Anthropic's FlashProve aims to prove an AI answer came from the model you paid for
Anthropic's research proposes attaching a cryptographic proof to each AI reply, so users can check that a cheaper model wasn't quietly substituted.
The short version
- FlashProve commits to a model's weights, then attaches a proof to every reply
- It uses zkSNARKs, a proof type that keeps the model's weights secret
- Anthropic's report names Starknet and Mantle as examples of the same technique
- StarkWare and Succinct appear as co-authors of the research
- Mantle is not a commercial partner and did not take part in training Claude
Why it matters
Today, anyone buying AI by the answer has to take the provider's word for which model ran. A workable proof would turn that promise into something checkable. It could matter most for businesses that resell AI access or depend on a specific model's quality.
The trust gap in AI services
When you use an AI service, you see only the text that comes back. The model itself runs on the provider's servers, so you cannot tell whether the flagship model answered or a cheaper, smaller one did.
TechFlow notes that this worry applies in particular to API routing services, which could swap in a cheaper model behind the scenes and pocket the price difference without users noticing.
What FlashProve does
Per TechFlow's account of Anthropic's research, the provider first makes a mathematical commitment to the specific weights of a neural network. Weights are the learned numbers that make up a model. The commitment cannot be changed afterward.
Each reply then ships with a proof that the full, committed model computed it faithfully. The proofs are zkSNARKs, a kind of zero-knowledge proof: it shows a statement is true without revealing the secret data behind it. Here, that means the weights stay private. Anthropic says the proofs are small and cheap to check, adding only milliseconds of overhead.
Why Mantle and Starknet appear
Anthropic's report says FlashProve rests on the same cryptographic approach that Ethereum Layer 2 networks use to verify state changes. A Layer 2 is a network built on top of Ethereum to process transactions more cheaply. The report names Starknet and Mantle as examples.
StarkWare, which built Starknet, and Succinct, which supplies the proving network behind Mantle, are both listed as co-authors. TechFlow says Mantle upgraded to OP Succinct in September 2025, becoming a true ZK rollup. It says that cut the state-finalization time to about 1 hour and withdrawals to under 6 hours, from 7 days.
What this does and does not mean
TechFlow is clear that Mantle was not involved in training Anthropic's models and that this is not a purchasing deal. Its argument is that years of real-money use pushed proof costs down, and that AI research can now borrow that engineering.
The piece is written from a crypto perspective, and its claims about Mantle's rank and costs come from TechFlow and the data sources it cites. They are not Anthropic's claims.
What to watch
- Will Anthropic or other providers offer proofs like this in real products?
- Can proofs stay cheap on very large models at everyday volumes?
Try it
If you resell or depend on a particular model, ask your provider in writing how it verifies which model served each answer.
Sources and links
- ReportingTechFlow: 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 TechFlow · 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.