Anthropic releases FlashProve model-verification research

Anthropic has released FlashProve, an open-source research prototype designed to let users verify that an AI response matches a specified model’s output for a given prompt. The system uses cryptographic proofs tied to a fingerprint of the model’s weights, which remain hidden from the verifier. ^[1]

In Anthropic’s test, a single H100 GPU generated proofs for the 400-billion-parameter Llama 4 Maverick model at about 110 tokens per second; each proof was roughly 19 MB. Anthropic says verification took about 17 seconds on a 24-core CPU, while generating proofs took substantially more GPU time than running ordinary inference. ^[1]

The company describes the release as intermediate research, not a finished system. Its post-hoc proof process adds computation and delays verification, and formal verification does not remove the need to audit the statements and executable code. TechCrunch reported there are no plans to deploy FlashProve at Anthropic or another lab. ^[1][2]