AI

Anthropic Launches FlashProve for Verifiable AI Inference

Tuesday, November 3, 2026Read Original

Details

  • Anthropic introduced FlashProve, a research prototype that cryptographically verifies whether a response came from a specified AI model.
  • The system uses zero-knowledge proofs, allowing providers to prove inference occurred with committed model weights without revealing those weights.
  • FlashProve was tested on Meta’s Llama 4 Maverick, a large mixture-of-experts model, using a single NVIDIA H100 GPU.
  • Anthropic reports proof generation at roughly 110 tokens per second, with proofs measuring about 19 MB for a 1,024-token request.
  • Users can verify proofs on an ordinary CPU without access to the model weights, creating an auditable link between a response and its claimed model.
  • The prototype currently requires about 70 to 95 times the GPU time of conventional inference and remains unsuitable for production-scale deployment.
  • Anthropic released the research code as an open-source prototype, positioning FlashProve as a foundation for future model provenance and compliance systems.

Impact

FlashProve addresses a growing trust problem as organizations rely on hosted models but cannot independently confirm which system generated an answer. Its cryptographic approach could support regulated AI, model licensing, and agent accountability without exposing proprietary weights. However, the substantial proving overhead, large proof files, and limited context window make near-term use more likely in high-value audits than everyday inference. The release also gives Anthropic a research position in verifiable AI infrastructure, although practical adoption will depend on major reductions in cost and latency.

Rift Dispatch