Anthropic open-sources FlashProve, a research prototype for proving which AI model produced an answer
Announced · Published · By Contractor AI Shop
Anthropic's Frontier Red Team released FlashProve under Apache 2.0, a research prototype that produces cryptographic proofs that a specific model generated a given output. Anthropic reports about 110 tokens/s proving on one H100, ~19 MB proofs per 1,024-token request, and 17-second CPU verification.
Key facts
- FlashProve is a research prototype released under the Apache 2.0 license, per Anthropic. Anthropic
- Anthropic-reported figures: ~110 tokens/s proving Llama 4 Maverick on one H100, ~19 MB proofs per 1,024-token request, 17-second CPU verification; limited to 1,024-token contexts and MoE models. Anthropic
- TechCrunch reports there are no plans to deploy it at Anthropic or any other lab yet. TechCrunch (Russell Brandom)
Our take: what it means for contractors
Low near-term relevance for contractors. It matters as a sign that "prove which model you actually ran" is becoming checkable, which could help later with vendor claims and audits.
What we don't know yet
- All speed and size figures are Anthropic's own
- Whether any vendor will adopt it
Sources
- Anthropic · Provable inference (FlashProve) · 2026-11-02 · Primary
- TechCrunch (Russell Brandom) · Anthropic is working on a system to prove AI models' identity · 2026-11-03 · Secondary
- Crypto Briefing (Diego Almada Lopez) · Anthropic FlashProve zkSNARK AI inference · 2026-11-03 · Secondary