synthesis‘s frontier-premium thread). So it reads as a fast, coding-strong model, which makes the cost anecdote below the more surprising.

The spoke’s read here was built second-hand, source by source; the resolution at the foot of this page is where it lands. It starts with a T4 blog anecdote, Njenga’s Fable 5 vs Sonnet 5 test, which — on a single "Hello" inside Claude Code — clocked Fable 5 at $0.4745 versus a cheaper Sonnet 5, a claimed ~3.2× cost gap, with Fable also slower (7 s vs 3 s) and caching less of the session (15.6k vs 24.8k cache-read tokens). That result is counterintuitive for a model pitched as the fast/cheap pick, and it rests on one unrepeated run — see the caveat on the source page and the open question in synthesis.

A second source now corroborates the “Fable 5 is expensive” read (2026-07-08). The Decoder’s Grok-4.5 pricing comparison (grok-4-5-price-vs-benchmarks) lists Fable 5 at $10/$50 per 1M tokens (input/output) — the most expensive of the four models it compares, above Opus 4.8 ($5/$25). That is ~3.3× Sonnet 5‘s $3/$15, closely matching the ~3.2× gap Njenga measured inside Claude Code — so the anomaly is looking less like a harness artifact and more like Fable 5’s actual list price. Two independent reads now agree Fable 5 is dear; both are non-Anthropic, so it’s corroboration, not proof. The Code Arena WebDev coding lead makes a premium-priced “fast” model less paradoxical: it may be a high-quality creative/coding model, not a budget tier.

Availability shock (2026-07): per openai-gpt56-grok45-clash, the Trump administration export-banned Anthropic’s Fable and Mythos models, citing risk of misuse for cyberattacks; Anthropic suspended them, then redeployed Fable 5 last week. So Fable 5’s availability has been governed by export controls, not only price — a first for a model in this spoke, and an ai-governance cross-spoke tie. (“Mythos” is the class above Opus flagged in the Sonnet 5 system card.)

Third price read + usage economics (2026-07-09). A buyer’s-guide post by Ruben Hassid, “Don’t use Claude Fable 5,” again lists $10/$50 per 1M tokens — so three independent third-party sources now agree on the sticker. It adds two things the earlier reads lacked: Fable 5 reportedly moves to pay-per-use credits after 2026-07-12, and concrete conversation costs (short Q&A ≈$0.15, 19 long turns ≈$6, 40 turns ≈$14) because the model re-reads the whole thread each turn — the token-accounting point (llm-api-pricing) as a felt user bill. Positioning: reserve Fable 5 for “expensive problems” (hard goals, deep research, planning) at high/max effort, one or two turns max; Hassid pitches Opus 4.8 as the cheaper-and-smarter workhorse for ordinary knowledge work. That matches the spoke’s “premium creative/coding tier, not a budget one” read.

Anthropic finally confirms the position, if not the sheet (2026-07-24). The Opus 5 launch post pitches Opus 5 as coming “close to the frontier intelligence of Claude Fable 5 at half the price,” with Opus 5 priced at $5/$25. Halving $10/$50 gives exactly that, so a T1 Anthropic source now implies the rate three third-party reads had already converged on — implication, not a published rate card. (Superseded 2026-08-03: a published rate card did exist, from launch — see the resolution at the foot of this page. The reasoning in this paragraph was sound on the evidence this spoke had; the evidence was incomplete.) More importantly it settles the ordering: Anthropic itself places Fable 5 above Opus 5 on intelligence and sells its own flagship as the discount. Fable 5 is the frontier tier of the family, priced like it. Anthropic also cites Opus 5 beating Fable 5 on OSWorld 2.0 and landing within 0.5% of it on CursorBench 3.2, so the capability gap it’s charging double for is narrow on the benchmarks it chose to show.

A capability read from outside Anthropic (2026-07-26, opus-5-arc-agi-3). ARC Prize puts “Fable-class” models at ~20% on ARC-AGI-3, well ahead of the 7.8% GPT-5.6 Sol record and well behind Opus 5‘s 30.2%. On Guanghan Ning’s private Witness benchmark the three are level: Fable 5 statistically ties Opus 5 (43.4) and kimi-k3. So on a novel-task benchmark the family’s premium tier sits second to the cheaper flagship, and on the private one the gap disappears. It’s the first non-Anthropic capability number on Fable 5 in this spoke; every earlier third-party read here was about price.

Resolved — Anthropic published the rate all along (found 2026-08-03). This page spent four sources triangulating a number the corpus already held. Anthropic’s own launch announcement of 9 June 2026 states it outright — “Pricing (both): $10 / $50 per Mtok,” framed as “less than half the price of Claude Mythos Preview” — and it is summarized in the sibling spoke as research-wiki/claude-fable-5-mythos-5-announcement. So the sticker was published, not inferred, a month before this page was created.

That does not retract anything above; it re-labels it. The three third-party reads and the Opus 5 arithmetic were not a substitute for a rate card, they were independent corroboration of one — and they agree with it exactly, which is a stronger result than either alone. What was wrong was only the claim that no first-party rate existed.

Still genuinely missing from any Anthropic source: context window, max output, and intended tier for Fable 5 (the announcement explicitly does not disclose them; the review supplies 1M ctx / 128K out / Jan-2026 cutoff). The caching half of the cost question is also untouched — it still rests on Njenga’s single unrepeated run. Volatile: date any cost claim.

Related

claude-opus-5 · claude-sonnet-5 · llm-api-pricing · anthropic · fable-5-vs-sonnet-5-token-cost · dont-use-fable-5-hassid · effort-level · opus-5-arc-agi-3 · arc-agi