Anthropic releases Claude Fable 5.1 and Mythos 5.1. Here is what changed

Anthropic has today announced Claude Fable 5.1 and Mythos 5.1. The company has called them “the world’s most advanced models for coding and knowledge work”. Prices remain unchanged at $10 per million input tokens and $50 per million output tokens. However, the company says the new models are 25% cheaper to run for most customers.

Claude is a family of models at different sizes and prices: Haiku is the small one, Sonnet the medium, Opus the large. Above Opus sits a tier Anthropic calls Mythos-class. We explained that ladder when Opus 5 launched in July at half the price of Fable 5.

Fable and Mythos are the same model with different safety filters:

1. Fable 5.1 declines or redirects requests that touch on cyberattacks or dangerous biology, and anyone with a paid Claude plan or an API account can use it.
2. Mythos 5.1 loosens those filters for vetted cybersecurity defenders and life scientists. Anthropic’s announcement says it is currently available only to a set of US organisations and gives no date for wider access.

Why two versions exist

A little history: On June 9th 2026, Fable 5 launched. A few days later on 12 June, the US Department of Commerce ordered Anthropic to block access for foreign nationals. This was after Amazon researchers found a way past Fable 5’s cybersecurity filter that got it to find software vulnerabilities and, in one case, write code to exploit one. Anthropic thereafter switched off both models worldwide for 19 days. They returned access on 30 June.

With 5.1, the companys says it has made the filters more precise. The cybersecurity filter now fires around 60% less often per session in Claude Code. Going forward Fable 5.1 may find software vulnerabilities, the defensive half of security work. However, it will not write exploits for them. Penetration testing and exploit generation still get redirected to Opus models. The biology filter now fires 85% less often on ordinary medical questions. However, life sciences research still goes to Opus 5.

Anthropic says two external firms and the testing company Gray Swan attacked the new filters and found no critical jailbreak.

Benchmarks

1. On Terminal-Bench-Science 0.1, a test of research tasks run from a command line, Fable 5.1 scored 52.6% against 24.7% for Fable 5.
2. On Terminal-Bench 4.0, a coding test, it scored 55.8% against 42.0%, with Mythos 5.1 at 60.9% and OpenAI’s GPT-5.6 Sol at 37.3%.
3. On AutomationBench, which measures business workflows, it scored 31.4% against 17.1%.

Its knowledge now runs to June 2026, five months later than Fable 5, according to Anthropic’s developer documentation. The 1 million token context window and 128,000 token output limit are unchanged.

Where the 25% saving comes from

When an AI model re-evaluates previously processed text, this can be an extensive codebase retained within a continuous conversation, it performs a “cache read.” This action is billed independently from processing new, fresh inputs. Anthropic has aggressively reduced this specific cost: while Fable 5 charged $1.00 per million cached tokens, Fable 5.1 reduces this to $0.25. Analysing four weeks of real-world usage in August, Anthropic says this translates to an overall cost reduction of roughly 25% for standard workloads, and up to 45% for extensive, automated tasks heavily reliant on cache reads.

To highlight the model’s capabilities, Anthropic provided testimonials from 22 early adopters. Notably, a senior portfolio manager at Millennium reported that Fable 5.1 successfully identified a rare, one-in-a-million crash caused by a vendor library bug—an issue that had evaded detection for up to five years. However, it is crucial to recognise that these success stories were curated by Anthropic and do not represent independent, third-party benchmarks.

Three Subtle Updates:

1. Anti-Distillation Restrictions: To prevent “distillation” (copying the AI’s capabilities to train rival models), new API accounts can no longer edit Claude’s past messages while retaining its underlying reasoning data.
2. Invisible Watermarking: Complying with the EU AI Act for models released post-2 August 2026, all outputs now feature a statistical watermark. It contains no user data and leaves the text unaltered. A detection tool is currently in private preview for regulators, fact-checkers, and researchers.
3. Enterprise Data Security: Later this year, Anthropic will launch “Enterprise Frontier Safeguards,” enabling companies to host their data on private clouds while Anthropic monitors for misuse. Until then, eligible enterprises can operate Fable 5.1 under a strict zero data retention policy.