Anthropic has released FlashProve, a prototype intended to let a user verify that a response came from a committed model without learning its weights. The research announcement frames this as a check on the computation behind an answer, rather than a test of whether the answer itself is true.
The model identity problem
A customer of a routing service usually cannot establish model identity from the response alone. FlashProve uses a zero-knowledge succinct argument to provide a certificate tied to the model's committed weights.
The goal is to make verification cheaper than rerunning the model and to avoid disclosing the weights. Hiding the model architecture is not one of this system's stated goals.
Performance comes with overhead
For a four-bit-quantised Llama 4 Maverick mixture-of-experts model, the announcement reports around 110 tokens per second of proof generation on one H100. A proof is approximately 19 MB per 1,024-token request.
Those numbers should be read alongside the cost comparison: proving consumes roughly 70–95 times the GPU time of the described vLLM inference baseline. A fast-looking token rate does not mean proving is as inexpensive as ordinary serving.
What is being released
The release includes executable code, a technical report, a protocol specification and Lean proofs for top-level mathematical statements. The code includes a prover, a C++ verifier and a Lean verifier.
The authors distinguish formal statements from the work of auditing whether those statements capture the intended guarantees and whether executable implementations match them. The everyday C++ verifier does not carry the same formal guarantee as the Lean verifier.
An intermediate result
FlashProve uses post-hoc proving rather than producing every guarantee inline with the original response. That introduces delay and recomputation considerations. The contribution is a research prototype and released auditing material, not a claim that ordinary production APIs now offer universal, low-cost proofs for every answer.