Claude Fable 5.1 and Mythos 5.1: What Changed
Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 on 1 September 2026. They are the same underlying model with different safeguards and access rules. Fable 5.1 is generally available. Mythos 5.1 is reserved for vetted cybersecurity and life-sciences work.
The useful business story is narrower than the launch headline. Fable 5.1 is better at long-running coding, multistep research, documents, spreadsheets, slides, dense visual material, and computer use. Its headline input and output prices have not changed. The cost reduction comes from much cheaper cache reads, and an API migration has three breaking changes that agent builders need to test before switching.
The Upgrade Is About Finishing Long Work
Anthropic positions Fable 5.1 for demanding reasoning and long-horizon agentic work. Its own documentation still recommends starting most workloads on Claude Opus 5, then moving to Fable 5.1 when evaluations show that Opus at higher effort is not enough. That is a useful correction to the instinct to put the newest model everywhere.
The capability gains concentrate in six areas: agentic coding across long sessions, document and office work, multistep research, vision over dense charts and PDFs, reasoning across a one-million-token context, and browser or desktop operation. The model can produce as many as 128,000 output tokens and uses adaptive thinking at every effort level.
Anthropic reports substantial gains on its launch evaluations. Fable 5.1 scored 52.6% on Terminal-Bench-Science 0.1 against 24.7% for Fable 5, and 73.4% on CursorBench 3.2 against 70.5%. Those are vendor-run launch numbers, not a guarantee for your workload. The more important claim to test is that medium effort can roughly match Fable 5 at lower cost. If that holds on your cases, the upgrade can improve both quality and unit economics.
The List Price Stayed Put, The Effective Cost Did Not
Base pricing remains $10 per million input tokens and $50 per million output tokens. A five-minute cache write costs $12.50 per million tokens, and a one-hour write costs $20. The change is the read price: $0.25 per million cached input tokens, one quarter of the Fable 5 rate.
Anthropic estimates that this reduces a typical Fable workload by about 25%, and a context-heavy agentic workload by as much as 45%. The estimate comes from four weeks of actual usage in August 2026 at the default effort settings. Your saving depends on architecture. A stateless request that never reuses a prefix gets little benefit. An agent that repeatedly reads the same repository map, tool definitions, instructions, or research corpus can get much more.
Effort is now a first-class cost control. Fable 5.1 supports low, medium, high, xhigh, and max effort, including a beta option to change effort within a conversation without invalidating the prompt cache. Start at high, evaluate the other levels against the same test set, and route the easy turns down. Do not assume that an effort label consumes the same amount of thinking as it did on Fable 5.
Three API Changes Can Break An Existing Agent
Changing the model ID from `claude-fable-5` to `claude-fable-5-1` is not the whole migration.
First, forced tool use is no longer supported. `tool_choice` with `any` or a named tool returns a 400 error. Keep automatic tool choice, use strict schemas or structured outputs for valid JSON, and state clearly in the prompt when the tool must be used.
Second, earlier Claude models cannot read thinking blocks created by Fable 5.1. A router that moves a conversation back to Opus, Sonnet, or Fable 5 will lose those blocks. The API drops unreadable thinking rather than billing for it, but the lower-tier model no longer has that reasoning context.
Third, changing the system prompt, tool list, or an earlier message can invalidate later thinking blocks. Agent harnesses that rebuild or edit history between turns need particular attention. Treat a conversation as append-only, pass thinking blocks back unchanged, and use the new mid-conversation mechanisms when instructions or tools must change.
Other changes are additive: per-message effort, turn-scoped system messages, readable progress updates between tool calls, and content provenance. Text generated by Fable 5.1 and Mythos 5.1 carries Anthropic's statistical watermark. Supported media returned through the Files API can carry signed C2PA credentials.
Fable And Mythos Are Access Policies, Not Two Capability Tiers
Fable 5.1 and Mythos 5.1 have the same core capabilities and pricing. The difference is which safeguards apply.
Fable 5.1 is available through the Claude API and Anthropic's supported cloud platforms. Its classifiers can return an HTTP 200 response with `stop_reason: "refusal"`. Production clients should handle that state explicitly and either stop, ask for a safer formulation, or use a permitted fallback. Anthropic says its revised cyber safeguards cause around 60% fewer interventions per Claude Code session than the original Fable 5 safeguards. Defensive vulnerability discovery is now allowed, while exploit generation and several dual-use activities remain redirected or restricted.
Mythos 5.1 is not a premium switch in the public model picker. It is identical to Fable 5.1 but has safeguards designed for approved professional work in cybersecurity and life sciences. Access runs through Anthropic's verification programs and is initially limited to selected organizations.
Both models normally carry a 30-day data-retention requirement and are not available under zero data retention unless Anthropic expressly authorizes it. Anthropic has announced Enterprise Frontier Safeguards, where monitoring data stays in customer-controlled cloud infrastructure, with a phased rollout beginning later in 2026. Eligible customers can receive zero data retention before that rollout. Procurement teams should treat this as an eligibility question, not as a default contract term.
Who Should Move Now
Move a Fable 5 workload to 5.1 if it depends on long agent runs, complex research, dense business documents, or repeated cached context. The lower cache-read price makes this the clearest upgrade case.
Test before moving if your harness forces tools, edits conversation history, routes down to older Claude models, or relies on constant narration between tool calls. The migration is manageable, but these behaviors are visible in production and deserve targeted tests.
Stay on Opus 5 or a smaller tier when it already meets the acceptance criteria. Anthropic itself recommends Opus 5 as the starting point for most workloads, at half Fable's base input and output price. Fable 5.1 should win a measured job, not a naming contest.
The release does not change the durable architecture: keep model IDs in configuration, log the model and stop reason for every output, preserve conversation history correctly, and evaluate cost per passed case. Our free AI API cost estimator lets you compare normal input, cache writes, cache reads, and output before changing a production route.
---
Sources: Anthropic's Fable 5.1 and Mythos 5.1 announcement and the Claude Platform migration and pricing reference, consulted 1 September 2026. Benchmark and savings figures are Anthropic's measurements and estimates, not independent results.