Anthropic releases a prototype for proving which model answered

Anthropic released FlashProve on November 2, saying it can verify output against committed model weights without revealing them. The company reports proof generation at 110 tokens per second for quantized Llama 4 Maverick, using roughly 70–95 times the GPU time of vLLM inference. Its research prototype covers integer computation with greedy decoding and contexts of up to 1,024 tokens; formal correctness guarantees apply to its slower Lean verifier, while the fast C++ verifier has been tested for agreement with it.

ThereforeAnthropic says customers and evaluators could check model identity without possessing the weights or rerunning the model themselves.