2 September 2026
Anthropic releases cheaper, less restrictive Claude Fable 5.1
First reported
AI Breakfast, TechCrunch and 3 others ran this on , a day before the other 4 sources picked it up.
- Claude Fable 5.1 costs about 25 percent less for typical work and up to 45 percent less for complex autonomous tasks, primarily through reduced pricing on cached data.
- Safety filters are less aggressive: cybersecurity false positives dropped 60 percent and biology-related false positives dropped 85 percent compared to prior versions.
- Fable 5.1 can now identify software vulnerabilities for the first time, though penetration testing and exploit generation still require the more restricted Mythos 5.1 model.
- Mythos 5.1, the restricted variant for cybersecurity and life sciences work, is only available to US organizations through special access programs.
Where they differ
The Rundown AIemphasized cost savings and reduced safety friction, while Artificial Analysis disputed the overall savings claim, noting that Fable 5.1 actually costs more per task than prior models when accounting for higher output token usage.
Simon Willisoncovered unrelated security and documentation changes. All newsletters that discussed the model release itself agreed on the core improvements.
What each one reported
Anthropic introduced Claude Fable 5.1 and Mythos 5.1 with stronger coding and research capabilities, lower effective pricing, and updated safeguards. Mythos uses the same underlying model with specialized access controls for advanced cybersecurity and life sciences work.
Anthropic released Fable 5.1 with cache reads costing 75% less at $0.25 per million tokens, reducing typical workload costs by 25% and agentic tasks by up to 45%. The model improved on agentic science tests from 24.7% to 52.6% and loosened safeguards that interrupted legitimate work, though some safety friction remains a problem.
Anthropic launched Claude Fable 5.1, a successor to Fable 5 that tops AI benchmarks with a score of 66 and shows major improvements on coding, research, and knowledge work tasks while reducing safety rejections by 60-85% on certain domains. The newsletter emphasises this marks the end of a quiet period from AI leaders and signals an acceleration in model releases, with OpenAI's Astra expected to follow soon.
Attackers used infostealer malware to steal Claude login sessions and drain user account usage. Anthropic responded by forcing sign-outs, stripping saved payment methods, and refunding charges.
Anthropic has reorganised their published Claude system prompts into an index with per-model pages showing both current and historical versions, with each page available in Markdown format for easy diffing by both humans and LLMs.
Reported by The Verge, VentureBeat, The Decoder, TechCrunch