Anthropic FlashProve proves which model answered with a zkSNARK proof

📅 November 3, 2026 • TechCrunch / Anthropic

Anthropic published a cryptographic answer to a question users normally cannot check: did the claimed model actually produce this response? In a research post titled “Progress on provable inference for large language models”, the company introduces FlashProve — a zkSNARK-based proof that an answer was generated by exactly the stated model, without revealing the weights. zkSNARKs are a cryptographic technique that lets one party prove a statement is true while keeping the underlying data hidden — here, that a specific generation came from specific weights that stay sealed. Per Anthropic, the proof travels with the response, so a user or an auditor can verify the origin of an output independently of the provider's assurances. The work lands as third-party routers and agent stacks increasingly sit between users and frontier models.

The published numbers show both feasibility and cost. On Llama 4 Maverick — a mixture-of-experts model with 400B parameters quantized to 4 bits — a single H100 GPU generates proofs at about 110 tokens per second, and each proof weighs roughly 19 MB for a 1024-token request. Verification is the cheap side: checking a proof takes 17 to 21 seconds on an ordinary CPU with no GPU at all. Proving is the expensive side: per the technical report, it adds 70 to 95 times the GPU time of a standard vLLM inference run. Both sides shape the threat model — cheap verification means anyone can check, while heavy proving forces providers to choose where to spend it.

The stated use cases target router opacity and benchmark trust. Per Anthropic and TechCrunch's coverage, a user can confirm that a routing layer did not silently substitute a cheaper model, and a lab can demonstrate that reported results were not inflated — the proof binds each answer to specific weights. Anthropic has released the code and a Lean formalization of the guarantees under Apache 2.0 on GitHub, alongside a technical report.

This is research, not a deployed product. The company states the system is not in production, and the cost figures explain why: at current overhead, proving every response would multiply inference budgets many times over. The work is positioned as a foundation for verifiable inference — a way to keep accountability intact as traffic to models moves through intermediaries that users cannot inspect.

IMPACT: model provenance shifts from vendor assurances to checkable math — if the 70-95x proving overhead falls as the field iterates, proof-of-model could become a standard attachment to paid API answers, especially behind resellers and routers.

Sources: techcrunch.com, anthropic.com, github.com

Continue the conversation: https://t.me/AIAutomationArchitecture · Bluesky