# Claude Fable 5.1 Launches With Cheaper Caching, Three Breaking Changes

Page: https://hirantha.me/blog/claude-fable-51-launches-with-cheaper-caching-three-breaking-changes
Published: 2026-09-08
Written by: hirantha.me Newsroom (AI-generated). An AI model wrote this from the sources listed below; how it is made: https://hirantha.me/about/newsroom
Topics: anthropic, claude, ai models, llm pricing, developer tools

> Anthropic's Claude Fable 5.1 cuts cached-token costs by 75 percent, but developers upgrading from Fable 5 must fix three breaking API changes.

## What shipped

Anthropic released two new models on September 1: Claude Fable 5.1, available to every API customer, and Claude Mythos 5.1, an identical model with relaxed safety restrictions offered only to vetted organizations inside the company's Project Glasswing program for cybersecurity and life-sciences work. Both replace Fable 5 and Mythos 5 respectively and share the same specifications: a 1 million token context window at standard per-token pricing across the whole window, 128,000 maximum output tokens, and reasoning ("thinking") that is always on rather than optional.

Anthropic pitches Fable 5.1 as the model to reach for on long-running agentic coding, multistep research, and document, spreadsheet, and slide work, ahead of the more general-purpose Opus 5 when evaluations on Opus 5 fall short at higher effort settings.

## The price didn't move, but the bill will

Input tokens stay at $10 per million and output at $50 per million, unchanged from Fable 5. The one price that moved is the cache read rate, which drops from $1.00 to $0.25 per million tokens, a 75 percent cut. Anthropic frames the practical effect as roughly 25 percent lower cost on typical workloads and up to 45 percent lower on agentic workloads, since those tend to re-read the same cached context repeatedly across a long session. Cache write prices and the 512-token minimum cacheable prompt length are unchanged.

## Three changes that will break existing integrations

Anthropic's migration documentation flags exactly three breaking changes for anyone already calling Fable 5:

1. **Forced tool use is gone.** Setting `tool_choice` to `any` or to a specific named tool now returns an HTTP 400 error. Because thinking is always on in these models, a forced tool call would skip that reasoning step and push the model's working-out into the tool arguments instead, which Anthropic says lowers argument quality. The fix is to keep `tool_choice` at its default `auto` setting combined with strict tool use, or to move schema enforcement to structured outputs.
2. **Thinking blocks only flow one way.** Every reasoning block records which model produced it. Fable 5.1 can read thinking blocks left by older Claude models, but no older model can read blocks Fable 5.1 produces. A conversation that later falls back to an earlier model loses its reasoning for any turns that ran on Fable 5.1.
3. **Editing history invalidates later reasoning.** Changing anything that comes before a Fable 5.1 thinking block, meaning the system prompt, the tools list, or an earlier message, causes the next request to fail, unless the caller opts into silently dropping the affected block instead. This check applies to accounts created on or after August 31, 2026; older accounts are only warned, not blocked, until they opt in. Anthropic's guidance is to treat a conversation as append-only and use new turn-scoped system messages for one-off instructions instead of editing history directly.

## What else is new

Five additions are opt-in rather than breaking. Developers can now change the reasoning effort level mid-conversation without invalidating the prompt cache; scope a system message to a single turn so it stops applying once the user replies again; request visible, plain-text progress updates between tool calls instead of Anthropic's default of hiding that reasoning entirely; and rely on the lower cache price above. Text the model generates also now carries Anthropic's statistical watermark by default, and images, audio, or video it produces through the code execution tool carry signed C2PA content-credential metadata when retrieved through the Files API.

Anthropic also documents behavior that changed without any code change at all: Fable 5.1 tends to issue tool calls one at a time in agent loops where Fable 5 often batched several, it narrates fewer progress updates during long tool runs, it reaches for memory instead of a search tool more often at low effort settings, and its default prose runs denser with less markdown formatting than earlier Claude models used.

## The numbers behind the upgrade

On Anthropic's own benchmark disclosures at maximum effort, Fable 5.1 scored 55.8 percent on Terminal-Bench 4.0 and 52.6 percent on the newer Terminal-Bench-Science 0.1; the more capable Mythos 5.1 scored 60.9 percent on Terminal-Bench 4.0. Fable 5.1 also posted 73.4 percent on CursorBench 3.2.0, 65.0 percent on Humanity's Last Exam with tool use enabled, and 77.9 percent on the partial OSWorld 2.0 computer-use benchmark, according to a summary of the launch materials published by MarkTechPost.

Among the customer testimonials Anthropic published at launch, Red Hat's Josh Boyer said the model "correctly identified root cause of every broken build tested across all effort levels," while Shopify's Ben Lafferty described it as "more comfortable with long unattended work, keeps records, reprioritizes, picks up where it left off."

## What it means for developers

Anyone with existing Fable 5 integrations should budget migration time before switching the model string, not after: search the codebase for forced `tool_choice` settings, audit whether the integration edits or reorders past turns in the messages array, and re-run evaluations rather than assuming parity given the documented behavior shifts around tool-call batching and formatting. Teams running long, cache-heavy agent sessions stand to see the biggest cost drop from the cache-read price cut, while teams with simple, short-lived requests will barely notice the pricing change either way.

## Sources

This article is based on Anthropic's official announcement and platform documentation, and on independent benchmark reporting, listed below.

## Questions

**What is the difference between Claude Fable 5.1 and Claude Mythos 5.1?**
They are the same underlying model with different safeguards: Fable 5.1 is generally available to all API customers, while Mythos 5.1 has relaxed restrictions and is offered only to vetted organizations in Anthropic's Project Glasswing program.

**What breaks when upgrading from Claude Fable 5 to Fable 5.1?**
Forced tool_choice settings now return an error, thinking blocks can no longer be read by older models once produced by Fable 5.1, and editing an earlier turn in a conversation invalidates later reasoning blocks.

**How much cheaper is Claude Fable 5.1 to run?**
Input and output pricing are unchanged, but cache reads cost 75 percent less, which Anthropic estimates saves roughly 25 percent on typical workloads and up to 45 percent on agentic ones.

**Does Claude Fable 5.1 behave differently even without code changes?**
Yes: Anthropic says it tends to make one tool call per turn instead of batching several, narrates fewer progress updates during long tasks, and defaults to denser, less-formatted prose.

**Where can I access Claude Fable 5.1?**
It is available on the Claude API, Amazon Bedrock, Claude Platform on AWS, Google Cloud, and Microsoft Foundry; Claude Mythos 5.1 requires approval through Project Glasswing.

## Sources

- [Introducing Claude Fable 5.1 and Claude Mythos 5.1](https://www.anthropic.com/claude-fable-and-mythos-5-1), Anthropic
- [What's new in Claude Fable 5.1](https://platform.claude.com/docs/en/models/fable-5-1/whats-new-fable-5-1), Claude Platform Docs
- [Anthropic Releases Claude Fable 5.1 and Claude Mythos 5.1: 52.6% on Terminal-Bench-Science and 75% Cheaper Cache Reads](https://www.marktechpost.com/2026/09/01/anthropic-releases-claude-fable-5-1-and-claude-mythos-5-1-52-6-on-terminal-bench-science-and-75-cheaper-cache-reads/), MarkTechPost