Anthropic FlashProve is an open-source system that lets an AI provider prove a reply came from one specific model, without showing anyone the model’s weights. Anthropic announced it on November 2, 2026, and it is a research prototype, not a product you can switch on today.
The idea is simple to state. When you pay for a big model, you currently have to trust the provider that it didn’t quietly swap in a cheaper one. FlashProve tries to replace that trust with math.
Quick facts before the details:
- Announced November 2, 2026, with code released under the Apache 2.0 license.
- Built with researchers at EPFL, MIT, StarkWare and Succinct, per AI Weekly.
- Tested on Llama 4 Maverick, a 400-billion-parameter model, running on a single H100 GPU.
- Proofs run about 19 MB per 1,024-token request and check in roughly 17 seconds on a CPU.
Those four lines tell you the shape of it: real, open and impressive, but still early.
What Anthropic FlashProve actually proves
Today, an API call is a black box. You send a prompt, you get text back, and you hope it came from the model you chose and paid for. Nothing in the reply proves it.
FlashProve adds a receipt. The provider commits to a specific set of model weights ahead of time. Then, for each reply, it produces a cryptographic proof that those exact weights produced that exact output.
The weights stay private the whole time. That’s what “zero-knowledge” means here: you can check the claim without learning the secret. Anthropic frames it as a way to catch model substitution, tampering with token streams, and cheating in independent tests, according to Blockchain.News.
How the proof check works in practice
Think of it as two steps. First, you download a small commitment, about 6.5 MB, that stands for the model’s weights. You only do this once.
After that, each reply comes with a proof of about 19 MB for a 1,024-token request. Your machine checks it on a normal CPU in about 17 seconds using 24 threads. No GPU is needed on your side.
AI Weekly also reports that the guarantees (completeness, soundness and zero-knowledge) were formally checked in the Lean proof assistant. That matters because cryptography bugs are easy to make and hard to spot.
The cost: proving is much slower than answering
Making the proof is the expensive part. On one H100, the prover handles a bit over 100 tokens per second. That sounds usable until you see the comparison.
Anthropic’s numbers put proving at roughly 70 to 95 times the GPU time of running the same model normally on vLLM. So a proof for every chat message would be very costly today.
Sampling matters too. The system works with integer models and greedy decoding only, which means no creative, randomized sampling yet. Real products usually use sampling.
What FlashProve does not cover
It’s easy to read “provable AI” as “trustworthy AI.” They aren’t the same thing. The table below separates what the proof claims from what it doesn’t.
| Question | Does FlashProve answer it? |
|---|---|
| Did this reply come from the committed model weights? | Yes. That is the core claim. |
| Can the provider keep its weights secret? | Yes, the weights stay hidden. |
| Is the answer correct, safe or unbiased? | No. It proves where the answer came from, not whether it is good. |
| Is the model architecture hidden? | No. Only the weights are hidden. |
| Does it stop side-channel leaks? | No. Anthropic lists this as a limit. |
| Does it handle long chats? | Not yet. Contexts are capped at 1,024 tokens. |
There’s one more caveat. Anthropic’s post is a progress update, and Blockchain.News notes the results haven’t been independently reproduced yet. Treat the speed and size figures as the team’s own measurements.
Why this could matter for you
Most readers won’t run a prover. But the goal behind it touches anyone who pays for AI or builds on it.
If you’re a developer, a proof would let you confirm that an API gave you the model tier you bought. If you’re a company, it could back up compliance claims or independent test results. The same instinct for checking rather than trusting shows up in tools like 1Password for Agents and the replay test behind OpenAI’s AI monitors.
Anthropic also ties the work to its Frontier Safety Roadmap, which set a goal of a provable-inference prototype by September 30, 2026. This release looks like that milestone arriving.
Frequently Asked Questions
Is FlashProve available to regular users?
Not as a product. It’s an open-source research prototype for developers and researchers, with code under Apache 2.0. No launch date for a commercial feature has been given.
Does FlashProve prove an AI answer is correct?
No. It proves which model weights produced the reply. A proven answer can still be wrong, unsafe or biased.
Does it only work with Claude?
No. The demo ran on Meta’s Llama 4 Maverick, a 400-billion-parameter open model, on one H100 GPU.
Why is the proof so big and slow?
Zero-knowledge proofs of huge computations are heavy by nature. Each 1,024-token request produces about 19 MB, and proving takes roughly 70 to 95 times the normal GPU time.
Should you care yet?
Not for daily use, and that’s fine. The real news is that someone proved a 400B-parameter model’s output on a single GPU and published the code, which many people doubted was feasible. If you build on AI APIs, bookmark it and watch for faster provers and longer contexts. If you just chat with AI, nothing changes for you this month.
Leave a Reply