Meta ships its frontier model Muse Wonder, Torvalds codes with Claude
Meta's Muse Wonder, Linus Torvalds writing kernel code with Claude Code, Anthropic's warning about small models copying themselves, and its crackdown on rivals copying Claude.
Meta’s frontier model, Muse Wonder, is out at Claude Opus prices It went out on Monday through Meta’s API at $4 per million tokens in and $20 out, with a bigger Full version kept for subscribers to its coding app. Artificial Analysis, which tests models independently, puts the Full version level with Claude Sonnet 5.5, with only Claude Opus 5.5 ahead. The big pitch is memory: it keeps track of what matters to you across conversations. Source
Linus Torvalds used Claude Code to triple Linux’s small-read speed The branch turned up on Sunday and credits Claude Code with the code, the test programs and the first draft of his write-up, all done under his direction. On his laptop it took one kind of file read from about 10 million a second to over 30 million. It isn’t merged yet. Source
Anthropic says small AI models may be months from self-replicating That means copying themselves onto cheap internet-connected gadgets and spreading on their own. Anthropic’s red team got Google’s Gemma running across 15 Android TV boxes that cost $55 each, and in a simplified test where breaking in was made easy, Zhipu’s GLM-5.2 spread to all eight machines in four of five tries. It published the research on Monday. Source
Anthropic says it’s cut rivals copying Claude by up to 99% Labs copy a model by training their own on its answers, and Anthropic says requests sent to Moonshot’s Kimi were secretly forwarded to Claude to do exactly that, one from someone it believes was linked to China’s military. It has banned tens of thousands of accounts, and since 24 September it charges for requests its filters block before Claude answers. It put the details out on Monday. Source