Did you get the model you paid for? Anthropic opens FlashProve
The research prototype binds responses to committed model weights, but its proof generation costs roughly 70–95 times the GPU time of plain inference in Anthropic's tests.
Paying for a particular AI model does not give you a way to check its identity from the answer alone. Anthropic’s FlashProve tackles that gap with cryptographic proofs. The company published the research on November 2, with code, a protocol specification and formal proofs under Apache 2.0. November 3 reporting has put the proposal in front of a wider audience.
The workflow starts with a provider publishing a commitment that pins down one set of model weights. After inference, the provider produces a proof tied to that commitment. A client can check the proof using the prompt, output and commitment, without receiving the weights or rerunning the large model. The architecture itself is not hidden.
A receipt for model identity comes at a cost
Anthropic reports testing a four-bit version of Llama 4 Maverick on a single H100, with 24 requests in a batch and 1,024-token windows. Its reported proving throughput is about 110 tokens per second in prefill-style processing. This is not ordinary chat generation speed: the company puts the GPU-time overhead at roughly 70–95 times the plain-inference comparison using vLLM.
Each proof checks one token in a request window, and the reported proof size is about 19 MB. Anthropic says CPU verification took 17 seconds on 24 threads in its evaluation. AITrending has read the released evidence but has not reproduced those measurements.
The practical attraction is an auditable answer to model substitution: a routing service or provider could not simply assert that a cheaper model’s output came from the committed weights. That does not establish that an answer is correct, useful or based on accurate external data.
Research code, with specific boundaries
The prototype proves an integer model’s greedy choice, not arbitrary floating-point or sampled decoding. It currently handles contexts only up to 1,024 tokens. Anthropic also distinguishes its Lean-verified protocol from the fast executable code, whose equivalence is not formally guaranteed.
The repository supplies a CPU verifier and example proofs. Running the prover requires its supported x86-64 Linux, CUDA and H100 stack; the full real-weight example also needs substantial host memory, disk space and access to the model’s tokenizer. The repository calls itself an unmaintained research artifact. This is an invitation to inspect and extend model-identity verification, rather than evidence that production Claude responses now arrive with such proofs.
