Anthropic has introduced FlashProve, an open-source prototype designed to allow users to verify that an AI response was generated by the model its provider claims to have used.
The company published the prototype’s code, a technical report and formal proofs in a research.
Currently, users have no reliable way of knowing whether an AI routing service has secretly switched to a cheaper model or whether an attacker has tampered with its responses.
FlashProve seeks to address this problem through a system called “provable inference”.
Under the approach, AI providers publish a fingerprint of their model weights, while each response comes with a cryptographic proof confirming that it was generated using those specific weights.
Anthropic said the verification process does not expose the model weights themselves, an important consideration for AI companies such as Anthropic and OpenAI, which keep their model weights private.
“This research post is meant to share intermediate progress, and not a completed research agenda.” Anthropic stated.
FlashProve relies on a type of zero-knowledge proof known as a zkSNARK. Anthropic tested the technology using Meta’s open-source Llama 4 Maverick, a 400-billion-parameter mixture-of-experts model.
The test showed that a single Nvidia H100 GPU could generate proofs at approximately 110 tokens per second, with each 1,024-token request producing a proof of about 19 megabytes.
However, generating the proofs requires 70 to 95 times more GPU computing time than running the model normally, Anthropic said.
Verifying a proof takes about 17 seconds on a standard CPU, without requiring a GPU or access to the model weights. Anthropic said the only previous proof of the same model took 14.3 hours to generate for 1,000 tokens.
