Anthropic’s AI proof tool needs up to 95 times more GPU time

Anthropic’s FlashProve can verify that a private AI model produced an output, but generating a proof used 70 to 95 times the GPU time of ordinary inference in a benchmark.

Anthropic released FlashProve on Nov. 2 as a research prototype for checking whether a hidden AI model produced a specific output. The system checks the computation without exposing the model’s private weights. It is not a verification feature available to Claude customers. Anthropic tested it with Meta’s Llama 4 Maverick.

The process starts with a provider publishing a cryptographic commitment to the model’s weights. The commitment acts as a fixed identifier while keeping the weights private. The provider then creates a proof that the committed model produced a stated output from a given input.

A verifier can check the proof without downloading the model or running it again. The method uses zero-knowledge techniques, which can confirm that a computation occurred without revealing the private information used to perform it.

The model’s identity depends on the commitment being trustworthy. FlashProve’s repository says the commitment ID should come from a trusted party, rather than the server being checked or files supplied with the proof. A proof tied to an untrusted commitment would not establish that the advertised model produced the output.

Anthropic’s test used a public, four-bit version of the 400-billion-parameter Llama 4 Maverick. Proof generation ran on one H100 GPU. Verification ran on a CPU and took 17 seconds with 24 CPU threads, or 21 seconds with one thread. Each proof was about 19 megabytes.

The reported cost compares FlashProve with vLLM, an inference engine running an FP8 version of the same model on eight H100 GPUs. The 70-to-95-times figure compares those specific implementations.

The report cites throughput of 110 tokens per second, but that number counts all tokens processed in each 1,024-token request. In the main benchmark, each proof covered one following output token. The throughput figure therefore does not represent 110 separately authenticated new tokens per second.

FlashProve can generate proofs after a response has been delivered by recomputing the transcript. This can delay detection if an output was changed before proof generation. The system currently supports contexts of up to 1,024 tokens and proves the exact greedy choices of an integer model. Floating-point calculations and sampled responses can produce different results.

A successful proof does not show that an answer is factually correct or safe. It only checks that the model performed the specified computation.

Anthropic’s formal security claims were checked in Lean, a system for verifying mathematical proofs. The faster executable code has not been formally shown to match that protocol. The report also says zero-knowledge properties for requests processed together in a batch have not been formally proved. The repository describes FlashProve as an unmaintained research artifact.

Articles by this author