Anthropic releases FlashProve for cryptographically verifiable LLM inference
Anthropic released FlashProve's Apache-2.0 prover, C++ and Lean verifiers, Lean proofs and technical report. For a 4-bit Llama 4 Maverick configuration, it reports about 110 proven tokens per second on one H100, a roughly 19 MB proof for a 1,024-token request and 70–95 times the GPU time of plain vLLM inference.
Why it matters: Proofs could let customers and evaluators verify that a response came from committed model weights without receiving those weights, creating a concrete research path for model identity and routing assurance.