Independent newsroom

Tuesday, November 3, 2026

The File

Every claim traces to a primary source. Uncertainty is labeled. Corrections are public.

AI · Tuesday, November 3, 2026

Anthropic posts a proof that a named model gave the answer

The File desk · Nov. 3, 2026, 6:33 a.m. ET

Confirmed1 primary sources

Status

Confirmed. Anthropic's Frontier Red Team posted on November 2, 2026, a research prototype called FlashProve for proving that a reply came from a specific model. The speed figures in the post are Anthropic's measurements on Llama 4 Maverick, not an independent test.
  • Confirmed
    The page is dated November 2, 2026, and describes FlashProve as an open-source prototype for provable inference.
  • Confirmed
    The post says one H100 chip produced proofs at more than 100 tokens a second, with a proof of about 19 megabytes for a 1,024-token request.
  • Partially confirmed
    Those speeds are the lab's own measurements. This desk did not run the code.
Anthropic's Frontier Red Team said on November 2, 2026, that it has a research prototype, called FlashProve, for proving a reply came from a specific model. On one large model, the lab says, a single graphics chip produced those proofs at more than 100 tokens a second.1
Someone talking to a model usually cannot check that the reply came from the model the company named. Anthropic describes a zkSNARK as a short certificate that a computation was done correctly. The certificate, it writes, "reveals no information beyond what is implied by the computation's correctness," and it is meant to be much faster to check than to redo the work. If the reply did not come from the committed model, the check fails. The proof is not supposed to show the model's weights, the numbers that make the model what it is.1
Why it matters (Estimate). Estimate. The post is for anyone who needs to know that the model they were promised is the one that answered, including an outside tester or a customer reaching a model through a middleman. Anthropic says a user can check a short proof without holding the model's weights. It also says this release is a research prototype, and the speed it prints was measured on Llama 4 Maverick, not described as a product customers can switch on today.
The post says FlashProve was measured on Llama 4 Maverick. It says one H100 chip generated proofs at more than 100 tokens a second, and that a proof for a 1,024-token request was about 19 megabytes. Those speeds are Anthropic's measurements; this desk did not run the code. The lab says the code, the Lean proofs, and the documentation are under the Apache 2.0 license, and it calls the post intermediate progress rather than a finished system.1
How we know

Notes

  1. 1.Anthropic, Frontier Red Team, "Progress on provable inference for large language models," page dated November 2, 2026. The post says a zkSNARK lets a prover convince a verifier, by a short certificate, that a computation was executed correctly, and that the certificate reveals no information beyond what is implied by the computation's correctness. It releases FlashProve and says, for Llama 4 Maverick, a 400-billion-parameter mixture-of-experts model at 4-bit quantization, a single H100 GPU generates proofs at over 100 tokens a second, with a proof size of about 19 megabytes per 1,024-token request, about 70 to 95 times the GPU time that vLLM takes to process prompts of the same length. It says the code, Lean proofs, and documentation are under the Apache 2.0 license, and that the post shares intermediate progress. The visible date line reads Nov 2, 2026. The page metadata published time is 2026-11-02T00:00:00.000Z and the modified time is 2026-11-03T10:44:38.000Z. https://www.anthropic.com/research/provable-inference

What is still unknown or disputed

  • Whether the speed holds on Anthropic's own models. The measurement in the post is on Llama 4 Maverick.
  • An independent rerun. This desk did not run the code.
  • What changed at the page's modified timestamp. The visible date is November 2, 2026. A machine timestamp reads 2026-11-02T00:00:00.000Z, which is 7:00 p.m. Eastern on November 1. A modified timestamp reads 2026-11-03T10:44:38.000Z, which is 5:44 a.m. Eastern on November 3. This desk did not open an earlier copy.

Primary sources

Every claim in this story is drawn from the documents below. If a fetch failed, that is recorded on the card.

  1. Source 1

    Progress on provable inference for large language models

    Anthropic, Frontier Red Team · November 2, 2026

    The lab describes FlashProve, a prototype that attaches a short proof that a reply came from a committed model, and reports more than 100 tokens a second on Llama 4 Maverick.

    https://www.anthropic.com/research/provable-inference