Claude Fable 5.1 launches, up to 45% cheaper for agents

Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 on September 1, three months after the Fable 5 line shipped in June 2026. The two models are the same underlying system running behind different safeguard layers: Fable 5.1 is generally available under the API name claude-fable-5-1, while Mythos 5.1 remains restricted to vetted partner organizations. Anthropic frames the release around three numbers: 52.6% on the Terminal-Bench-Science 0.1 benchmark for scientific-research agents, more than double Fable 5's 24.7%; a 75% cut to the price of cache reads, from $1.00 to $0.25 per million tokens; and savings of up to 45% on complex agentic workloads that lean heavily on cached context.

Fable 5.1 is available today on the Claude API, Amazon Bedrock, Claude on AWS, Google Cloud and Microsoft Foundry. Mythos 5.1, by contrast, ships only to organizations registered with Project Glasswing, Anthropic's vetting program for partners working in cybersecurity or life-sciences research. Both models carry a 1-million-token context window, a 128,000-token maximum output, and adaptive thinking that stays on by default — there is no separate reasoning toggle to switch off.

What the benchmarks actually show

On Terminal-Bench-Science 0.1, an agentic benchmark built around real scientific-research tasks, Fable 5.1 scores 52.6%, ahead of Anthropic's own Opus 5 at 29.0%, Fable 5 at 24.7%, and OpenAI's GPT-5.6 Sol at 22.4%. Anthropic itself cautions that the standard error runs 3.5 to 4.5 points per model, so the ranking matters more than the exact gap between any two scores. On Terminal-Bench 4.0, a command-line coding benchmark, Fable 5.1 reaches 55.8% while Mythos 5.1 reaches 60.9% — the difference between two otherwise identical models, which Anthropic attributes directly to the cost of its safeguard interventions, an unusually candid admission from a model vendor.

• Terminal-Bench-Science 0.1: Fable 5.1 52.6% vs. Opus 5 29.0%, Fable 5 24.7%, GPT-5.6 Sol 22.4%
• Terminal-Bench 4.0: Fable 5.1 55.8%, Mythos 5.1 60.9%
• Humanity's Last Exam: 60.9% without tools, 65.0% with tools
• CursorBench 3.2.0: 73.4%
• AutomationBench 31.4%, OSWorld 2.0 (strict) 41.7%

Why the price just dropped

Base pricing is unchanged, at $10 per million input tokens and $50 per million output tokens. The cut is on cached content: reading from the prompt cache now costs $0.25 per million tokens instead of $1.00, a ratio of 0.025 times the base input price against 0.1 times on every other Claude model. Anthropic measures this as roughly 25% lower cost on a typical workload and up to about 45% on agentic tasks that repeatedly reuse a large context window — the exact pattern behind long coding sessions or multi-step research agents. Batch processing, for workloads that can tolerate delay, stays at $5 and $25 per million tokens.

Three breaking changes for anyone with a live integration

Anthropic also documents three API changes that are not backward compatible, aimed squarely at teams that have already built on Claude. Forced tool use is gone: setting tool_choice to any or to a named tool now returns a 400 error, and Anthropic tells developers to use auto mode combined with strict tool use or structured outputs instead. Thinking blocks are now model-bound — Fable 5.1 can read the reasoning blocks produced by earlier Claude models, but no earlier model can read its own reasoning back, which breaks router or fallback setups that silently downgrade to an older model mid-conversation. And editing earlier turns now invalidates thinking blocks entirely: injecting or deleting per-turn reminders, or rebuilding the system prompt or tools array mid-conversation, throws an error instead of being silently accepted. The check applies to accounts created on or after August 31, 2026, and Anthropic's suggested fix is to move to turn-scoped system messages and server-side context editing.

• No forced tool use: tool_choice 'any' or 'tool' now returns HTTP 400
• Thinking blocks are model-bound: older Claude models cannot read Fable 5.1's reasoning back
• Editing past turns now errors: rebuilding system prompts or tool arrays mid-conversation is blocked

Anthropic also lists real regressions alongside the gains. Parallel tool calling is more variable, so an agent loop may now issue one tool call per turn where Fable 5 batched several; the model narrates its own reasoning less; it answers more often from memory at low reasoning effort instead of checking; and it now favors rewriting a whole file over making a targeted edit. On safety, cyber safeguards now permit identifying software vulnerabilities but still block exploit development, which Anthropic says cuts safeguard interventions inside Claude Code by roughly 60% per session; biology safeguards trigger 85% less often on benign questions. Penetration testing, exploit generation and binary-based vulnerability scanning still redirect automatically to Opus models. Anthropic's system card rates Mythos 5.1 'low-risk' for accelerating its own research and development beyond current trends, but the alignment picture is mixed: Mythos 5.1 cooperates with human misuse and accepts unverifiable claims of authorization somewhat more readily than Opus 5, while being less likely to ignore explicit constraints, hallucinate inputs, or falsely claim to have completed a task.

It's the strongest coding model we've used, but now it's fast, token-efficient, and crucially actually speaks like a normal person.

Early users echoed that shift in tone. Box CEO Aaron Levie said an agent running on Fable 5.1 picked up on subtleties and ambiguities in a dataset that the same test had missed with Fable 5. Anthropic paired the release with three research results the models produced before launch: Mythos 5.1 designed protein binders with roughly a 50% hit rate across 12 targets, against a 10-to-15% norm for the field; Fable 5.1 assembled a Venus elevation map at 2-to-3-kilometer resolution from existing photographs; and custom GPU kernels the model wrote sped up seven open-source genomics models by as much as 2.5 times. On data handling, Anthropic repeated that it has never trained on enterprise data without explicit permission, and announced Enterprise Frontier Safeguards, a high-privacy service — previously withheld from Fable over security concerns — that will let customers run the models on their own infrastructure starting this fall, keeping misuse monitoring in place but letting the customer control how it happens. Standard retention stays at 30 days; zero data retention still requires separate authorization. All output continues to carry a statistical text watermark, and generated files carry C2PA provenance credentials.

For a US company already building on the Claude API, the immediate task is not the price cut — it is an audit of the three breaking changes before pointing any router or fallback logic at claude-fable-5-1. Any code path that sets tool_choice to any or to a named tool will start returning errors and needs to move to auto mode with strict tool use; any agent that falls back to an older Claude model mid-session will silently lose the reasoning Fable 5.1 produced, since that model cannot read it back; and any pipeline that edits system prompts or tool definitions mid-conversation — a common pattern in longer-running agentic systems — needs to switch to turn-scoped system messages before the check is enforced on new accounts. Once that audit is done, the 75% cut on cache reads is real money for teams running long coding sessions or research agents that keep reusing the same large context: Anthropic's own estimate of up to 45% lower cost on agentic workloads is the kind of number that changes a build-versus-buy calculation for in-house tooling. And with Enterprise Frontier Safeguards arriving this fall, regulated companies that had ruled out Fable over data-residency concerns have a real reason to look again.