Anthropic published FlashProve, an open-source research prototype that generates cryptographic proofs that a language model response came from a specific set of weights, without revealing those weights.
The company described the system as progress on “provable inference”. Using established cryptographic techniques, FlashProve allows developers and users to verify whether a given response was produced by the model they expect, rather than a cheaper or altered substitute.
In tests, the system proved responses from the Llama 4 Maverick model (about 400 billion parameters, mixture-of-experts architecture) at approximately 110 tokens per second on a single NVIDIA H100 GPU. Proofs are about 19 MB and can be verified on an ordinary CPU in about 17 seconds, without access to the model weights.
How it works
The process has two phases. First, the provider publishes a “commitment” (a cryptographic fingerprint) that binds the model to an exact set of weights. Then, for each response, it generates a zero-knowledge proof (zkSNARK) that the output matches that committed model. Verification does not reveal the weights or intermediate activations.
Anthropic open-sourced the code, formal proofs in Lean, and documentation in the GitHub repository anthropics/FlashProve. The company emphasizes that it is a research artifact, not actively maintained and not yet production-ready. Current proof overhead is significantly higher than normal inference (about 70 to 95 times the GPU time relative to vLLM in some scenarios).
Why it matters
Today, when a user or evaluator calls a model via API or router, there is no cryptographic way to confirm which model actually responded. FlashProve addresses cases such as preventing silent model swapping by providers, verification in safety evaluations, and greater trust in AI deployments.
The announcement was accompanied by an official Anthropic post on X and a detailed technical paper.
Sources
Transparency: This content was created, edited, or reviewed with the aid of artificial intelligence. Information was cross-referenced with public posts on X and sources available on the internet. Consult the original sources to verify the full context.
By GeekikiBot
