FlashProve: Anthropic ships a cryptographic receipt for which model actually answered you
You pay a router for a frontier model. Did you get one? Today there is no way to know, since a good cheap model can pass for it on any given answer. Anthropic released FlashProve on Tuesday, an open source system that attaches a proof to a response saying this exact committed model produced it, without revealing the weights.
The numbers are the news. On Llama 4 Maverick, a 400B-parameter model at 4-bit quantization, a single H100 generates proofs at about 110 tokens per second. A proof is about 19 MB per 1,024-token request and checks in 17 seconds on a 24-thread CPU. The earlier proof of the same model that Anthropic cites took 14.3 hours for 1,000 tokens and produced a 93.6 GB proof that took almost a day to verify. Proving still costs about 70 to 95 times the GPU time of running the model, which is why the design proves after the fact, on a subset of tokens.
The method is its own story. Anthropic says Claude produced the zkSNARK improvements and wrote the protocol spec. It also says Claude makes optimization decisions that are hard to evaluate, so the top-level security claims are stated in Lean, with a verifier in Lean too. An AI builds a system too intricate to review by eye, and the machine-checked statements are what you audit. Code, proofs and docs are Apache 2.0.
Now the limits, which Anthropic lists itself. It is a research prototype. It proves an integer version of the model under greedy decoding, so a floating-point or sampled reply may not match. Contexts top out at 1,024 tokens, and an agent loop runs hundreds of thousands. It was measured on an open model, not on Claude. So nobody's coding agent gets receipts next quarter. Anthropic also pitches this for verifying compliance with AI pacing rules, while noting it cannot show that hardware is not training on the side.
Provenance had a second update the same day. Claude's text watermark now covers Mythos 5, Opus 4.7 and Sonnet 4.6, as Anthropic retrofits older models under its signing of the EU Article 50(2) code. On Sonnet 4.6 it may not hold if you set temperature, top_p or top_k to non-default values, and Bedrock rollout finishes by November 10. One mark says who wrote this text. The other would prove which model served this call. The second one is the one agent buyers should be asking routers about.
Post: https://www.anthropic.com/research/provable-inference Code: https://github.com/anthropics/FlashProve Watermark: https://support.claude.com/en/articles/16266773-how-claude-marks-ai-generated-content
← Back to all articles
The numbers are the news. On Llama 4 Maverick, a 400B-parameter model at 4-bit quantization, a single H100 generates proofs at about 110 tokens per second. A proof is about 19 MB per 1,024-token request and checks in 17 seconds on a 24-thread CPU. The earlier proof of the same model that Anthropic cites took 14.3 hours for 1,000 tokens and produced a 93.6 GB proof that took almost a day to verify. Proving still costs about 70 to 95 times the GPU time of running the model, which is why the design proves after the fact, on a subset of tokens.
The method is its own story. Anthropic says Claude produced the zkSNARK improvements and wrote the protocol spec. It also says Claude makes optimization decisions that are hard to evaluate, so the top-level security claims are stated in Lean, with a verifier in Lean too. An AI builds a system too intricate to review by eye, and the machine-checked statements are what you audit. Code, proofs and docs are Apache 2.0.
Now the limits, which Anthropic lists itself. It is a research prototype. It proves an integer version of the model under greedy decoding, so a floating-point or sampled reply may not match. Contexts top out at 1,024 tokens, and an agent loop runs hundreds of thousands. It was measured on an open model, not on Claude. So nobody's coding agent gets receipts next quarter. Anthropic also pitches this for verifying compliance with AI pacing rules, while noting it cannot show that hardware is not training on the side.
Provenance had a second update the same day. Claude's text watermark now covers Mythos 5, Opus 4.7 and Sonnet 4.6, as Anthropic retrofits older models under its signing of the EU Article 50(2) code. On Sonnet 4.6 it may not hold if you set temperature, top_p or top_k to non-default values, and Bedrock rollout finishes by November 10. One mark says who wrote this text. The other would prove which model served this call. The second one is the one agent buyers should be asking routers about.
Post: https://www.anthropic.com/research/provable-inference Code: https://github.com/anthropics/FlashProve Watermark: https://support.claude.com/en/articles/16266773-how-claude-marks-ai-generated-content
Comments