Anthropic open-sources FlashProve for provable AI inference Anthropic

Anthropic has released FlashProve, an open-source cryptographic tool and protocol that lets AI providers prove which model generated an answer without revealing model weights. The code and formal Lean 4 verification proofs are available under the Apache 2.0 license on GitHub.

The system addresses model substitution, security auditing, and compliance in multi-provider environments. Without provable inference, enterprise routing layers cannot verify if an API response came from the requested model or a cheaper substitute. FlashProve attaches a lightweight cryptographic proof to each response, allowing any user to verify authenticity on a standard CPU.

What it is

• Cryptographic proof. FlashProve compiles an interactive oracle proof into a non-interactive zero-knowledge argument (zkSNARK) to verify model computations.
• Weight privacy. The provider publishes a cryptographic commitment ID of its weights. Model weights remain fully secret during verification.
• Single GPU proving. The prover runs on a single NVIDIA H100 GPU and generates proofs for a 400B-parameter mixture-of-experts model (Llama 4 Maverick, 4-bit) at 110 tokens per second across 24 concurrent requests.
• Fast CPU verification. Proof files measure approximately 18.8 MB for a 1,024-token context window. A client verifies the proof on a consumer CPU in roughly 17 seconds using cp_verify.
• Formally verified soundness. Core protocol properties and soundness theorems are formally specified and proved using the Lean 4 proof assistant.
• Agent-optimized development. Anthropic used Claude to explore optimizations across cryptographic literature and structure Lean proof lemmas.

Sources

• Anthropic Research: Progress on provable inference for large language models
• GitHub: FlashProve repository