In today’s issue:
Anthropic ships Claude Fable 5.2
Google launches Gemini 4 Hydrogen
CMU trains agents to write memories
AlphaProof Code verifies Signal’s code
Singularity-Bench tests agents building AI
OpenAI tracks CoT monitorability decline
Meta releases Muse Ember
Gemini Decisions API launches
Amazon swaps Sonnet for Strands Decider
Continual-ARC tests agent learning
And all the top AI dev news, papers, and tools.
Top Stories
Claude Fable 5.2
Anthropic released Claude Fable 5.2 and Claude Mythos 5.2, which share the same underlying model. Fable carries stronger cyber and biology safeguards and is available to everyone, while Mythos is limited to vetted organizations in Anthropic’s Cyber and Life Sciences Verification Programs.
Agentic coding: Fable 5.2 scores 49.1% on FrontierCode 2.0 at Xhigh effort, the highest of any released model Anthropic tested, against 45.5% for Opus 5.5 and 42.8% for GPT-6 Astra. At High effort, it roughly matches the best Opus 5.5 score for about half the cost per task.
Computer use and robotics: It reaches 89.3% on OSWorld 2.1 (offline set, partial credit) and 64.9% on RoboDojo Sim, where Fable 5.1 scored 27.6%. On the Remote Labor Index, it scores 41.3% against 21.3% for Opus 5.5.
Agent teams: A team of 1,000 Fable 5.2 agents formalized Fermat’s Last Theorem in 30 hours, against 11 days for Anthropic’s first formalization. Opus 5.5 with the same setup completed 26% in 18 hours.
Security: Mythos 5.2 scores 89.6% on CVE Rediscovery and 76.3% on CyScenarioBench. In adversarial sandbox-escape tests, it used planted credentials or files in 3.1% of scenarios, down from 21% for Mythos 5.1.
Pricing and speed: Fable 5.2 costs $8 per million input tokens and $40 per million output tokens, down from $10 and $50, and generates output up to 30% faster than Fable 5.1. Medium is the default effort in Claude apps and Claude Code.