Anthropic FlashProve: prove which model answered, at 70x cost

Anthropic released FlashProve, an open-source prototype that proves which model produced an AI response, in a research post dated November 2 and covered by TechCrunch on November 3. A provider publishes a short fingerprint of its exact model weights, then produces a cryptographic proof for each request that those weights generated those tokens. The buyer checks the proof on a laptop. Nobody has to trust the vendor's word about which model ran. The catch is the price: about 70 to 95 times the GPU time of normal inference.

What actually happened

FlashProve comes from Anthropic's Frontier Red Team with researchers from EPFL and MIT. It is a zkSNARK, a zero-knowledge proof, built for mixture-of-experts models. The numbers from the post and the GitHub repo:

• Model tested: Llama 4 Maverick, 400 billion parameters, public 4-bit weights. Not Claude.
• Prover speed: about 110 tokens per second on one H100. A batch of 24 requests takes about 223 seconds.
• Overhead: about 72 times the GPU time of vLLM on the same batch, up to about 94 times at vLLM's highest load.
• Proof size: about 19 MB per 1,024-token request. The one-time weight commitment takes about 39 minutes.
• Verification: a 6.5 MB download, then 17 to 21 seconds on a CPU. No GPU.
• Guarantees: completeness, soundness and zero knowledge are proved in Lean. Everything is Apache 2.0.

The limits are real. Context is capped at 1,024 tokens. The proof covers an integer version of the model under greedy decoding, so sampled outputs may not match. Proving happens after the fact, not in line. The repo is marked "not maintained," and Anthropic calls it "intermediate progress, and not a completed research agenda." The post also says Anthropic met its own roadmap goal through a separate, closed project that signs its inference traffic and recomputes it on a trusted server.

For scale, the prior academic system on the same model ran at 0.02 tokens per second with a 93.6 GB proof. FlashProve is far faster, and still two orders of magnitude from free.

Why it matters for your business

Most small businesses do not call Anthropic or OpenAI directly. They buy through a wrapper: a support-bot vendor, a CRM add-on, an agent platform. The invoice says "powered by a frontier model." You cannot check that. A vendor under margin pressure can route your traffic to a cheaper model tonight. FlashProve shows the industry is working on a proof. You cannot buy one this year.

What you can do now costs nothing:

• Log the model ID on every call. Every major API returns the exact model version. If your vendor will not expose it, ask why.
• Pin versions in your contracts. "A frontier model" is not a term. "claude-opus-5-5 or a successor we approve in writing" is.
• Run a weekly eval. Twenty fixed prompts, scored the same way each time. A silent swap shows up as a step change.

Sources: Anthropic, Progress on provable inference for large language models, FlashProve on GitHub, TechCrunch.