Claude Fable 5

Sources#

• A Field Guide to Fable: Finding Your Unknowns
• Claude Fable 5 and Claude Mythos 5
• Claude models explained: choosing the best model for your use case
• Claude Opus 5 System Card
• Gemma 4 Technical Report
• Rewriting Bun in Rust

Summary#

Claude Fable 5 is Anthropic's first generally-available Mythos-class model (launched June 2026) — a tier of Claude models that "sit above our Opus class in capability." It is described as state-of-the-art on nearly all tested capability benchmarks (software engineering, knowledge work, vision, scientific research), with the lead widening "the longer and more complex the task." Fable 5 is the same underlying model as Claude Mythos 5 — the two differ only in their safeguards (footnote: Fable from Latin fabula, "that which is told," akin to the Greek mythos). Fable ships the safeguards on; Mythos has them lifted in some areas. The safeguard architecture — classifiers that route risky queries to Opus 4.8 instead of refusing — is documented in Capability-Gated Model Fallback.

Status (as of the 2026-06-14 clip): access suspended. The source page now leads with a banner: "We are suspending access to Claude Fable 5 and Claude Mythos 5. We apologize for this disruption to our customers and are working to restore access as soon as possible." The launch announcement below is preserved; the suspension is the live state.

Pricing and identity#

• $10 per million input tokens / $50 per million output tokens — explicitly "less than half the price of Claude Mythos Preview." Same pricing for Fable 5 and Mythos 5.
• API model id: claude-fable-5 (via the Claude API / platform.claude.com).
• "Fable 5's capabilities exceed those of any model we've ever made generally available."

Capability highlights#

The announcement leads with Fable 5 (general access); the science-heavy results run under Mythos 5 (bio safeguards lifted) and are compiled in Autonomous Scientific Discovery. Fable's headline demonstrations:

• Software engineering. Stripe reported Fable 5 "compressed months of engineering into days" — a codebase-wide migration in a 50-million-line Ruby codebase done in a day that "would otherwise have taken a whole team over two months by hand." On Cognition's FrontierCode eval (passing hard coding tasks while meeting production-codebase standards), Fable 5 scores highest among frontier models "even at medium effort," and is more token-efficient than past Claude models.
• Knowledge work. Highest score of any model on Hebbia's Finance Benchmark for senior-level reasoning (document reasoning, chart/table interpretation, problem solving); IMC reported Fable 5 "aced their trading-analysis evaluations nearly across the board."
• Vision (new SOTA). Extracts precise numbers from detailed scientific figures; rebuilds a web app's source code from screenshots alone. Crucially, it "needs less scaffolding" — previous Claude models struggled to play Pokémon FireRed even with helper harnesses, but Fable 5 beat FireRed with a minimal, vision-only harness (raw game screenshots, no maps/navigation aids/game-state). This is a flagship case of Harness Shrinkage as Models Improve.
• Memory and long-context. Stays focused "across millions of tokens" and improves outputs using its own notes. In Slay the Spire, file-based persistent memory improved Fable 5's performance 3× more than it improved Opus 4.8's, and Fable reached the final act 3× more often.
• Autonomy. Fable 5 and Mythos 5 "can work autonomously for longer than any previous Claude models" — see Task Time-Horizon Scaling.

(The head-to-head benchmark table vs other leading models is published only as an image in the source and is not transcribed here.)

Third-party corroboration (July 2026)#

Everything above is vendor-claim. One external number now exists: DeepMind's Gemma 4 report reproduces the Arena Text leaderboard as of 2026-06-19, where Claude Fable 5 ranks 1st at Elo 1508 ±9 — the top model overall, listed as the closed-model reference point against which open models are measured. Blind side-by-side human ratings, published by a competitor with no incentive to flatter.

This is better evidence for Fable 5's standing than the launch post, and it is the wiki's only non-Anthropic measurement of the model. It bears on capability-as-preferred-by-humans, not on the software-engineering and science claims, which remain uncorroborated. The gap to the best open model (GLM 5.1, 1475) is 33 Elo; to the best open dense model (Gemma 4 31B, 1451) it is 57. See The Open-Weight Frontier Gap.

Practitioner reception: the bottleneck moved onto the human#

Thariq Shihipar (Claude Code team), writing a month after launch (practitioner-opinion, one expert's experience):

"Fable is the first model where I find the quality of the work is bottlenecked by my ability to clarify its unknowns."

The claim is a bottleneck relocation rather than a benchmark: earlier models were limited by what they could do, Fable by what you managed to tell it. It is the human-side counterpart to the vendor's harness-shrinkage demonstrations above (Pokémon on raw pixels, Slay-the-Spire memory gains) — as the model needs less scaffolding, the residual scaffolding that matters is the kind that transfers your context into it. See Unknowns as the Agentic Bottleneck.

The largest published deployment: the Bun port (pre-release Fable)#

Jarred Sumner ported Bun from 535,496 lines of Zig to Rust in 11 days using a pre-release version of Fable 5 (Rewriting Bun in Rust, 2026-07-08, case-study) — the biggest single documented Fable workload in the vault: 50 dynamic workflows, peak 64 concurrent instances, 6,502 commits, 5.9B uncached input + 690M output + 72B cached-read tokens ($165k), merged with 100% of a 1M+-assertion test suite green.

Three things it establishes that the launch post does not. First, a capability discontinuity claim with a named task: Boris Cherny reports the same rewrite failed on every prior generation "even with the steering," and Sumner's own expectation was that it wouldn't work ("at first, I didn't expect it to work"). Second, a price for a frontier-scale outcome — the counterfactual Sumner states is three engineers for a year, which the team says it would never have paid. Third, a failure profile: 19 known regressions shipped, all in the class of constructs that are syntactically identical across Zig and Rust but semantically different (release-erased debug_assert!, bounds-check differences, comptime format strings) — the model reproduced surface form faithfully and missed semantics that live in the build configuration.

Weight it accordingly: Sumner is an Anthropic employee describing an Anthropic-owned codebase, on a pre-release build that is not the GA model. It is a detailed build log, not a benchmark, and there is no comparison arm against Opus or a competitor.

When to pick Fable over Opus (vendor guidance, July 2026)#

Anthropic's model-selection guide gives a rule that is explicitly not benchmark-driven (Cost-per-Task Over Cost-per-Token, vendor-claim): Opus and Fable have "similar benchmark scores," but "in real-world situations, larger models such as Fable tend to have more wisdom, creativity, and writing skills." The stated rule of thumb — if your evals show Opus struggling, Fable is the answer; if Opus already clears the bar, its speed and price make it the better choice. A vendor conceding that its own benchmarks fail to separate two adjacent classes is a first-party construct-validity admission (Measuring Beyond Accuracy Saturation).

The same guidance positions Fable as the advisor in the cheap-worker/strong-advisor pattern: Sonnet 5 coached by a Fable 5 advisor reaches within 10% of Fable 5's SWE-bench Pro score at 63% of the price of running Fable 5 for the whole task.

Safeguards (why Fable exists)#

Fable is the safeguarded SKU. Because Mythos-class capability "could be misused to cause serious damage" (especially cybersecurity, biology), Fable ships with classifiers covering cybersecurity, biology & chemistry, and distillation. When a classifier fires, the response is handled by Opus 4.8 rather than refused, and the user is told. Anthropic tuned the classifiers conservatively (they "sometimes catch harmless requests," triggering in "less than 5% of sessions"); more than 95% of Fable sessions involve no fallback at all, and for those Fable's performance "is effectively the same as that of Mythos 5." Full treatment: Capability-Gated Model Fallback.

Alignment: the automated alignment assessment found Mythos 5's misaligned behavior "low, and similar to that of Opus 4.8"; "given that they are the same underlying model, Fable 5's level of alignment will be similar." A 30-day data-retention requirement now applies to all Mythos-class traffic (safety use only).

Availability and rollout#

• Fable 5 "available everywhere today"; fully available from launch on the Claude API and consumption-based Enterprise plans.
• Staged rollout on subscription plans: free on Pro/Max/Team/seat-based Enterprise from launch through June 22; from June 23 it requires usage credits; Anthropic "aim[s] to restore Fable 5 as a standard part of subscription plans" once capacity allows.
• Demand expected "very high, and difficult to predict" — the stated reason for staged access. (See the suspension note above for the live state.)

Connections#

• Claude Opus 5 — inherits Fable 5's cyber classifier stack minus the source-code vulnerability block, and trips it far less often (5% of API calls on FrontierBench vs Fable 5's 42%)
• Mythos Model — the model tier and its first member (Mythos Preview); Fable 5 is the first general-access Mythos-class model
• Claude Mythos 5 — the same underlying model with safeguards lifted; the cyber/bio sibling
• Claude Opus 4.8 — the fallback model: safeguarded Fable queries are answered by Opus 4.8 instead of refused
• Capability-Gated Model Fallback — Fable's defining safeguard architecture (classifiers + fallback-not-refuse + 30-day retention)
• Harness Shrinkage as Models Improve — Pokémon-FireRed-vision-only and Slay-the-Spire-memory are flagship harness-shrinkage demonstrations
• Autonomous Scientific Discovery — the science results (run under Mythos 5, the bio-safeguards-lifted form of the same model)
• Task Time-Horizon Scaling — "work autonomously for longer than any previous Claude models" pushes the autonomy-duration curve
• Anthropic — vendor
• Claude Code — the agentic runtime Mythos-class coding gains flow through
• Claude Sonnet 5 — the opposite end of the safeguard spectrum: Sonnet 5's default cyber safeguards are explicitly "less strict than those launched with Fable 5," which block a much wider range of cyber tasks and fall back to Opus 4.8
• Unknowns as the Agentic Bottleneck — Thariq Shihipar's claim that Fable is the first model where the human's unarticulated context, not the model, sets output quality
• Thariq Shihipar — the practitioner account; used Claude Code with Fable to edit the Fable launch video end to end
• The Open-Weight Frontier Gap — Fable 5 is the rank-1 closed reference point in DeepMind's Arena table; the corroboration comes from a competitor
• Gemma 4 — the competitor's report that publishes the number
• Cost-per-Task Over Cost-per-Token — where Fable sits in Anthropic's selection framework: the top class, chosen when Opus visibly struggles, and the recommended advisor for a cheaper worker model
• Dynamic Workflows: An Algebra for Agents — the largest published Fable workload: the 11-day Bun Zig→Rust port on a pre-release build
• Jarred Sumner / Bun — the practitioner and the codebase behind that deployment

Open Questions#

• Why was access suspended after launch? The source banner gives no reason (capacity? a safety finding? the UK-AISI jailbreak progress noted in Capability-Gated Model Fallback?). Not in source.
• Exact benchmark numbers vs GPT-5.x / Gemini are image-only in the source; not transcribed.
• How much of Fable's general-access experience is actually Fable vs Opus-4.8 fallback for security-research-adjacent users whose queries trip the conservative classifiers?

Sources#

• Claude Fable 5 and Claude Mythos 5 — Anthropic, "Claude Fable 5 and Claude Mythos 5" (June 2026; AAV edit June 9, 2026)
• A Field Guide to Fable: Finding Your Unknowns — Thariq Shihipar, 2026-07-04 (practitioner-opinion): the bottleneck-relocation claim and the launch-video build account
• Gemma 4 Technical Report — Table 4 (empirical): third-party Arena Text standing, rank 1 at Elo 1508 ±9 as of 2026-06-19
• Claude models explained: choosing the best model for your use case — Anthropic, July 2026 (vendor-claim): the Opus-vs-Fable decision rule and the advisor-strategy result
• Rewriting Bun in Rust — Jarred Sumner, bun.com (2026-07-08, case-study): the Bun Zig→Rust port on a pre-release Fable 5 — scale, cost, and the 19-regression failure profile
• Claude Opus 5 System Card — Anthropic, July 24, 2026 (empirical): Fable 5 as a comparison arm throughout, and the classifier flag-rate contrast quoted above (§8.5). Parse hazard: this PDF's raw markdown shifts table rows — model names land inside value columns across the §4 safeguards tables (4.1.1.A, 4.2.B, 4.3.1.B, 4.3.2.A, 4.4.2.B, 4.4.3.B), the §5.1 agentic-safety tables (5.1.1.A–5.1.3.A) and Table 8.13.6.A, so a row read literally can hand one model's score to another. Figures quoted here were reconciled against the PDF on 2026-08-03; never quote a table row from the raw markdown unchecked