Anthropic studies provable inference

Anthropic’s research looks at a basic trust problem: when you send a request to an AI model, you can’t independently check that the response came from the model the provider says it did. A routing service could substitute another model, or a token stream could be tampered with, and the output alone may not reveal it.

The proposed answer is provable inference. The provider would send a cryptographic proof alongside a response, letting a user check that it came from a specified model without seeing the model’s weights. Anthropic describes FlashProve, an open-source prototype, and reports that it can generate proofs for a large model. But the reported proof generation takes about 70 to 95 times the GPU time of running the model itself with vLLM. Anthropic describes this work as intermediate progress, not a completed research agenda.

For an operator, the idea matters most where you need evidence about which model handled a request, such as when using a third-party routing service or checking an AI provider’s commitments. The research drew 95 points and 34 comments on Hacker News, but that’s a signal of interest, not proof of business adoption. Teams using AI can consider what evidence they need about which model handled a request. Read Anthropic’s research.