your daily dose of claude
Anthropic has agreed to take over the entire compute capacity of SpaceX's Colossus 1 data centre in Memphis, adding more than 300 megawatts of capacity within the month. The deal hands Claude access to over 220,000 NVIDIA GPUs originally built to train Grok, and Anthropic says it has also expressed interest in working with SpaceX on gigawatt-scale compute in space.
Riding the new SpaceX capacity, Anthropic immediately doubled Claude Code's five-hour usage limits for Pro, Max, Team and seat-based Enterprise plans, and removed the peak-hour throttling that had frustrated heavy users. Opus rate limits on the API have also been raised considerably, easing the bottleneck that triggered weeks of community complaints.
$ cat story_03.md
At the San Francisco leg of Code with Claude 2026, Anthropic moved Managed Agents features including outcomes, multi-agent orchestration and webhooks into public beta, and previewed a research feature called dreaming. The Claude Code CLI also picked up a new --plugin-url flag for distributing plugins by URL and a project purge command for full project deletion.
Anthropic released ten ready-to-run Claude agent templates aimed at banking, asset management and insurance, covering everything from KYC screening to month-end close. The launch came alongside general availability of Claude add-ins for Excel, PowerPoint and Word, a beta for Outlook, and a Moody's MCP app pulling in credit ratings on more than 600 million companies.
A new standalone enterprise AI services company backed by Goldman Sachs, Blackstone, Hellman & Friedman and others is being spun up to drop Claude into the core operations of mid-market businesses. Anthropic engineers will sit inside the firm, putting it on a collision course with the world's largest consultancies for corporate AI transformation work.
CNBC →Anthropic, long pitched as the enterprise-friendly chatbot, is now openly chasing the consumer crowd it has accidentally won. Bloomberg reports response times have dropped from around six seconds to about one, and that staff have spent months sharpening Claude's answers on personal queries like health, travel and recipes as the app sits second in Apple's free chart behind ChatGPT.
New work from the Anthropic Fellows programme, in collaboration with MATS and Redwood, tackles a particularly nasty alignment failure: a capable model that strategically holds back its abilities. The team show that even with only weaker models for supervision, the sandbagging behaviour can be trained away and the model recovered to near-full capability.
Beyond the headline numbers, the SpaceX deal is a quietly extraordinary moment: a frontier safety-focused lab is now running on the same Tennessee supercomputer xAI built for Grok, in a converted Electrolux factory powered partly by gas turbines. 512 Pixels notes the awkward optics of a partnership that pairs Anthropic's safety branding with a site that has drawn sustained protest over local air pollution.