Meta ships its first frontier model, Muse Wonder
Meta's Superintelligence Lab released Muse Wonder and Muse Wonder Full on Monday, its first real attempt at a frontier-class model after the lightweight Muse Spark. Trained at Meta's gigawatt-scale Prometheus data centre, the pair is pitched at personal and professional agents, built around "Active Memory" and "Active Focus" for long, multi-step work. Meta's own model card puts them level with the current top models on coding and ahead on its MemoryBench recall test. Wonder is live now on the Meta Model API at $4 per million input tokens and $20 per million output, while Wonder Full is kept for Muse Code subscribers only. The practical upshot: there's now a fourth serious lab selling frontier agent models, and it's pricing at the top of the market. Meta's announcement
What else matters
Anthropic details its agents going off-script
Anthropic published a report on unintended actions by Claude during testing and internal use. One run of Claude Haiku 4.5 filed an invented tip on a Philadelphia police homicide form, and a research model submitted 20 incomplete US visa applications. Anthropic says it has cut live internet access for evaluations and tightened its web tools. Anthropic's report
Mistral Large 4 lands as a 1-trillion-parameter preview
Mistral opened a public API preview of Mistral Large 4, a 1.05-trillion-parameter multimodal model with a 1-million-token context window and native tool calling, through Mistral Studio and OpenRouter. Open weights are promised later this month, so the API phase is effectively a paid benchmarking window before the download arrives. Launch write-up
UK government announces an AI risk package
The UK is extending a frontier-AI cyber-defence programme it says has already found and fixed more than 1,000 vulnerabilities in government systems. It will also look at strengthening the Cyber Security and Resilience Bill against AI loss-of-control risk, how serious AI incidents should be reported, and backing for UK AI verification work. Unite.AI on the package
Broadcom and Intel say money, not chips, is the bottleneck
Speaking in Singapore on Monday, Broadcom's Hock Tan said access to capital, rather than shortages of chips, memory or power, may now be the biggest constraint on the AI build-out. Intel's Lip-Bu Tan pointed to high borrowing costs and doubts over whether AI spending is earning its keep. Financial News report
Cloudflare absorbs the Deno team
The whole Deno team is joining Cloudflare to merge its celld project into workerd, making Workers and Durable Objects easier to self-host on your own servers. The Deno runtime gets a final year of maintenance releases, Deno Deploy shuts down in six months, and the JSR registry carries on under Cloudflare. Cloudflare's post
