Claude Fable 5.1 and Mythos 5.1 cut cache reads 75%

On September 1, 2026 Anthropic shipped Claude Fable 5.1 and Claude Mythos 5.1. They are the same weights with different safeguards: Fable 5.1 is generally available; Mythos 5.1 is limited to trusted-access programs for cybersecurity and life sciences. List prices did not move ($10 input / $50 output per million tokens). Cache reads did: $0.25 per million, down 75% from Fable 5's $1.00. Anthropic says that cut makes typical token-billed workloads about 25% cheaper than Fable 5, and highly agentic work up to about 45% cheaper, measured on four weeks of August 2026 usage at default effort.

This piece is for Claude Code users, API callers, and anyone who still has claude-fable-5 pinned. It covers what shipped, a before/after table from Anthropic's launch post, what that does to a session bill, and the exact commands to update Claude Code, switch models, and set a default. It is not a recap that tells you to go read VentureBeat.

What shipped

Fable 5.1 is the public model. Mythos 5.1 is the same model with lighter cybersecurity and biology safeguards, available only through Anthropic's trusted access programs (a Cyber Verification Program and a Life Sciences Verification Program run with the US government). Anthropic's launch post is https://www.anthropic.com/claude-fable-and-mythos-5-1. Both are on the Claude API, Amazon Bedrock, and Google Cloud Vertex AI. In Claude Code, Fable 5.1 needs version 2.1.255 or later; 2.1.257 makes the fable alias resolve to 5.1. Effort defaults to high in Claude Code and on the API, and to medium in Claude Cowork and on claude.ai.

Enterprise Frontier Safeguards (EFS) is the other launch item: customer-controlled storage that Anthropic describes as equivalent to zero data retention, rolling out to enterprise in phases later this fall. Until EFS is live, eligible customers can run Fable 5.1 with zero data retention. Cybersecurity safeguards cut false positives by 60% versus the prior generation, in part because Fable 5.1 is allowed to discover vulnerabilities but not to develop exploits.

What improved

The table below is Anthropic's own launch accounting. Input and output list prices are unchanged. The only price line that moved is cache reads. Benchmark rows are from the same post (Fable 5.1 with production safeguards on).

Those benchmark jumps are why a 75% cache-read cut can still raise cost per task if you pin max effort and the model writes twice as many output tokens. Anthropic's 25% / 45% savings figures were measured at default effort on mixed Claude Enterprise, Claude Code, and API traffic in August 2026. Your bill follows cache-hit rate and effort, not the headline 75% alone.

What you gain

If you already cache the system prompt, tool schemas, and repo prefix — which Claude Code does every turn — Fable 5.1 makes those rereads four times cheaper than Fable 5. A long agentic session is mostly cache reads; that is why Anthropic's agentic number is 45% rather than 25%. At low or medium effort, Anthropic says Fable 5.1 lands at or above Fable 5 quality for less money. Cognition said it was moving Devin's Opus 5 traffic onto Fable 5.1 on launch day because cache-read pricing finally made a Fable-class model economical for code review.

You also get a later knowledge cutoff (June 2026 vs January 2026) and the qualitative change partners described: fewer shortcut fixes, longer unattended runs, more readable traces. None of that replaces measuring output tokens per turn on your own harness. GitHub issue 91623 recorded Fable 5.1 emitting about 2× the output tokens per turn of Fable 5 on one Max 20x machine after launch — savings on cache can be eaten by longer completions if you leave effort at high or max.

How to get it

Update Claude Code first so the picker lists Fable 5.1. Native installs use claude update. If you installed via npm, update the global package, then confirm you are on 2.1.255 or later.

npm install -g @anthropic-ai/claude-code
claude update
claude --version

Switch the current session. After v2.1.255 the fable alias points at 5.1. Use the ID if you want to be explicit.

/model fable

/model claude-fable-5-1

Pin a new session from the shell, or persist a default. The /model picker also saves the choice. ANTHROPIC_MODEL overrides for one shell. Do not invent a config key — if claude config set model is not in your build, use the picker or the env var.

claude --model fable
export ANTHROPIC_MODEL=claude-fable-5-1
claude config set model claude-fable-5-1

API callers change the model string from claude-fable-5 to claude-fable-5-1. Cache-write prices are unchanged ($12.50 / 5-minute, $20 / 1-hour). Mythos 5.1 is not a /model flag on a regular Max plan; it is a trusted-access SKU. If the picker says Fable requires usage credits, that is the Max 50% Fable cap, not a missing update.

What to watch next

Confirm EFS availability on your org's Anthropic console this fall; until then the zero-retention path is the documented stand-in for eligible customers. Watch whether /model fable stays on 5.1 or gets moved again. The September 14 Claude Code weekly-limit raise (25% for Pro, Max, Team, and seat-based Enterprise) is a separate change from this cache-price cut. If a Max week still evaporates in a day, log cache hits and output tokens per turn before you file a metering bug.

Verify every number on Anthropic's launch page, not on a syndication blurb. Pin claude-fable-5-1 in staging, run your own eval pack at high and at medium effort, and only then change the default. Mythos 5.1 stays behind verification programs; do not promise it to a team that is not enrolled.