Anthropic advances proofs for model inference
Anthropic introduced FlashProve, an open prototype of cryptographic proofs for checking whether a response came from a specific model.
- FlashProve produces cryptographic evidence that a claimed model generated a response.
- The demonstration uses an H100 GPU and has large proofs and meaningful compute costs.
- Anthropic describes the post as intermediate progress that can be improved.
Context and analysis by Noosigma AI
Editorial summary and context based on the cited source. Claims and results attributed to the organization are reports from that source; they do not constitute independent validation by Noosigma AI.
The proposal lets a provider attach a verifiable proof to a response without exposing the model’s weights. This could help customers and evaluators detect model substitutions in routing services.
Why it matters
In a test with the 400-billion-parameter Llama 4 Maverick quantized to 4-bit weights, the company reports proof generation at about 110 tokens per second on one H100 GPU. Each 1,024-token response produced a proof of roughly 19 MB, with GPU cost estimated at 70 to 95 times the vLLM runtime used as a reference.
This is a research prototype: the artifacts are released for evaluation and further development, and the current cost remains high. The system hides model weights but is not designed to conceal the model architecture.
What to check
Source: Anthropic, November 2, 2026.