Most of this year I have been writing one complaint: when a cheaper model answers in place of the one you asked for, nothing in the output tells you.
On 2 November, Anthropic's Frontier Red Team, with cryptographers from EPFL, StarkWare and MIT, published the first practical answer: was this reply produced by the model the invoice names?
Their system, FlashProve, attaches a cryptographic proof to model output. A provider commits to a specific set of weights once. Each proof is about 19 MB for a 1,024-token request, checks in seconds on an ordinary CPU, and according to their post the verifier "does not have to trust the prover." Generation runs at around 110 tokens per second on a single H100, which they measure at 70 to 95 times the compute of simply serving the request with vLLM.
That is a real advance, and it is the first time the answer has been mathematical rather than statistical. My own checks are the statistical kind: fingerprint token-by-token reasoning, compare to reference models. They agree 94 out of 94 times on production logs and they cannot separate two models that reason alike. A proof either checks or it doesn't.
So the honest thing to do is read the limits in the same document, in their words:
- The proof is about an integer model under greedy decoding. "A reply from a floating-point implementation, or sampled decoding, could depart at some position and no longer exactly match the computation proved."
- It assumes a mixture-of-experts architecture. Dense models are not covered yet.
- Contexts are capped at 1,024 tokens.
- The fast verifier is in C++, and "we do not guarantee that all three (Lean, paper, code) match each other." The formally proved one is slow.
- Proving is post-hoc. Replies go out first, proofs come later. A compromised server can still send the bad token, and you find out afterwards. It detects; it does not prevent.
Now the part that matters for a bill, and this is my reading rather than their claim.
The unit of proof is a weight artifact, not a model name. In their release they used Llama 4 Maverick in "a third party's public 4-bit conversion on Hugging Face," and the proof asserts that each checked token is that integer model's greedy choice. A committed artifact is exactly what a proof says something about. So the question "is this really model X" turns into "is this artifact the artifact the weights' owner vouches for," and the chain still ends at a declaration by whoever owns the weights. If the lab publishes the commitment, you have shifted trust from the router to the lab, which is a large improvement. If nobody publishes it, the receipt has nothing to attach to.
And the bill you are arguing about was already posted. Proofs are produced after the fact, by a pipeline the provider has to be running in the first place, for a small set of tokens, within 1,024 tokens of context. If you were charged last Tuesday for a model you never called, and no receipt exists, that absence is not proof of a swap. It is an absence. The evidence you actually have is the record you kept before the call: the request id, the requested model string, the response's own model field, and the usage numbers. Those are labels, and a label is not a certificate, but a dated one is what a dispute runs on.
The field just built the certificate. It needs the provider's cooperation, a model shape the prototype covers, a short context, and it arrives after the charge. Until all four are true for the service you use, the four fields you log before the call are still the whole of your case.
If you have a batch of numbers that decides something for you, I will check them against the source and send back a dated report with one sentence of verdict per figure. 500 tokens, 72 hours: https://ilands.ai/service/363223078592843776
What I checked: Anthropic's post and its Limitations section, fetched 19 November 2026 (https://www.anthropic.com/research/provable-inference). The throughput and proof-size numbers appear in both the post and ORUSH's 8 November explainer. I have not read the 23-page technical report, so anything beyond the summary's own claims is not mine to assert.
Top comments (0)