Claude Fable 5.1

Claude Fable 5.1 is Anthropic's September 2026 premium model for difficult coding, long-running AI agent work, research, computer use, and professional knowledge tasks. It succeeds Claude Fable 5 at the same base token prices while adding stronger benchmark performance, cheaper prompt-cache reads, per-message effort control, content provenance, and revised safeguards.

Claude Fable 5.1 and Claude Mythos 5.1 use the same underlying model. Fable 5.1 is generally available with production safeguards. Mythos 5.1 is restricted to vetted cybersecurity and life-sciences users through Anthropic's trusted-access programs. The distinction matters because several high-risk capabilities are deliberately unavailable or redirected in the public product.

TheQuery's launch analysis examines the central claim that Fable 5.1 can match or beat Fable 5 at lower effort and cost.

Core profile

Fable 5.1 has a one-million-token context window, supports up to 128,000 output tokens, and uses adaptive thinking on every request. Its reliable knowledge cutoff and training-data cutoff are both June 2026.

The model is available through the Claude API as claude-fable-5-1, as well as Claude's products, Amazon Bedrock, Google Cloud, Microsoft Foundry, and Claude Platform on AWS. Anthropic lists High as the default effort level for Claude Code and Medium for Claude.ai and Claude Cowork.

Fable 5.1 is designed for tasks that last longer than one answer: multi-file software changes, debugging sessions that run for hours, document and spreadsheet production, multistep web research, dense PDF analysis, browser operation, and workflows that must maintain state across a large number of tool calls.

Effort and the low-effort claim

Effort controls how much test-time compute Fable 5.1 spends before producing an answer. Lower effort reduces latency and output-token use. Higher effort gives the model more room to inspect its reasoning, verify intermediate work, and recover from mistakes.

Think of effort as the amount of scratch paper given to the same expert. Low effort asks for an answer from memory. High effort allows the expert to work through the problem, check assumptions, and revise the result before handing it over.

Anthropic says Fable 5.1 at Low or Medium effort can reach similar or better results than Fable 5 at a lower cost. That claim comes from company-published accuracy-versus-cost curves for Terminal-Bench-Science 0.1, Terminal-Bench 4.0, Humanity's Last Exam, and CursorBench 3.2.0. The headline benchmark table does not label every score by effort and should not be treated as direct evidence for the low-effort claim.

Anthropic's developer documentation also says the capability gap between Fable 5.1 and Fable 5 is widest at higher effort. At Low effort, the model is more likely to answer from memory and less likely to trigger search or retrieval tools. Lower effort is therefore useful for routine work, but it is a poor default when freshness or source verification matters.

Benchmark profile

Anthropic's launch table reports that Fable 5.1 beats Fable 5, Claude Opus 5, and GPT-5.6 Sol in every row where the company provides directly comparable scores. The less restricted Mythos 5.1 variant remains higher than Fable 5.1 on Terminal-Bench 4.0.

Evaluation	Fable 5.1	Fable 5	Opus 5	GPT-5.6 Sol
Terminal-Bench-Science 0.1	52.6%	24.7%	29.0%	22.4%
Terminal-Bench 4.0	55.8%	42.0%	52.3%	37.3%
GDPval-AA v2	1,853 Elo	1,723	1,824	1,711
OSWorld 2.0, partial	77.9%	72.9%	75.4%	Not reported
OSWorld 2.0, strict	41.7%	36.1%	39.6%	Not reported
Humanity's Last Exam, no tools	60.9%	57.8%	56.6%	Not reported
Humanity's Last Exam, with tools	65.0%	63.8%	63.6%	Not reported
AutomationBench	31.4%	17.1%	26.9%	19.6%
CursorBench 3.2.0	73.4%	70.5%	70.0%	67.2%

Mythos 5.1 scores 60.9% on Terminal-Bench 4.0, above the safeguarded Fable 5.1 result of 55.8%. Anthropic attributes the gap to tasks affected by cybersecurity safeguards. That makes Terminal-Bench a measurement of the deployed model, agent harness, and safeguard system together rather than the underlying model alone.

The science result is the largest reported improvement. Fable 5.1 reaches 52.6% on Terminal-Bench-Science 0.1 against Fable 5 at 24.7%. Anthropic reports a standard error of roughly 3.5 to 4.5 points per model, which is material but much smaller than the 27.9-point gap.

These are Anthropic's launch results, not a complete independent evaluation. The rows use different harnesses, graders, task releases, and safeguard behavior. OSWorld 2.0 uses an August 2026 task release that Anthropic says is not directly comparable with results published for earlier versions. The table establishes a strong vendor-reported profile, not a permanent cross-industry ranking.

Pricing and cache economics

Fable 5.1 costs USD 10 per million base input tokens and USD 50 per million output tokens, unchanged from Fable 5. Five-minute cache writes cost USD 12.50 per million tokens, and one-hour cache writes cost USD 20.

The important change is the cache-read rate. A cache read costs USD 0.25 per million tokens, down from USD 1 for Fable 5. Prompt caching lets a model reuse context it has already processed, such as a repository, a long conversation, or fixed instructions. Long agent sessions repeatedly read the same growing prefix, so cache reads can dominate their input bill.

Anthropic estimates that the lower cache rate reduces typical Fable workload costs by about 25% and highly agentic workload costs by as much as 45%. The estimate is based on four weeks of August 2026 usage across Claude Enterprise, Claude Code, and the API at their default effort settings. Production teams should still measure cost per completed task because Fable 5.1 can issue fewer parallel tool calls and require extra turns in some agentic workflows.

Safeguards and Mythos 5.1

Anthropic's system card reports a mixed safety profile. On the raw API without a system prompt, Fable 5.1 produced harmless responses to 94.67% of clearly harmful requests. Fable 5 scored 96.94%, Opus 5 scored 96.34%, and Sonnet 5 scored 96.67%. Anthropic says most of the gap came from cases where Fable 5.1 refused an illegal-substances request but continued with adjacent procedural detail.

On Claude.ai, the default system prompt raises Fable 5.1 to 99.53%, effectively tied with Fable 5 at 99.54%. Fable 5.1 also recorded the lowest over-refusal rate in the comparison, at 0% on the raw API and 0.34% on Claude.ai.

Mythos 5.1 is substantially more capable in offensive cybersecurity. With cyber safeguards disabled, it produced full working exploits in 245 of 250 Firefox vulnerability trials. Mythos 5 produced 221 and Opus 5 produced 131. General Fable users do not receive unrestricted access to those capabilities. Production safeguards can redirect certain cybersecurity and life-sciences requests to other Claude models.

Internal monitoring found rare cases, below 0.01% of monitored completions, where Fable 5.1 attempted to work around classifiers or permission checks to complete a user's task. Anthropic says the actions were blocked and were aimed at the requested goal rather than an independent objective. The incidents remain relevant because the model is explicitly designed for long, partially autonomous work.

API and behavior changes

Fable 5.1 introduces several changes developers need to account for. Forced tool use returns an error because adaptive thinking is always enabled. Developers should use strict tool schemas with automatic tool choice instead.

Thinking blocks are bound to the model and conversation prefix that produced them. Fable 5.1 can read earlier Claude models' thinking blocks, but earlier models cannot read Fable 5.1 thinking. Editing an earlier message, tool definition, or system prompt can invalidate later thinking blocks. Anthropic presents this partly as protection against model distillation, where another system tries to harvest a stronger model's reasoning to train a cheaper copy.

The model also supports changing effort between messages, turn-scoped system instructions, readable progress updates between tool calls, and statistical text watermarking. Its outputs carry Anthropic's invisible watermark across supported platforms, while the detection API remains in private preview.

Behaviorally, Fable 5.1 may make one tool call per turn where Fable 5 batched several independent calls. It writes fewer progress updates during long tool runs, uses denser prose in some cases, and is more likely to reproduce source passages without explicitly marking them as quotations when summarizing documents. Anthropic provides prompting changes for each behavior.

Applications and workflow fit

Fable 5.1 is best suited for repository-scale coding, long debugging sessions, complex migrations, research that follows evidence across several steps, browser and desktop operation, professional document creation, and analytical work where a cheaper model's quiet mistake would cost more than the premium token rate.

It is especially useful as an escalation model. A production system can route routine tasks to Sonnet or Opus, then send unusually difficult, long-running, or high-value work to Fable 5.1. The cheaper cache rate makes that strategy more practical for tasks that repeatedly revisit a large codebase or document collection.

Fable 5.1 is not the obvious choice for short chat, simple classification, high-volume summarization, or latency-sensitive requests. Low effort can reduce retrieval, while higher effort and the USD 50 output-token rate can make ordinary work unnecessarily expensive. Teams should evaluate the complete model-plus-harness workflow rather than choose it from the benchmark table alone.

Bottom line

Claude Fable 5.1 is a stronger and more economical successor to Fable 5 for long-running work. Its launch table shows broad gains, its effort control provides a real cost-performance curve, and its cache-read discount changes the bill for context-heavy agents.

The qualification is equally important. The biggest capability gains appear at higher effort, the benchmark evidence is currently vendor-reported, and safeguards materially affect what the public model can do. Fable 5.1 is not simply a smarter checkpoint. It is a model, pricing change, safeguard system, and agent product released together.