Research & models Nov 4, 2026

Anthropic open-sourced FlashProve, a provable-inference prototype, on November 3; a single H100 generates zero-knowledge proofs for a 400-billion-parameter model at about 110 tokens per second.

Generating proofs takes about 70 to 95 times the GPU time of ordinary inference, and only Anthropic has reported these figures. Provable inference lets a user who can't see a model's weights verify that a response really was computed by the committed set of weights. A zero-knowledge proof is a cryptographic proof that can be checked quickly without revealing the computation itself. The test model was Llama 4 Maverick, a 400-billion-parameter mixture-of-experts model, quantized to 4 bits. The proof for a 1,024-token request is about 19 MB. Verification runs on CPUs only and takes 17 seconds on 24 cores (Anthropic Research, 2026-11-03). zkLLM, the earlier system Anthropic lists for comparison, ran at about 2.5 tokens per second on a 13-billion-parameter model. Model size, quantization and hardware all differ between the two, so the comparison only gives a sense of scale. FlashProve proves only that the output came from the committed weights, not that the content is correct. It currently covers only a 1,024-token context, integer arithmetic and greedy decoding, which takes the most probable token at each step. Anthropic's own list of limits also says the fast verifier has no formal guarantee and that proof-generation time may leak information. Anthropic names the uses as letting regulators check that hardware really serves the committed model, and that a router hasn't quietly swapped in a cheaper model. No third party has re-run the performance figures; the code is open-sourced under Apache 2.0. How we checked →