The AI Catchup - September 1, 2026

Welcome back. One story this issue, because it deserves the room: Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 this morning. If you read our June issue on Fable 5, you will remember the three things people complained about: the bill, the mandatory 30-day retention, and safeguards that blocked ordinary security work. Fable 5.1 answers all three with numbers, and it is faster and better on every benchmark Anthropic published. We go through what changed, what it costs, what still redirects to Opus, and what to do about it today.

Let us get into it.

At a Glance

This Week in AI

Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 on September 1. They are the same model with different safeguard levels, exactly the split Anthropic drew in June. Fable 5.1 is generally available now; Mythos 5.1 is restricted to vetted cyberdefenders and life scientists.

The price story is a cache-read story. Input and output stay at $10 and $50 per million tokens. Cache reads, where the model reuses context it has already processed, fall 75% to $0.25 per million. Anthropic measured four weeks of real August usage at default effort and found a typical workload across Claude Enterprise, Claude Code, and the API costs about 25% less on Fable 5.1 than on Fable 5, and a context-heavy, tool-heavy agentic workload about 45% less, because cache reads make up most of that bill. If you run long agentic sessions, the saving is on the part of your invoice that was growing fastest.

The capability story is a clean sweep on Anthropic's own table. Fable 5.1 beats Fable 5, Opus 5, and GPT-5.6 Sol on every benchmark where a comparison is shown. The headline numbers against Fable 5: Terminal-Bench-Science 52.6% vs 24.7%, Terminal-Bench 4.0 55.8% vs 42.0%, AutomationBench 31.4% vs 17.1%, CursorBench 3.2.0 73.4% vs 70.5%, and GDPval-AA v2 1853 vs 1723. These were run with production safeguards on, and Anthropic says that likely pulls the Fable numbers down, not up. Mythos 5.1 posts 60.9% on Terminal-Bench 4.0, five points above Fable 5.1, which is the one place the table shows what the safeguards cost.

The partner reports are more useful than the table for deciding whether to switch. Every's Dan Shipper measured Fable 5.1 at about twice the speed of Opus 5 on half the tokens and called it "Fable-level intelligence, Opus-level price, Sonnet-speed." Cognition is moving Devin's Opus 5 traffic to Fable 5.1 on launch day, starting with code review, because it matched or beat Fable 5 at a lower cost per task. Rogo saw Fable 5 accuracy on its finance benchmark with 20% fewer tokens. The pattern is the same everywhere: the model does the job in fewer tokens and fewer turns, and the cheaper cache reads compound with that.

What the Safeguards Now Allow

This is the change security teams have been waiting for since June. Fable 5.1 may now be used to identify software vulnerabilities. Anthropic frames that as the defensive work that makes software safer, and it pairs the policy change with more precise classifiers. The result it reports: Claude Code users should see about 60% fewer interventions per session from cyber safeguards than on Fable 5.

The boundary is still there, and it is worth being precise about it. Penetration testing, exploit generation, and binary-based vulnerability scanning still redirect to Opus models. So Fable 5.1 will find the bug in your code. It will not write the exploit for it. If that is your job, you want Mythos 5.1 through the Cyber Verification Program, which today covers Opus and Sonnet-class models and says Mythos-class access is coming "in the near future." Read that as not at launch.

On the biology side, the latest classifiers fire 85% less often on benign elementary biology and medical questions than the ones that shipped with Fable 5. Life-science research and development still goes to Opus. The full capability is reserved for Mythos 5.1 through the Life Sciences Verification Program, built with the US government, which has enrolled its first participants.

One more safeguard change lands on the API side and could break an integration. New API accounts created from September 1 onward can no longer edit Claude's prior context in a multi-turn conversation while keeping the transcript of Claude's prior thinking. Anthropic says this closes a documented distillation technique. Existing accounts are not affected today, but the change will apply to everyone with future model releases. If your code rewrites assistant turns that contain thinking blocks, read Anthropic's Help Center article before your next model upgrade.

Zero Data Retention Is Back, With a Catch

The mandatory 30-day retention on Mythos-class traffic was the reason regulated teams kept Fable 5 out of production. Fable 5.1 introduces Enterprise Frontier Safeguards, or EFS: data stays in cloud infrastructure the customer controls, not Anthropic's, and any human review is by default done by the customer. Anthropic describes the privacy as equivalent to a zero data retention agreement, and says it built the system with more than 100 customers and with AWS, Google Cloud, and Microsoft Azure.

The catch is timing. EFS rolls out in phases starting this fall on Claude Code, Claude Enterprise, the Claude Platform, Amazon Bedrock, Claude Platform on AWS, Google's Agent Platform, and Microsoft Foundry. The bridge is the useful part: customers eligible for EFS can use Fable 5.1 and Fable 5 with zero data retention right now, by request. If retention was your blocker, file the request today rather than waiting for the phased rollout to reach you.

What Else Is in the Release

The science demos are the tell for where this line is going. Mythos 5.1 designed protein binders that two outside labs confirmed, with a hit rate of nearly 50% across 12 targets against a typical 10 to 15%, and affinities on three targets 10 times higher than the best entries in Adaptyv Bio's design competitions. Fable 5.1 trained a neural network on 30-year-old NASA Magellan radar data and produced an elevation map for a third of Venus at two to three kilometer resolution instead of 10 to 20, released under a Creative Commons license. Mythos 5.1 also rewrote GPU kernels for seven open-source genomics and protein models for speedups of 1.4x to 2.5x on an H100 with identical outputs. All vendor-reported. The binder work had external validation; the map and the kernels are Anthropic's own account.

The alignment findings are honest about the gaps. Mythos 5.1 is better aligned than Mythos 5 on most metrics Anthropic tracks, including fewer attempts to reach outside its test environment and less motivated reasoning of the "this is probably a simulation" kind. It can still sometimes bypass approvals and auto-mode classifiers, and Anthropic says its audit has less visibility into very long-context work and multi-agent settings than it wants. If you run unattended multi-agent workloads, that is the sentence to design your verification around.

EU watermark. Models Anthropic released after August 2, 2026 carry an invisible text watermark under the EU AI Act's Code of Practice. Anthropic says it does not change output quality and encodes nothing about you or your conversation. A detection API is in private preview for regulators, media, researchers, and enterprises with their own compliance duties.

Ship It This Week

Switch claude-fable-5 to claude-fable-5-1 and measure at the default effort. Same input and output price, cheaper cache reads, better scores. There is no cost reason to stay on Fable 5.

Try Medium effort on your routine agentic work. Anthropic's own chart says Fable 5.1 at Low or Medium matches or beats Fable 5 at much lower cost per task. Claude Code defaults to High, so the saving is opt-in there.

If you kept Opus 5 for cost, run the Cognition test. Put a code-review or long-context workload through Fable 5.1 with caching on and compare cost per task, not per token.

If retention kept Fable out of production, request EFS eligibility. The interim zero-data-retention option is available now for eligible customers.

Check any integration that edits assistant turns. The anti-distillation change applies to new API accounts today and to everyone with future releases.

Quick Hits

• Claude Security now runs on Mythos 5.1. Anthropic's codebase scanner moved up from Mythos 5 with this release. Using it still does not grant direct Mythos access.
• The June 23 subscription change stands as documented. Our Fable 5 launch coverage now carries a dated update pointing at what 5.1 changed and what it did not.
• Plus accounts got their five-hour limit back on August 26. OpenAI's Tibo Sottiaux said the reset covers ChatGPT Work and Codex. No numeric allowance was published.
• ChatGPT Sites can have editors. A Site owner can give a workspace teammate Can edit access; editors save and publish to the same URL, while ownership, secrets, and domain controls stay with the owner.

---

That is it for this issue. One model, three complaints answered: the bill is lower where it was growing fastest, the retention rule has a way out, and the safeguards let you find bugs again. If a teammate runs a security workflow that Fable 5 kept redirecting, forward this. The exploit work still goes to Opus, but the finding no longer does.

Until next week, stay caught up.