Anthropic's FlashProve aims to expose resold AI models that are not what they claim
Anthropic released open-source FlashProve, which lets anyone check on an ordinary CPU in 17 seconds whether an answer came from the model they paid for.
The short version
- FlashProve attaches a cryptographic proof of about 19MB to each AI answer.
- A normal CPU checks the proof in 17 seconds, with no model weights and no GPU.
- The test used one H100 GPU with 80GB of memory, streaming a 232GB model through it.
- Claude wrote most of the code and math proofs; humans fixed 6 top-level security definitions.
- Everything is open source under the Apache 2.0 license; speed claims are not independently tested.
Why it matters
Many people buy AI access through resellers and have no way to know which model really answers. If proofs like this become standard, buyers and auditors could check instead of trusting. For now, the numbers come from Anthropic's side and still need outside testing.
The problem with model resellers
Some services sell access to a top-tier AI model but may quietly run a smaller, cheaper one behind the scenes. Customers cannot see the machines or the model's internal settings, so they have little proof either way. Asking a chatbot who it is does not help, because that answer is itself just generated text.
As PEdaily reported, Anthropic's new FlashProve is aimed at exactly this gray area, sometimes called "wrapper AI" swapping.
How the three-step check works
First, before launch, the provider publishes a unique fingerprint of the model's weights (the numbers that make up the model). Changing even one number would break the match, yet the fingerprint reveals nothing about the weights themselves.
Second, for each request, the provider's GPU produces a proof of about 19MB showing the answer came from that fingerprinted model. Third, anyone can run a verifier program on a normal CPU and get a true-or-false result in 17 seconds, without the weights or a GPU.
The method is a zkSNARK, a kind of zero-knowledge proof: it convinces a checker something is true without revealing the secret behind it.
Proving a huge model on one GPU
The team tested it on Meta's Llama 4 Maverick, a mixture-of-experts model. That design activates only a small slice of the model for each word, so the proof work scales with the active part. The quantized weights take 232GB, while the H100 has only 80GB of memory, so the system streams the weights through in segments.
The headline's 5000x speedup is a vendor claim, and we have no independent benchmark for it.
Claude built most of it
Most of the code, the paper and a huge body of math proofs were written by Claude, according to the source. The proofs use Lean 4, a language whose checker confirms each step. Humans mainly set 6 top-level security definitions.
These cover completeness (honest providers pass), soundness (cheating is practically impossible) and zero knowledge (no weights leak). The code, proofs and documents are open source under Apache 2.0.
What to watch
- Will independent researchers reproduce the 17-second and speedup claims?
- Will resellers and API providers actually adopt proof-based verification?
Try it
When using a discounted or third-party AI app, ask the provider whether it can prove which model it runs, and compare its answers with the official service on a few of your own tasks.
Sources and links
- ReportingPEdaily: the article we sourced
Liked it? Say why
Add a line for other readers.
No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.
Listen to the AI Wire headlines
Listen to the AI WireListen to our daily AI shows every morning: on your commute, in the shower, wherever you are
Listen to the daily showsFirst reported by PEdaily · Read the original. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.