Claude Fable 5.1 is generally available. Mythos 5.1 is the same weights behind trusted access.

Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 on September 1, 2026. Fable 5.1 is generally available. Mythos 5.1 uses the same weights with more permissive safeguards, and Anthropic limits direct access to trusted programs for cybersecurity and life sciences.

The version number is the easy part. The buying decision is the stack around it: which safeguard profile you actually get, whether your data stays on Anthropic’s side or in your cloud, and whether a cheaper cache-read rate changes the bill for Claude Code and other agentic loops. Anthropic also announced Enterprise Frontier Safeguards the same day. EFS is not generally available.

Same weights, two safeguard stacks, two access paths

Anthropic’s system card states the two names are configurations of one model. Fable 5.1 is the general-access stack. It blocks a set of high-risk, dual-use tasks in biology and cybersecurity. Mythos 5.1 relaxes some of those domain-specific blocks for vetted users.

Treat them as two products with one capability floor, not as two independent model families.

• Claude Fable 5.1: generally available today on Anthropic’s API as `claude-fable-5-1`, and Anthropic lists Amazon Web Services, Google Cloud, and Microsoft Azure among the platforms that carry it.
• Claude Mythos 5.1: same weights, more permissive safeguards, trusted access only. Anthropic currently limits that access to a set of U.S. organizations and says it is coordinating with the U.S. government on a wider set of domestic and international partners.
• Cyber Verification Program: Anthropic says the CVP already offers certain Opus- and Sonnet-class models with reduced cyber safeguards for defensive work, and that Mythos-class access is planned for the near future, not as a same-day swap.
• Life Sciences Verification Program: Anthropic says it designed LSVP so life-science professionals can use Mythos 5.1 with research-oriented biology safeguards, and that it enrolled the first participants with the U.S. government.

Fable 5.1 defaults to High effort in Claude Code and to Medium effort in Claude Cowork and on Claude.ai, per Anthropic’s launch page. If you compare runs across surfaces, start by matching effort. Then compare the Claude profile with ChatGPT and Gemini on the same jobs, not on vendor charts alone.

Company-reported scores, with Anthropic’s own caveats

Anthropic published a comparison table on the launch page and labeled the numbers as its own setup. Fable 5.1 ran with production safeguards on. Anthropic says that when those safeguards intervened, Fable 5.1 and Fable 5 scored zero on OSWorld 2.0, Fable 5 scored zero on AutomationBench, and other intervened cybersecurity and biology tasks went to Claude Opus 4.8 and Claude Opus 5.

Use the table as a vendor scorecard, then run the jobs you actually pay for. Our 2026 agent evaluation framework is the better place to write those jobs down before you change a production string.

• Terminal-Bench-Science 0.1: Anthropic reports Fable 5.1 at 52.6% and Fable 5 at 24.7% in its setup. It also reports a standard error of ±3.5 to 4.5 points per model, and it notes the public leaderboard (three trials per task, Claude Code harness) showed different Fable 5 and Opus 5 scores that still sat within noise of its reproduction.
• Terminal-Bench 4.0: Anthropic reports Fable 5.1 at 55.8% and Mythos 5.1 at 60.9%, against 42.0% for Fable 5 in the same table.
• CursorBench 3.2.0: Anthropic reports 73.4% for Fable 5.1 against 70.5% for Fable 5.
• Cyber: Anthropic and the system card call this the strongest cyber capability Anthropic has released. The system card still places Mythos 5.1 in the lower Frontier Compliance Framework cyber-offense category (Tier 1, not the autonomous Tier 2). The launch page says Anthropic found no critical-severity jailbreak for Fable’s cyber safeguards.
• Biology: Anthropic says Mythos 5.1’s chemical and biological capability is higher than Mythos 5 and still below the next Responsible Scaling Policy tier. The system card assigns a CB-1 judgment and says the model does not cross CB-2.

Anthropic’s launch page also reprints partner comments. On that vendor page, Jane Street’s Craig Falls says Fable 5.1 “solves more of our coding problems than Fable 5 or Opus 5” in Jane Street’s internal benchmarks and “remains readable over long, multi-step tasks.” The same page says Millennium’s tests found the cause of a rare crash that its engineers and other models had not explained after several years. Those are partner statements on a vendor page, not independent audits.

The price cut is on cache reads, not on the headline input rate

Anthropic kept Fable 5.1’s usage-based input and output rates at Fable 5’s $10 per million input tokens and $50 per million output tokens. It cut cache-read pricing by 75%, to $0.25 per million tokens, wherever usage is billed by token. Cache reads are tokens the model rereads after it has already processed and stored them. That is most of the bill on long, tool-heavy loops.

Anthropic estimates typical token workloads cost about 25% less than Fable 5, and it estimates highly agentic, context-heavy work can fall by about 45%. It measured those indexed costs at default effort over four weeks of August 2026 usage across Claude Enterprise, Claude Code, and the API. Highly agentic work is the slice where cache reads dominate. Treat both figures as company estimates, then re-price your own traces.

Enterprise Frontier Safeguards are not generally available

EFS is Anthropic’s answer to the retention tradeoff that started with Fable 5. Anthropic says it began 30-day retention so it could correlate misuse across sessions and accounts, and that it does not train on enterprise data without explicit permission. Many regulated buyers still could not accept provider-held logs. EFS keeps automated misuse detection and stores the activity data in infrastructure the customer controls.

• Customer-owned storage: activity data used for monitoring can live in the customer’s cloud account, such as Amazon S3, Azure Blob Storage, or Google Cloud Storage, under the customer’s keys, access policies, and audit logs.
• ZDR-equivalent privacy, with misuse detection: Anthropic describes EFS as the privacy of a zero-data-retention policy plus automated review across a rolling window of traffic. Flags go to the customer. Anthropic says customer human review is the default, and that Anthropic employees do not have to read the traffic.
• Not generally available: Anthropic is rolling EFS out in phases, starting later this fall. It is taking access requests by form, not shipping a self-serve toggle.
• Interim ZDR: until EFS is ready, Anthropic says eligible customers receive zero data retention on Fable 5 and Fable 5.1.
• Surfaces Anthropic lists for EFS: Claude Code, Claude Enterprise, the Claude Platform, Amazon Bedrock, Claude Platform on AWS, Google’s Agent Platform, and Microsoft Foundry.
• Commercial terms Anthropic states: it does not charge for EFS. Customer-owned storage, customer-managed keys, and fully automated review are each opt-in. None of those controls, Anthropic says, change model behavior, API pricing, or rate limits. The customer’s cloud provider bills storage, reads, writes, and egress.

Write EFS as a roadmap item with an interim ZDR clause, not as a control you already have. If a vendor questionnaire still asks who holds the logs, the accurate answer for a non-eligible account is still Anthropic’s retention policy, not customer-owned storage. Pair that question with the data-handling items in our OWASP agentic vendor rubric.

Fable 5.1 may find source vulnerabilities. It still must not write exploits.

Anthropic changed the general-access cyber policy in one narrow way. Fable 5.1 may identify software vulnerabilities in source code, including at general availability. Anthropic still blocks exploit development. The launch page says Fable’s updated cyber safeguards produce about 60% fewer false positives than the prior Fable 5 stack, and that Claude Code users should see about 60% fewer cyber-safeguard interventions per session versus the previous Fable 5 safeguards.

• Allowed on Fable 5.1, per Anthropic: source-code vulnerability discovery and other defensive work that improves software security.
• Still redirected to Opus-class models on Fable 5.1: penetration testing, exploit generation, and binary-based vulnerability scanning.
• Still blocked on Fable 5.1, per the system card: vulnerability discovery against compiled binaries, which Anthropic treats as closer to offensive technique.
• Still blocked on Fable 5.1: professional biology and drug-development queries. Anthropic says those stay on Opus-class models for general users.

This is a policy change, not a playbook. Do not use Fable, Mythos, or Claude Security to develop exploits. If a safeguard redirects a request, treat that as a hard stop. If your security team needs reduced cyber safeguards, Anthropic’s stated path is the Cyber Verification Program, then Mythos when that program includes it. Ask the vendor how it logs a redirect, who can override it, and what the fallback model is allowed to do.

Life-science access stays gated. Claude Security now runs on Mythos 5.1.

Anthropic’s scientists program note is explicit: Fable models continue to block professional biology and drug-development queries because of dual-use risk. Biology and chemistry researchers on the free and discounted scientist seats still get Opus-class models for that work. Anthropic says it is building a life-sciences access program for Mythos-class models with the U.S. government, has enrolled the first participants, and expects to widen enrollment.

Separately, Anthropic says Claude Security, the Claude Enterprise product that scans codebases for vulnerabilities and suggests patches for human review, now runs on Mythos 5.1. The system card says that product is available to all Claude Enterprise customers even when those customers cannot call Mythos 5.1 directly. Human review of suggested patches stays in the product description. Treat the scan as an input to your own triage, not as a ship decision.

What to do this week if Claude is already on the shortlist

A model-string swap is cheap. A safeguard, retention, and price-mix change is not. Do the inventory before you point production agents at `claude-fable-5-1`.

1. List every workload that would move to Fable 5.1, and mark which ones are token-billed, which ones run in Claude Code at High effort, and which ones still need Opus because they are professional biology, pentest, or exploit-adjacent.
2. Ask Anthropic, in writing, whether your account is eligible for interim zero data retention on Fable 5 and Fable 5.1, and when EFS would reach your region and cloud. Do not write EFS into a control matrix as if it is live.
3. Re-price a real August-style trace with the new cache-read rate. If cache reads are a small share of tokens, the company-estimated 25% cut will not show up.
4. Re-run your own coding and science evals at a fixed effort level. Do not accept Terminal-Bench or CursorBench as a substitute for the jobs in your agent evaluation framework.
5. If you still hold OpenAI model strings that sunset this fall and winter, fold Fable 5.1 into the same inventory you use for the 2026 OpenAI deprecation wave. Changing one vendor’s default does not retire the other vendor’s shutdown dates.
6. Keep ChatGPT and Gemini on the same shortlist as Claude. Anthropic’s table includes GPT-5.6 Sol as a comparison column. It does not replace a bake-off on your data, your tools, and your retention rules.

Once you finish the inventory, open the Claude and Claude Code profiles, add ChatGPT and Gemini to the same list, and save it. The next useful click is a side-by-side on price, retention, and who is allowed to call which safeguard stack, not another vendor chart.