Details
- Anthropic released FlashProve, an open-source research prototype for proving that a response came from a specific AI model.
- The system uses cryptographic techniques to bind proofs to an exact set of model weights while keeping those weights private.
- Model providers first publish a short commitment, or fingerprint, for the selected weights; they can then generate a proof for each response.
- Users can verify the proof and confirm that the committed model produced the supplied output, rather than relying on the response alone.
- Anthropic tested FlashProve on Meta’s 400-billion-parameter Llama 4 Maverick mixture-of-experts model with 4-bit quantization.
- The prototype generated proofs at more than 100 tokens per second on one H100 GPU, with proofs of about 19 MB for a 1,024-token request.
- Proof generation required roughly 70 to 95 times the GPU time of vLLM processing an equivalent prompt, limiting immediate deployment.
- FlashProve is research software rather than a production Anthropic service, and its code and technical specification are publicly available.
Impact
FlashProve targets a growing trust problem in AI: users cannot normally verify which model generated an answer. Its open-source design could support auditable enterprise, regulated, and agentic systems, but the substantial compute overhead and large proof files make routine use impractical today. The work gives Anthropic a visible role in model provenance and may pressure rival providers to explore similar verification mechanisms as synthetic content, model substitution, and compliance requirements increase.