Claude Fable 5.1 and Mythos 5.1 are now generally available: same price, much cheaper caching

On 1 September 2026, Anthropic released Claude Fable 5.1 and its restricted sibling Claude Mythos 5.1. The headline pricing is unchanged, $10 per million input tokens and $50 per million output tokens, but a significant cut to cache read pricing means most teams building with the API will pay noticeably less without changing a line of code. There are also a handful of breaking API changes worth knowing about before you upgrade.

What actually changed on pricing

The number that matters most is the cache read price. On every previous Claude model, a cache hit costs 10% of the standard input price. On Fable 5.1, that drops to 2.5%, or $0.25 per million tokens, a 75% reduction.

Anthropic derived its savings estimates from four weeks of real August 2026 usage across Claude Enterprise, Claude Code, and the API. For typical workloads, the reduction comes out at around 25%. For heavily agentic pipelines, where the model reads long cached contexts repeatedly across many turns, the saving can reach around 45%.

If your workload has no reusable prefix, you save nothing, because you were never paying for cache reads to begin with. But if you are running multi-step agents, coding assistants with large shared system prompts, or any workflow where the same context gets referenced repeatedly, this is a genuine cost reduction that arrives automatically.

Claude Sonnet 5 also gets a small pricing footnote: the introductory $2/$10 per million input/output pricing introduced at launch has been made permanent. The scheduled increase to $3/$15 from 1 September 2026 will not happen.

Where you can use it

Fable 5.1 is available from 1 September 2026 on the Claude API, Amazon Bedrock, and Google Cloud Vertex. On claude.ai, it is available to Pro, Max, Team, and Enterprise subscribers. Free plan users do not get access to Fable-class models.

For Max-tier and premium seats, up to 50% of weekly usage limits can be spent on Fable models at no additional charge. Pro-tier and standard Team or Enterprise seats use API-rate usage credits.

The model carries a 1 million token context window, 128K maximum output tokens, adaptive thinking on by default, and a knowledge cutoff of June 2026. Benchmark performance is strong: 55.8% on Terminal-Bench 4.0 for coding, up from Fable 5’s score, and ahead of both Opus 5 at 52.3% and OpenAI’s GPT-5.6 Sol at 37.3%.

Claude Mythos 5.1 and Project Glasswing

Fable 5.1 and Mythos 5.1 are the same underlying model. The difference is the safety layer sitting in front of it. Fable 5.1 runs Anthropic’s full production safety classifiers. Mythos 5.1 runs with cybersecurity and biology safeguards relaxed, and is available only to organisations vetted through Project Glasswing.

Glasswing has two invitation-only verification tracks. The Cyber Verification Programme covers defensive security professionals. The Life Sciences Verification Programme, developed alongside the US government, is launching as an invite-only beta for advanced life sciences researchers. As of 1 September 2026, Mythos 5.1 is available to a set of US organisations, with Glasswing being extended to approximately 150 new organisations across more than fifteen countries. No date has been given for broader access.

Practically speaking, Mythos 5.1 scores 60.9% on Terminal-Bench 4.0 versus Fable 5.1’s 55.8%, reflecting the absence of production safeguard interventions. The cybersecurity filter also fires around 60% less often per session in Claude Code for vetted users. The model will identify software vulnerabilities for defensive purposes, but will not write exploits.

If you work in security research or life sciences and think your organisation qualifies, the place to start is the Glasswing programme page.

Anti-distillation enforcement for new API accounts

From 1 September 2026, API accounts created on or after that date face a new restriction. The API now verifies that a thinking block is sent back with the same system prompt, tools, and messages that produced it. If they do not match, the request returns an error.

This closes off a publicly documented distillation technique where researchers would illicitly extract a model’s reasoning by editing the prior conversation context while keeping the thinking transcript intact. Anthropic traced over 16 million Claude exchanges to distillation campaigns run through roughly 24,000 fake accounts in February 2026, so the motivation is clear.

Accounts created before 1 September 2026 are not affected immediately. Mismatches are recorded but not enforced unless you explicitly opt in via thinking.block_binding.prefix_mismatch_behavior. Anthropic has said the restriction will apply to all accounts at future model releases, so this is the direction of travel regardless of when your account was created. Mythos 5.1 does not run the check.

If you build and distribute tools, Anthropic specifically flags that your API key is probably on an older account, but your users’ keys on newer accounts will hit the enforcement before you do.

Breaking changes if you are already on Fable 5

Three changes will break existing integrations:

1. Forced tool use now returns an error. If you pass tool_choice: {type: "any"} or similar, the API will reject it rather than silently falling back.
2. Earlier models cannot read Fable 5.1 thinking blocks. If you are routing thinking block outputs into a different model downstream, that will fail.
3. Editing earlier turns invalidates thinking blocks. Any integration that rewrites mid-conversation turns, including context compaction or injected system reminders, will need adjustment.

Five additive changes arrive alongside these: per-message effort control (beta), turn-scoped system messages (beta), readable progress updates between tool calls (beta), the lower cache read price, and content provenance watermarking.

Content provenance and watermarking

All Fable 5.1 output carries a statistical text watermark and C2PA credentials on generated files. This is not optional. It complies with the EU AI Act for models released after 2 August 2026 and contains no user data. The watermark does not alter the visible text.

The short version for your planning

If you are already using Fable 5 in production, check whether your integration edits earlier turns mid-conversation. If it does, you will need to adjust before migrating to 5.1. If your integration is straightforward, you should see cache-driven cost reductions with no changes required.

If you are in defensive security or life sciences research, Glasswing verification is worth pursuing. The capability difference between Fable 5.1 and Mythos 5.1 is meaningful for those specific domains.

If you are evaluating Claude for the first time, Fable 5.1 is the most capable generally available Claude model, with the most recent knowledge cutoff of any Claude release.