Anthropic publishes FlashProve, a prototype for provable inference
Anthropic's page, dated 2 November 2026, introduces FlashProve. In one case, an H100 generates proofs at about 110 tokens per second. It is a research prototype, not a product.
The page “Progress on provable inference for large language models” is dated 2 November 2026. The team is the Frontier Red Team. The system is called FlashProve. Anthropic describes an open prototype for proving that a response came from the model a provider claims to be serving, without revealing the weights.
The case measured on the page is Llama 4 Maverick, a 400-billion-parameter mixture of experts, 4-bit weights, integer arithmetic in prefill. One H100 generates proofs at about 110 tokens per second. The proof is about 19 MB per 1,024-token request. The page puts that at about 70 to 95 times the GPU time of running the model on vLLM. It says earlier open systems are orders of magnitude slower. In a comparison the page makes, zkLLM in 2024 proved a 13-billion model at about 2.5 tokens per second on one A100.
The page says the code, the Lean proofs, and the documentation are released under Apache 2.0. It links a repository at github.com/anthropics/FlashProve. This desk did not open the code or the technical report.
The limits the page lists: there is no formal guarantee that the fast executable code matches the Lean. The proof is of an integer model under greedy decoding, not floating point and not sampling. Proved context runs up to 1,024 tokens. Hiding the architecture is not a goal. Hiding the weights is. Side-channel attack is out of scope for this release. The page calls the release intermediate progress, not a finished agenda.