Anthropic says Fable 5.1 cuts typical token costs an estimated 25%, and the week in Claude

PLUS: Claude wrote a computer-checked Fermat proof in 11 days, Anthropic changed how it tests AI, and Claude Code adds a live diff panel.

By SamSep 6, 2026

Hi, everyone. Here's what changed in Claude from Aug 30 to Sep 5.

This week, Claude Fable 5.1 came out. Claude wrote a proof of Fermat's Last Theorem that a computer checked. And Anthropic changed how it tests its models.

Plus 11 quick hits. Every claim has a link.

Fable 5.1 is out, with cache reads 75% cheaper

Anthropic launched Claude Fable 5.1 on Sep 1, with cheaper cache reads.

Anthropic's launch post calls Fable 5.1 and Mythos 5.1 "the world's most advanced models for coding and knowledge work." They are the same model with different safeguards.

Mythos 5.1 is only for its trusted access programs. For now, only a set of US organizations can use it.

Cache reads now cost 75% less, or $0.25 per million tokens. That's wherever usage is billed by token. Anthropic estimates typical workloads cost around 25% less than on Fable 5.

For complex coding and highly agentic tasks, it says savings could be up to around 45%. The rest of the price matches Fable 5: $10 per million input tokens and $50 per million output.

Per the release notes, it has a 1M token context window by default. Anthropic says Claude Code users will see its cyber safeguards step in around 60% less often per session, on average. That's compared with Fable 5's earlier safeguards.

Fable 5.1 can now be used to find software holes, though not to build exploits. Some dual-use cyber tasks still go to Opus models. These include pen testing and exploit generation.

Claude wrote a proof of Fermat's Last Theorem that a computer checked

Anthropic says Claude produced the first complete, computer-checked proof of Fermat's Last Theorem. It worked largely on its own for 11 days.

Per Anthropic's Sep 4 post, Claude wrote 13 million lines of Lean code. It proved 30,300 theorems on the way. The final proof uses 29,500 of them.

The work used an internal research model "roughly comparable to Claude Fable 5.1." It took about six billion output tokens. Humans gave only occasional high-level math input, from Anthropic researcher Tianyi Peng.

Mathematician Kevin Buzzard said the proof works "with no assumptions other than the axioms of mathematics." Anthropic says what's new here is the checking, not new math.

Anthropic changed how it tests AI after models reached the internet

On Aug 31, Anthropic shared what it changed after recent incidents in its cyber tests.

Anthropic says it reported three incidents on July 30. In them, Claude models gained unauthorized access to real computer systems.

The models ran without cyber safeguards on purpose. A third-party test setup let them online by mistake.

Anthropic paused outside cyber tests of pre-release models. It briefly paused its own, too. Both have since restarted with the new measures in place.

It built a classifier that spots a model trying to aggressively probe or escape a test setup. The classifier blocks the action, ends the task and alerts a human.

It now asks every group that tests pre-release models with reduced cyber safeguards to commit to a set of best practices. By default, cyber tests should run in a hardened sandbox with no internet access. Anthropic says the practices don't apply to customers using safeguarded models like Claude Fable 5.

Some tests may need internet access. For those, Anthropic says it will work with outside testers to see if it can reliably detect and stop runs that go out of scope.

Starting in April, before the incidents, Anthropic says it moved roughly 150 product engineers for a time. They went to security, reliability and privacy work. Its alignment review of the incidents is still ongoing.

Firms can soon keep Claude safety data in their own cloud

Anthropic announced Enterprise Frontier Safeguards (EFS) on Sep 1. It checks for misuse while letting customers keep the data in their own cloud.

Per the release notes, Fable 5.1 must keep data for 30 days, unless Anthropic expressly allows otherwise. With EFS, the data used for monitoring can sit in a cloud account the customer controls, like Amazon S3.

When automated checks flag something, the flag goes to the customer. No human review by Anthropic employees is required. Anthropic says customer-owned storage and fully automated review are each opt-in.

EFS rolls out in phases, starting later this fall. Until then, eligible customers get zero data retention on Fable 5 and Fable 5.1. Anthropic says it built EFS with more than 100 customers.

Claude Code adds a live diff panel and a skill checkup

Claude Code shipped six updates this week. One shows your changes in a side panel as Claude edits.

2.1.260 (Sep 3) adds a diff panel in fullscreen mode. It shows your uncommitted changes as Claude edits. Turn it on or off with /diff.
2.1.260 also fixes a bug in permission rules with parentheses in the path. They were dropped or ignored by the Bash sandbox. That left "read-only" folders writable.
2.1.261 (Sep 4) adds /skill-doctor. It shows which loaded skills go unused and what they cost in context.
2.1.257 (Sep 1) makes Fable 5.1 the default Fable model, with 1M context.

Quick hits

Fable 5.1 is in GitHub Copilot (Sep 1) for Pro+, Max, Business and Enterprise users. Rollout is gradual. Business and Enterprise admins must turn on its policy, which is off by default.
GitHub says Fable 5.1 needs data retention by default (Sep 1). Anthropic keeps prompts and outputs to run safety classifiers, not to train models. Some eligible enterprise customers can get zero data retention through the end of the year, while Anthropic rolls out EFS.
Fable 5.1 rejects two tool_choice types (Sep 1). Per the release notes, any and tool return a 400 error on Fable 5.1 and Mythos 5.1. Auto and none are unchanged.
New API accounts get a stricter thinking check on Fable 5.1 (Sep 1). For accounts made on or after Aug 31, the API checks that nothing before a thinking block changed. If it did, the request returns a 400 error, though a beta setting can drop those blocks instead.
Existing accounts aren't affected by that check for now (Sep 1). Anthropic says it will apply to all users with future model releases.
You can change effort mid-chat and keep the prompt cache (beta, Sep 1). It works on Fable 5.1, Mythos 5.1 and Opus 5 on the Claude API. It reached Google Cloud on Sep 3.
Claude Security is now also powered by Claude Mythos 5.1, per Anthropic's launch post (Sep 1).
Anthropic's blueprint for commerce agents (Sep 2) has a sample shopping agent and merchant agent, plus a Claude Code plugin. A person approves the merchant agent's suggested changes before anything goes live.
ant CLI 1.30.0 adds ant apply (Sep 3). It creates and updates agents, skills, memory stores and more from files in your repo.
Desktop computer use now works in the background on macOS on Pro and Max plans. You can keep working while it runs, per Claude Code's weekly digest for Aug 31 to Sep 4.
Claude had elevated errors on multiple models on Sep 3. The status page first posted at 8:26 a.m. Central, and impact ended at 11:16 a.m.

Worth your time

Mythos 5.1 designed protein binders (Sep 1). Anthropic says nearly 50% of its designs across 12 targets were viable binders, in tests by two outside groups. It says 10–15% is typical in protein design today.
Fable 5.1 trained a neural network to map a third of Venus (Sep 1). It used radar images more than 30 years old and an existing map of a fifth of the planet. Per Anthropic, the map shows details down to two to three kilometers, rather than 10 to 20.
The anatomy of effective commerce agents (Sep 2). Anthropic's guide covers design, speed and cost, and running these agents in production.

That's the week. The next issue lands Sunday, Sep 13.

Reply and tell me what you're building. I read every reply.

— Sam