Anthropic Launches Claude Fable 5.1 With Major Cost Cuts as OpenAI Readies Powerful Astra Model With Critical Cyber Rating

Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 on September 1, calling them the world's most advanced models for coding and knowledge work. Fable 5.1 is built for complex, long-running agentic tasks and is available immediately on AWS, Google Cloud and Microsoft Azure, while Mythos 5.1 targets cyberdefenders and life scientists and is available only through trusted access programs.

The performance gains are substantial. On Terminal-Bench-Science 0.1, Fable 5.1 scored 52.6%, more than double the 26-plus percent posted by its predecessor Fable 5, and on Terminal-Bench 4.0 it scored 55.8% versus 42.0% for Fable 5. Cache read costs are 75% lower than Fable 5, which Anthropic says cuts practical costs by roughly 25% for typical workloads and up to 45% for highly agentic ones. The company cited customer feedback as the driving force behind both the performance improvements and the pricing changes.

On safety, Anthropic said its cybersecurity safeguards now flag benign requests about 60% less often and the fallback rate on basic biology and medical questions is down around 85%. The company said it found no evidence of a critical-severity jailbreak for either model's safeguards. Anthropic is also introducing Enterprise Frontier Safeguards that give enterprise customers privacy equivalent to zero data retention, rolling out in phases starting this fall. Fable 5.1 can now be used to discover software vulnerabilities, and Anthropic said it expects to open enrollment for scientists soon. The new models also include a block on AI distillation, aimed at preventing competitors from copying the models' capabilities.

Separately, OpenAI announced that its forthcoming Astra model is the first it has rated as reaching the "Critical" cybersecurity threshold under its preparedness framework. OpenAI said Astra can discover previously unknown vulnerabilities and develop exploits across well-protected systems autonomously, and during testing it found and chained together two zero-day vulnerabilities, which OpenAI is disclosing to the relevant maintainers. The most advanced cyber capabilities will initially be restricted to a small group of testers. Following the Hugging Face incident, OpenAI paused some frontier training and strengthened Astra's refusal behavior, misuse protections and monitoring, restarting one large frontier reinforcement-learning run on August 28 while some smaller experimental runs remain on hold. OpenAI described Astra as representing a significant increase in cybersecurity capabilities compared to GPT-5.6 Sol.