AI developer Anthropic introduced FlashProve, a cryptographic system for verifying that AI responses come from the claimed model. The prototype generated proofs at about 110 tokens per second for Meta's Llama 4 Maverick, a model with 400 billion parameters. Anthropic published interim results to invite outside audit and improvement.
How FlashProve works
The system publishes a digital fingerprint of a model's weights. It then creates a mathematical proof for each response. Users can check that proof without seeing the model's internal parameters.
Test performance
Anthropic tested the system on one NVIDIA H100 GPU. Each proof was about 19 megabytes per 1,024-token request. Checking a proof took between 17 and 21 seconds on a CPU.
Potential uses
The tool could help customers confirm they receive the model they paid for. It could also detect model substitution and help verify work attributed to an AI. Regulators and independent researchers may use it to check which models companies run.
FlashProve is not ready for wide deployment. Proving currently uses about 70 to 95 times more GPU time than ordinary processing. Anthropic released code, documentation, and Lean proofs under the Apache 2.0 license.