Playbooks · Coding
FlashProve: proofs of which model answered you
Written November 9, 2026
You cannot tell from the words alone which model wrote a reply. A routing service could swap in a cheaper model, or an attacker could tamper with the token stream, and the output would still read fine. Anthropic's FlashProve attaches a cryptographic proof to the answer — at 70 to 95 times the cost of serving it.
The problem
When you send a request to an AI model today, you have no way of knowing for certain that the model you selected is the one that generated the response. That matters in three cases Anthropic names in its research post: independent evaluators who test models before and after release need to know the model they are evaluating is the model being served; clients using a model through a third-party routing service want assurance the tokens came from the model they paid for; and if an attacker modifies the token stream while pretending to be a given model, the output alone gives no way to detect it. Even publishing chat logs proves nothing — a developer could fake them with a cheaper model.
FlashProve, from Anthropic's Frontier Red Team, is an open-source prototype for provable inference: a system that lets a provider prove a response came from a specific, committed set of model weights, without revealing those weights. The post is dated November 2, 2026; some coverage of the release is dated November 4. The repository itself was created October 30.
What it proves — and what it does not
FlashProve is built on a zkSNARK, a zero-knowledge succinct argument of knowledge. It works in two phases. In the offline phase, the provider publishes a commitment to the model weights — a short fingerprint that pins down one exact choice of weights while the weights themselves stay hidden. In the online phase, the provider produces a proof file with each reply, proving the statement: the committed model produces these output tokens when evaluated. Anyone holding the commitment, the prompt, and the reply can check it, without the weights and with limited compute. A proof file is accepted if it is honest, cannot be forged without breaking widely believed cryptographic assumptions, reveals nothing about the weights beyond what the response itself implies, and is small and cheap to check.
What it proves is origin, not correctness. A proof says the committed model produced these tokens. It says nothing about whether the tokens are true, good, or safe — a correct proof of a wrong answer is still a proof. The model architecture is not hidden either: hiding the weights is the goal; hiding the architecture is not. And a proof does not show the provider's hardware was used only for inference — Anthropic notes that it verifies the hardware is serving inference, not that it is not training on the side.
What it costs
Anthropic measured FlashProve on Meta's Llama 4 Maverick, a 400-billion-parameter mixture-of-experts model in public 4-bit weights, on a single H100 GPU. All figures are Anthropic's measurements:
| Measure | Figure |
|---|---|
| Proving speed | About 110 tokens per second on one H100 |
| Proof size | About 19 MB per 1,024-token request |
| Checking a proof | 17 seconds on 24 CPU threads (21 seconds on one thread), no GPU needed |
| Overhead vs plain inference | About 70–95 times the GPU time vLLM takes for the same requests |
| Commitment download | About 6.5 MB, once |
| Offline phase | About 39 minutes, once per model; the commitment is good for about 16 million requests |
For scale, the only earlier proof of the same model that Anthropic knows of, VerInf, took 14.3 hours for 1,000 tokens with a 93.6 GB proof that took almost a day to check. FlashProve is orders of magnitude faster than that — and still 70 to 95 times slower than serving the model plain. Anthropic frames the release as intermediate progress, not a completed research agenda, and says it has no plans to deploy FlashProve at Anthropic or any other lab.
The limits that keep it a prototype
- Short contexts. The system proves inference over contexts of up to 1,024 tokens — short against chat sessions that accumulate hundreds of thousands or millions.
- Greedy decoding of an integer model. A proof asserts each checked token is the model's greedy choice under an exact integer version of the published weights. A reply from a floating-point implementation, or sampled decoding, could depart at some position and no longer match the computation proved.
- Mixture-of-experts only. The system assumes an MoE architecture and uses that structure for its efficiency gains. Dense models are not covered.
- Side channels are out of scope. The proof itself reveals nothing about the weights, but measuring the time it takes to produce a proof could potentially reveal information. Anthropic considers that out of scope for this release.
- Not maintained. The repository describes itself as a research artifact: version 0.16.0, not maintained, not accepting contributions. The code, Lean proofs and documentation are released under the Apache 2.0 license, with the security properties formally stated in Lean — but Anthropic gives no guarantee that the Lean protocol, the paper, and the fast executable code match each other.
Who would use it, then? Evaluators and auditors who need to know the model they tested is the model that was served, and regulators who need to check that only approved models are served, without rerunning the model themselves. For everyone else, the honest summary is the one Anthropic gives: a fast prototype that shows provable inference is practical, at a cost no production service would pay today.
Limits
- Every performance figure on this page is Anthropic's own measurement, on its own hardware, for one model (Llama 4 Maverick at 4-bit weights). No proof was built or run on this machine, and no independent measurement is cited.
- The publication date is carried as a conflict: Anthropic's post is dated November 2, 2026, while some coverage of the release is dated November 4. The repository was created October 30, 2026.
- FlashProve proves which committed weights produced a reply. It does not prove the reply is correct, safe, or the best answer, and it does not prove the provider's hardware did nothing else.
Related playbooks
- BotPsy 0.8B: an agent model that fits on-device
- Check whether your session is really on Muse Wonder or Wonder Full — and which route you are paying through
- DeepSeek, Exa, Brave, Tavily: which web-search API should your agent actually pay for now that search is metered
- DeepSeek's cheap coding model changed under the same name: retest Flash 1017 before you rely on it
Sources
- Progress on provable inference for large language models — Anthropic
- FlashProve — Anthropic, GitHub
- FlashProve: an open-source library for provable inference — Anthropic technical report
Back to: Coding playbooks hub · Also: Astra did 40% less work on safety tasks · OpenAI: agents tried to escape their sandboxes