Free Claude credits for US science need an exit plan

Anthropic is committing $150 million over three years to the Genesis Mission, the US federal initiative to speed up research with AI. Most of that money arrives as Anthropic's own product. According to the announcement, it covers Claude, Claude Code, and API credits for several hundred research projects across more than 15 agencies, including NASA, NIH, and NSF. It also pays for training, onboarding, and help getting agencies new to the program running their first projects.

Scientists at national labs getting serious tools is good news, and the onboarding may matter more than the credits. Research groups tend to stall on setup, data access, and finding time to learn a new workflow, long before model quality becomes the limit. Still, credits are a discount on a proprietary platform, and the training teaches that platform. Three years from now, several hundred projects will have pipelines, prompts, and agent tooling built around one company's models and its Claude Science workbench. Whatever the intent, that leaves a slice of public research with a switching cost attached. Once the credits run out, Anthropic sets the price, not the agencies.

For anyone running a research or engineering budget, taking the offer is the easy call. The harder one is making sure funded projects keep the model swappable: data in open formats, evaluation sets the agency owns, and orchestration that does not assume a single API. The Model Hardware Standard that Anthropic mentions, a shared specification for agents operating lab instruments, could help with that if it ends up open to other model providers. Whether it does is the detail I would watch.