Claude Fable 5.1: Anthropic's Fable Flagship, Fully Specified

Claude Fable 5.1 is the generally available Fable-class flagship, live since September 1, 2026, and it sits in the middle of three names readers routinely mix up. The no-suffix Claude Fable 5 is the June 2026 original that Fable 5.1 supersedes at the same per-token price. Claude Fable 5.2 is a third name that has never shipped and has no vendor page anywhere — including the vendor's own documentation, pricing table and newsroom, none of which mention it. And Claude Mythos 5.1 is not a separate model either: it is the same weights as Claude Fable 5.1 sold through a restricted access program. This page is the reference for the one model that a reader can actually call today: its envelope, its full rate card, what the vendor documents about effort and thinking, every published benchmark figure and the setting it was run at, the independent read, and what it costs on OrcaRouter.

The Fable family: three names, two models, one suffix trap

The suffix confusion is not cosmetic — the three strings differ by a single digit and lead to different envelopes, different prices and, in one case, to nothing at all. Sorting them out first is what makes the rest of this page readable.

• Claude Fable 5 — released June 9, 2026 as Anthropic's first Fable-class model. It is still served and still priced identically to its successor: $10 per million input tokens and $50 per million output, with cache reads at $1.00 per million. It is the model the .1 release replaces, and it is not the model this page is about.

• Claude Fable 5.1 — this model. Released September 1, 2026, same $10 / $50 headline as Claude Fable 5, with cache reads cut to $0.25 and a shorter list of behavioral changes than the version bump suggests. Anthropic's own documentation describes it as the version to reach for "for demanding reasoning and long-horizon agentic work."

• Claude Fable 5.2 — does not exist as a purchasable or documented model. Anthropic's newsroom lists no such announcement; its model overview, pricing page and Fable 5.1 model page do not carry the name; Artificial Analysis has no entry for it; and OrcaRouter's catalogue returns "model not found." Write-ups you may have seen are pre-release reporting, which is a different article from this one and is linked below rather than re-litigated here.

• Claude Mythos 5.1 — the same underlying model as Claude Fable 5.1, with safeguards tuned for cybersecurity and life-sciences work instead of general release. It is available only through Anthropic's trusted-access programs (Project Glasswing and the Cyber Verification Program), currently only to a set of US organizations, and it is priced identically to Claude Fable 5.1. It is not in OrcaRouter's catalogue, and this page is not about it.

What shipped on September 1, 2026, and what the model actually is

Anthropic shipped Claude Fable 5.1 and Claude Mythos 5.1 in a single announcement on September 1, 2026, describing them as "the world's most advanced models for coding and knowledge work." That is a vendor claim, quoted as such. The verifiable part of the release is narrower and more useful: one model, two safeguard profiles, and a price change concentrated entirely in cached input.

The envelope, per Anthropic's own model page and API documentation: a 1,000,000-token context window at standard per-token pricing — no long-context premium, so a 900,000-token request bills at the same rate as a 9,000-token one — and up to 128,000 output tokens. Input is text and images; output is text. Tool use, vision, structured outputs and multilingual work are all supported. The reliable-knowledge and training-data cutoffs are both June 2026.

The API surface: claude-fable-5-1 on the first-party Claude API and on Claude Platform on AWS, Google Cloud and Microsoft Foundry; anthropic.claude-fable-5-1 on Amazon Bedrock. Anthropic commits to supporting it "not sooner than September 1, 2027," which is the horizon to plan a migration against rather than a deprecation date.

One deployment constraint is worth knowing before you standardize on it: Claude Fable 5.1 runs with production safeguards that can refuse a request outright, and Anthropic's fallback guidance matters in practice — a refusal returns a normal HTTP 200 whose stop reason is refusal, not an error. The announcement also sets out a data-retention path: Anthropic says its new Enterprise Frontier Safeguards system, which keeps data in customer-controlled cloud infrastructure, will be phased in for enterprise customers from later this fall, and that until then eligible customers can run Fable 5.1 with zero data retention.

The full rate card, and where the money actually moved

Every figure in this section is from Anthropic's own pricing documentation, read on September 28, 2026. Base token pricing is unchanged from Claude Fable 5: $10.00 per million input tokens and $50.00 per million output tokens. The change is in caching.

• Input — $10.00 per million tokens, identical to Claude Fable 5.

• Output — $50.00 per million tokens, identical to Claude Fable 5.

• Cache read (a hit) — $0.25 per million tokens, down from $1.00 on Claude Fable 5. Anthropic states this as a 75% reduction, and it is the entire mechanism behind the release's cost story: a cache hit now costs 2.5% of the standard input price instead of 10%.

• Cache write — $12.50 per million tokens for a five-minute window, $20.00 per million for a one-hour window. These are unchanged from Claude Fable 5, so the write side of the trade is not part of the discount.

• Batch API — a 50% discount on both directions, so $5.00 per million input and $25.00 per million output for asynchronous work.

• Data residency — requesting US-only inference carries Anthropic's standard 1.1x multiplier on every token category.

Anthropic models the result in its announcement as roughly 25% lower cost on typical workloads and up to around 45% on highly agentic ones, measured at default effort over four weeks of August 2026 usage. Those two percentages are vendor estimates from Anthropic's own usage data, not independently audited figures, and they hold only for workloads that re-read context. A single-turn request with no cache reuse costs exactly what it did on Claude Fable 5.

Effort, thinking, and the setting each benchmark was run at

This is the part of the spec sheet most coverage skips, and it is the part that decides whether two published scores are comparable at all.

Claude Fable 5.1 uses adaptive thinking, and thinking is always on — Anthropic's model page lists it as "Adaptive (always on)," and there is no mode that turns it off. The older fixed-budget form of extended thinking is not merely discouraged here: the parameter is rejected. On Claude 4.7-and-later models, a request that sets thinking: {type: "enabled", budget_tokens: N} returns a 400 error, and Anthropic's documentation names Claude Fable 5.1 as one of the models affected. The replacement is a depth dial rather than a token budget.

That dial is output_config.effort, and Claude Fable 5.1 supports all five levels: low, medium, high, xhigh and max. The default is high, which is also what you get by omitting the parameter entirely. Anthropic's own guidance is to start at high, step up to xhigh or max for the most capability-sensitive agentic and coding work, and step down to medium or low for routine or latency-sensitive work once your evals show quality holds. Anthropic also states that at Low or Medium effort Fable 5.1 reaches results similar to or better than Claude Fable 5 at much lower cost — and that the model defaults to High effort in Claude Code and to Medium in Claude Cowork and on Claude.ai, so the same model name can behave differently depending on which surface you call it from.

Now the benchmark-setting question, answered precisely. Anthropic publishes each of these benchmarks as an accuracy-versus-cost curve plotted at all five effort levels, and its headline comparison table does not attach a single effort level to a single number. The one figure in Anthropic's material that is pinned to an effort level is CursorBench 3.2.0: the partner quote in the same announcement reports 73.4% "at max effort," which matches the 73.4% in the table. Everything else should be read as the curve, not the point — which is also why a bare score comparison against a different lab's model, evaluated at a different effort and a different harness, tells you very little.

The benchmark sheet, and who ran it

Every number in this section is Anthropic-reported and has not been reproduced independently. Two caveats from Anthropic's own footnotes belong with them: the Terminal-Bench-Science 0.1 standard error is ±3.5–4.5 points per model, and the OSWorld 2.0 scores are on the benchmark authors' August 2026 task release, so they are not comparable to previously published OSWorld 2.0 results.

• Terminal-Bench-Science 0.1 — Claude Fable 5.1: 52.6% vs Claude Fable 5: 24.7%, Claude Opus 5: 29.0%, GPT-5.6 Sol: 22.4%. The "more than double" framing holds, and Anthropic notes its harness reproduces the public leaderboard's Opus 5 result (30.0%) and Claude Fable 5 result (21.4%) within noise.

• Terminal-Bench 4.0 — Claude Fable 5.1: 55.8% vs Claude Fable 5: 42.0%, Claude Opus 5: 52.3%, GPT-5.6 Sol: 37.3%. Claude Mythos 5.1 reaches 60.9%, a gap Anthropic attributes to tasks where the less precise cyber safeguards on the gated model intervened.

• Humanity's Last Exam — Claude Fable 5.1: 60.9% without tools and 65.0% with them, vs Claude Fable 5 at 57.8% / 63.8% and Claude Opus 5 at 56.6% / 63.6%.

• GDPval-AA v2 — Claude Fable 5.1: 1,853 vs Claude Fable 5: 1,723, Claude Opus 5: 1,824, GPT-5.6 Sol: 1,711.

• AutomationBench — Claude Fable 5.1: 31.4% vs Claude Fable 5: 17.1%, Claude Opus 5: 26.9%, GPT-5.6 Sol: 19.6%.

• CursorBench 3.2.0 — Claude Fable 5.1: 73.4% at max effort, vs Claude Fable 5: 70.5%, Claude Opus 5: 70.0%, GPT-5.6 Sol: 67.2%.

• OSWorld 2.0 — Claude Fable 5.1: 77.9% partial and 41.7% strict, vs Claude Fable 5 at 72.9% / 36.1% and Claude Opus 5 at 75.4% / 39.6%.

• Partner-run, first-party-reported — Browserbase measured 82% of its hardest browser-agent tasks completed in about ten minutes each, against 74% for Claude Opus 5 and 57% for Claude Fable 5; Crosby's RedlineBench contract-redlining score moved from 47.9 to 57.0 over Fable 5; Samaya's FrontierFinance rubric moved from 49.2% to 55.9%.

Read the sheet with one structural fact in mind, because Anthropic states it plainly: Claude Fable 5.1 and Claude Fable 5 were both evaluated with production safeguards enabled, and on tasks where those safeguards intervened both models scored a zero — on OSWorld 2.0 for each, and on AutomationBench for Claude Fable 5. Anthropic says cyber tasks that triggered an intervention were completed by Claude Opus 4.8 and biology tasks by Claude Opus 5, and that this "likely reduces the performance" of the Fable models on those benchmarks. A refusal, in other words, is scored as a failure, which is a fair reading of an agent that stops — and also a reason the numbers move when the safeguards change.

The independent read: what a third party measured, and at what effort

Independent figures here come from Artificial Analysis, whose Intelligence Index is the most widely cited third-party aggregate and who publish the exact configuration each result was produced under. As of September 28, 2026, its page for Claude Fable 5.1 carries the Artificial Analysis Intelligence Index v4.3.2 at 53 under the configuration "Adaptive Reasoning, Max Effort, Default Fallback," at a weighted cost of $7.63 per Intelligence Index task and 69.7 output tokens per second.

The effort ladder is the interesting part, and it is the thing a single headline number hides. Artificial Analysis publishes a separate page for each effort level: 47 at low ($2.37 per task), 49 at medium ($2.98), 51 at high ($3.91), 53 at xhigh ($5.98) and 53 at max ($7.63). Two things follow. First, the effort dial is real and it is priced: dropping from max to high gives up two index points and costs 49% less per task, and dropping to low gives up six points for 69% less. Second, xhigh and max land on the same headline score at different costs, which is the ordinary shape of these curves and a reason to measure your own workload rather than buy the top setting by default. The same index revision puts Claude Fable 5 at 50 ($8.75 per task), Claude Opus 5 at 51 ($5.86) and Claude Opus 5.5 at 58 ($5.98).

Two honesty notes on those comparisons. They are all one vendor's harness on one index revision, so they are comparable to each other and to nothing else. And they do not reproduce Anthropic's ordering: on Anthropic's vendor-reported sheet Claude Fable 5.1 clears its whole comparison field, while on Artificial Analysis's index a cheaper Anthropic model released three weeks later scores five points higher. Both can be true, because the index and the vendor sheet weight different work — and because a headline index is a summary of ten evaluations, not a verdict on your task.

Calling it: the version question under one key

The practical reason this page exists is that "claude fable 5.1" is a query people type on its own, usually to answer one of three things: which of the Fable names is current, what it costs, and how to try it without standing up a second vendor contract. The first two are answered above from Anthropic's own pages. The third is where a router earns its place.

Claude Fable 5.1 is on OrcaRouter as anthropic/claude-fable-5.1 at Anthropic's list price with 0% markup — $10.00 per million input tokens, $50.00 per million output, cache reads at $0.25 and cache writes at $12.50 — alongside Claude Fable 5 in the same catalogue. Both are reachable from one key, one base URL and one billing line, through either the OpenAI-compatible /v1/chat/completions endpoint or Anthropic's own /v1/messages shape, so an existing Claude SDK client can point at the new model by changing two values. That is the cheapest possible version test: run the same prompt against Claude Fable 5 and Claude Fable 5.1 at high and xhigh effort, compare cost per completed task rather than cost per request, and let your own evals decide whether the .1 is worth moving for. Because the markup is zero, a vendor price change — as with the cache-read cut on September 1 — is live on our side the same day rather than waiting on a re-rate.

If you are choosing between the Fable line and something cheaper, note that Anthropic itself points most workloads at Claude Opus 5 in its own documentation, reserving Fable 5.1 for demanding reasoning and long-horizon agentic work. And if what you actually need is the gated biology or cyber capability, that is Claude Mythos 5.1 through Anthropic's access programs — not something this page, or our catalogue, can route you to.

Why this is a reference page and not a launch story

Claude Fable 5.1 shipped on September 1, 2026 — four weeks before this page was written — so the launch is not the news, and this piece is deliberately not a second announcement. The argument for the page is different and, we think, more durable: it is the canonical reference for a model that is live in our catalogue today, the page a reader reaches after searching the model's own name, and its warrant is standing search demand that we can see first-hand rather than the freshness of a launch. The September 1 release date is used here to date the model, not as an event to report. Our launch-day coverage and the pre-launch leak reporting remain separate pieces, and this page links to them rather than restating either.