Anthropic's FlashProve proves which model answered you

You have no way to know which model answered your last API call. FlashProve is a serious attempt to change that.

Anthropic, with researchers from EPFL, StarkWare and MIT, published a technical report this month on FlashProve, which it calls "an open-source library for provable inference."

What FlashProve proves, and what it costs

The report states the problem in one line: "When a client sends a prompt to a model provider, the client sees only the generated tokens and has no guarantee that the provider uses the claimed model."

FlashProve works in two steps. Once per model, the provider publishes a short commitment to its model weights. Then, for each request, it produces a proof that the committed model generated the output. The proof is zero-knowledge, meaning it reveals nothing about the weights, and succinct, meaning you can check it far faster than running the model yourself.

The team tested it on Meta's Llama 4 Maverick, a 400B-parameter mixture of experts model, using one H100 GPU, according to the report. Its figures:

• The one-time setup per model takes about 39 minutes.
• A batch of 24 requests takes about 223 seconds to prove, or 110 tokens per second.
• Each proof is 18.78 MB.
• Checking a proof takes 17 seconds on a client CPU, with no GPU and no weights.
• Per GPU, proving a batch takes about 72 times longer than simply running the prefill on it.

72 times.

Why the narrow version still matters

The limits are real, and the report lists them. Each proof covers a window of 1,024 tokens and checks one token, the one right after the prompt. It only proves the model's greedy choice, the single most likely next token, so ordinary sampled chat output isn't covered. The prover runs on one GPU.

The report is also candid about how it was built. The GPU prover code and the CPU verifier code were generated by AI from formal specifications in Lean and are not proved equivalent to them. A longer protocol document in the repository was AI-generated and, the authors write, "we did not review it in detail."

So no, this isn't something your provider will turn on next week. It's a working proof that the problem can be solved without trusted hardware. Today, when a provider says you're getting the big model, you believe them or you don't. FlashProve turns that into a check an ordinary CPU can run.

The Decoder reports the code is on GitHub under an Apache 2.0 license.

Who should care now

If you buy inference from resellers, routers or hosts serving open models, this is the tool that could one day prove you got what you paid for. The report names the case directly: a provider "could instead use a smaller/weaker model, or even switch models between requests."

If you're in procurement or compliance, start asking vendors how they would show which model served a request. Most will have no answer yet. That's useful to know.

What to watch

The report promises a carefully reviewed paper later. Watch for sampled decoding support, longer contexts, and whether any provider offers proofs to customers.

The first one that does will have a sales pitch nobody else can match.