Here is today's AI news. Story 1. Anthropic. Anthropic released Claude Fable 5.1 and the restricted-access Claude Mythos 5.1, reporting major gains in agentic scientific-research performance and separating access by safety safeguards.. Anthropic released Claude Fable 5.1 for general use alongside restricted-access Claude Mythos 5.1, which share the same model weights but use different safeguards. Anthropic reported Fable 5.1 scoring 52.6% on Terminal-Bench-Science 0.1, versus 24.7% for Fable 5. Mythos is limited to vetted U.S. cybersecurity and life-sciences organizations, signaling a new model for deploying frontier AI with risk-based access. Story 2. OpenAI. OpenAI said its upcoming Astra model crossed its highest cybersecurity-capability threshold, requiring stronger safeguards and an initially limited rollout.. OpenAI said its upcoming Astra model is the first to reach its “Critical” cybersecurity threshold. With the right tools, it can find unknown flaws and develop zero-day exploits in hardened systems without human guidance. The capability could strengthen defensive testing but also enable serious attacks, prompting stricter isolation, encrypted model checkpoints and continuous monitoring. Advanced cyber access will initially be limited to vetted testers through Daybreak. Story 3. Google. Google released Gemini 3.8 Flash and a restricted Gemini 3.8 Flash Cyber variant, upgrading reasoning and coding performance while reserving cyber capabilities for trusted defenders.. Google announced Gemini 3.8 Flash on Sept. 2, 2026, calling it its most capable Flash model for software engineering, agentic workflows and complex reasoning. It supports a 1-million-token context, 64K-token outputs and adjustable thinking levels, priced initially at $0.75 per million input tokens and $3.75 per million output tokens. A cybersecurity variant, Gemini 3.8 Flash Cyber, detects vulnerabilities and develops patches but is limited to vetted governments, critical-infrastructure operators and maintainers through Google’s Fairwind Program. Story 4. Meta. Meta launched Muse Spark 1.3, a new model focused on coding, long-horizon agents, and real-world computer-use tasks.. Meta unveiled Muse Spark 1.3, its latest AI model, aimed at coding, long-horizon agent workflows, and computer-use tasks. Meta says it improves long-context handling and follows complex instructions more reliably. The update matters because it strengthens Meta’s push into developer tools and autonomous agents, competing with OpenAI and Anthropic in real-world software work. Story 5. European Union AI Office. The EU AI Office sent its first formal general-purpose-AI enforcement information requests to OpenAI, Anthropic, and Google under the EU AI Act, covering security practices, independent evaluations, and post-deployment monitoring.. **EU targets leading AI labs in first enforcement step** The EU AI Office has sent its first formal information requests under the AI Act to more than 30 general-purpose AI providers, reportedly including OpenAI, Anthropic and Google. The requests seek details on model security, independent evaluations, post-deployment monitoring and training-data summaries. The move turns voluntary compliance discussions into legally enforceable oversight. Firms must respond accurately: incomplete or misleading answers could bring fines of up to €15 million or 3% of global annual turnover. That's all for today's AI news briefing.