The Agent Lab
Ashrey

Claude Fable 5.1 and Mythos 5.1: 75% Cheaper Cache Reads, Forced Tool Use Removed, Watermarked Outputs

Anthropic shipped Fable 5.1 and Mythos 5.1 on September 1 with three breaking API changes, 75% cheaper cache reads, and mandatory EU AI Act watermarks. Here's what developers building on the API need to change.

On September 1, Anthropic released claude-fable-5-1 and claude-mythos-5-1 — the first model pair under the new Fable/Mythos naming scheme. Base token pricing stays at $10 per million input and $50 per million output. But three breaking API changes, a 75% cut to cache-read costs, and mandatory statistical watermarks on all text output mean existing integrations need attention.

Fable 5.1 is generally available on the Claude API, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry. Mythos 5.1 remains invite-only through Project Glasswing for vetted cybersecurity defenders and life-sciences researchers. Retirement for Fable 5 is "not sooner than September 1, 2027."

The Three Breaking Changes

If you already call Fable 5, three changes return errors where they previously succeeded:

  • Forced tool use is gone. Setting tool_choice to {"type": "any"} or a named tool now returns a 400 error. The model decides when to call tools.
  • Earlier models cannot read Fable 5.1 thinking blocks. A router that falls back from Fable 5.1 to an older model mid-conversation drops those blocks. Dropped blocks are not billed.
  • Editing earlier turns invalidates thinking blocks. Changing the system prompt, tools, or any message before a Fable 5.1 thinking block can return a 400 or silently drop the block if you opt in. Treat history as append-only.

The forced-tool-use removal is the most immediate migration. Code that pins tool_choice: { "type": "any" } or { "type": "tool", "name": "..." } to guarantee a tool call will now error. The documented fix: stop forcing tools. Let the model decide. If you need steering, use the new per-message effort control or turn-scoped system messages (both beta).

Cache Reads: $0.25/M Instead of $1.00/M

The headline economics change: prompt cache reads dropped from $1.00 to $0.25 per million tokens — a 75% reduction. This makes long-context, multi-turn workflows dramatically cheaper when the same prefix is reused.

For a 200K-token cached prompt read 50 times, the cache-read bill falls from $50 to $12.50. Base input/output pricing is unchanged, so the savings only apply to the cache_read line item.

This matters most for agents that reuse large system prompts, few-shot examples, or codebase context across many turns. The new rate is 0.025× the standard input rate instead of the usual 0.1×.

Mandatory Watermarks: EU AI Act Compliance

Every text response from Fable 5.1 and Mythos 5.1 now carries Anthropic's statistical watermark — a numerical signal detectable by a private API (in preview for eligible parties). This is not optional. It applies on every platform where the model runs: the Claude API, Bedrock, Vertex, Foundry, and claude.ai.

Images and video generated through the code execution tool carry signed C2PA Content Credentials when retrieved through the Files API.

The watermark does not change the meaning, quality, or readability of the output. But it means any text produced by these models is cryptographically attributable. If you're building a product that resells or redistributes model output, factor this into your compliance review.

Five Additive Beta Features

Beyond the breaking changes, five new beta capabilities shipped:

  • Per-message effort control — Set thinking: { "type": "enabled", "budget_tokens": N, "effort": "high|medium|low" } per message to control compute mid-conversation without busting the prompt cache.
  • Turn-scoped system messages — A system message that applies only to the next turn. Use { "role": "system", "content": "...", "cache_control": { "type": "ephemeral" } } to inject context for one exchange without rewriting the whole system prompt.
  • Visible progress updates — thinking: { "display": "updates" } surfaces the model's reasoning as readable text between tool calls instead of hiding it.
  • Lower cache-read price — Already covered above.
  • Content provenance — The watermark and C2PA credentials on generated media.

These are additive — they don't break existing code — but they change how you can steer the model. The effort control and turn-scoped system messages are the practical replacements for the removed forced-tool-use pattern.

Migration Checklist

  1. Grep for tool_choice. Any type: "any" or named-tool choice will 400. Remove or replace with effort control.
  2. Audit history editing. If your code rewrites the conversation (edits prior messages, swaps system prompts, changes tool sets mid-thread), Fable 5.1 thinking blocks will invalidate. Treat history as append-only.
  3. Check router fallbacks. A fallback from Fable 5.1 to an older model drops thinking blocks. If your router does this, expect lost context or 400s.
  4. Update cost models. Cache reads are now $0.25/M. Recalculate any workload that leans on prompt caching.
  5. Review watermark obligations. If you redistribute output, the watermark travels with it. The detection API is in private preview; contact Anthropic if you need access.

What Fable 5.1 Improves

The model scores 52.6% on Terminal-Bench-Science (up from 24.7% for Fable 5), with gains in long-horizon coding, research, and computer use. Anthropic also claims roughly 60% fewer cybersecurity false positives — Fable 5.1 can discover vulnerabilities but declines exploit generation, routing those requests to Opus.

Mythos 5.1 loosens the cyber/biology filters for vetted users. Both models share the same base capabilities; the difference is the safeguard tier.

What to Watch

  • Mythos access. Project Glasswing is invite-only. No public waitlist timeline.
  • Watermark detection API. Private preview only. General availability date unannounced.
  • Eval stability on new betas. Effort control, turn-scoped system messages, and visible updates are beta. Results vary between runs; the --runs flag in plugin evals mitigates variance but increases cost.
  • Opus routing for exploit requests. Fable 5.1 redirects exploit-generation to Opus. If your workflow expects a refusal, you may get an Opus response instead.

Practical Next Steps

  1. Run your test suite against claude-fable-5-1. Expect 400s on forced tool use and any history-rewriting patterns.
  2. Replace tool_choice: { "type": "any" } with per-message effort control (thinking: { "effort": "high" }) or turn-scoped system messages.
  3. Recache your prompt templates to benefit from the 75% cache-read discount.
  4. If you distribute model output, add a watermark detection step to your pipeline once the API is generally available.

If this was useful, you can support my open-source work on Ko-fi or check out my services.