Anthropic Prototype Proves Which AI Model Wrote an Answer

Anthropic has published FlashProve, an open-source prototype that can prove which model wrote the answer you got. TechCrunch reports the post went up on Tuesday, November 3.

The code is on GitHub.

If you pay for a top model through someone else’s tool, The Decoder notes you can’t check today which model wrote the answer, and a routing service could swap in a cheaper model.

How FlashProve works

The lab that runs a model publishes a fingerprint of it, and every answer then comes with a compact proof of which model wrote the answer. That fingerprint doesn’t reveal the weights.

So closed labs like Anthropic and OpenAI could take part without giving away the thing they guard most.

It runs on zkSNARKs, a kind of zero-knowledge proof. No trusted hardware needed.

TechCrunch says it should hold up even when the impostor is the lab itself, trying to swap in a cheaper model to cut costs or a pricier one to inflate its results.

Here are the numbers from Anthropic’s technical report, as The Decoder sums them up:

• About 110 tokens of proofs per second on one H100 GPU, tested on Meta’s Llama 4 Maverick (400 billion parameters)
• A proof of about 19 MB that a normal CPU checks in 17 seconds
• Roughly 70 to 95 times the cost of running the model without proofs

Heavy? Yes. Still, Anthropic says the only comparable open-source system took over 14 hours for 1,000 tokens, with a proof of more than 90 GB.

What it can’t do yet

Anthropic calls FlashProve a research prototype, and the limits are real:

• It only proves contexts of up to 1,024 tokens, far below what current models take in.
• It hides the weights, but not the model’s architecture.
• It only works when the model picks the single most likely next token, not with the randomized answers chatbots give you every day.

Those are big gaps. So don’t expect a “verified” badge in ChatGPT or Claude soon.

According to TechCrunch, no lab, Anthropic included, plans to deploy it. The researchers call it intermediate progress and want others to take the code and improve it.

The Decoder adds that Anthropic’s Frontier Safety Roadmap aimed for an inference verification prototype by September 2026; Anthropic says a trusted-server project already met it.

That said, FlashProve tries the same thing with cryptography alone.

Why it matters

Think about the tools you pay for that say they’re “powered by” a top model:

• the writing tool that drafts your blog posts
• a support chatbot on your site
• the AI gateway your developer set up

You take their word. Everyone does.

My read: it matters most if you buy AI secondhand. If a tool you pay for bills you for a frontier model and swaps in a cheaper model, your output gets worse and you never know why.

A proof of which model wrote the answer, checked on your own laptop, would turn “trust us” into something you test.

TechCrunch gives other cases: a developer who says a frontier security model scanned their library for flaws, or a firm saying a suitable model approved its business plan.

How would you verify that today? You can’t.

But it’s still early.

It’s also a different job from Anthropic’s hidden text watermark, which marks Claude’s text so you can spot it later.

What should you do now? Nothing in your stack changes this week, so here’s where I’d start:

• If you build on AI APIs, watch the FlashProve repo on GitHub.
• Ask the tools you pay for which model, and which version, answers your requests.
• Ask your providers whether they’d publish model fingerprints if this matures.

Common Questions

What is Anthropic’s FlashProve and what does it do?

FlashProve is an open-source research prototype from Anthropic that creates a cryptographic proof of which model wrote the answer to a prompt. Anthropic published it on November 3, 2026, with the code on GitHub under the Apache 2.0 license.

Can I use FlashProve to check which model wrote the answer in ChatGPT or Claude?

Not yet. No AI lab, Anthropic included, plans to deploy FlashProve, and the prototype only handles contexts of up to 1,024 tokens where the model always picks its most likely next token.

Does FlashProve reveal an AI model’s weights?

No. The lab publishes a fingerprint of the model, and the proofs don’t expose its weights, though they don’t hide the model’s architecture.