You pay for one model and you get an answer. How do you know it came from that model? You don't. Anthropic says it plainly: a user today cannot check that a response came from the specific model the developer claims is responding. It could be a different one, secretly swapped in by a routing service to save cost, or an attacker could have tampered with the token stream. The answer alone won't tell you, even when it is good.
Fact
The facts: on 2 November 2026 Anthropic (the Frontier Red Team) published research on provable inference for large language models and a system it calls FlashProve. According to Anthropic, for Llama 4 Maverick (400B parameters, mixture of experts, 4-bit quantized weights, integer arithmetic in prefill) a single H100 generates proofs at around 110 tokens/second, with a proof of about 19 MB per 1,024-token request. That is about 70-95 times the GPU time of running the model itself on vLLM. The proofs hide the weights of the model; hiding the architecture is not a goal of the system. The code includes a prover and two verifiers: a fast one in C++ and a provably correct one in Lean. The code, Lean proofs and documentation are released under the Apache 2.0 license. Anthropic says the post shares intermediate progress and is not a completed research agenda.
The medicine is called a zkSNARK. In Anthropic's words, it is a system in which a prover convinces a verifier, by way of a short certificate, that a given computation was executed correctly. The certificate reveals nothing beyond that correctness and is exponentially faster to verify than the computation itself. The provider commits to one model and then proves that every response is its own.
The price sits in Anthropic's numbers: about 70-95 times the GPU time of the model itself on vLLM. That is why it is a prototype for now. The company adds that earlier open-source systems are orders of magnitude slower, so part of the road is covered, but not all of it.
The proof does not grade the answer. It only says whose it is.
What interests me more is how they guard against their own model. Anthropic says Claude is good at optimization but often makes decisions that are hard to evaluate, especially in large, complex systems. So for the guarantees it provides auditable top-level statements in Lean, a formal verification language. A machine speeds it up, and the audit happens in Lean.
The weights stay hidden. The architecture is not meant to be. The code is released with a clear goal: that others take it and improve it. The test is whether anyone brings down those 70-95 times. Until then it is a prototype.
Official primary sources
Anthropic - Progress on provable inference for large language models, 02.11.2026The visual is generated code-art, no third-party images.