BUSINESSNew tool · · 2 min read · Source: Anthropic
Anthropic FlashProve: proof of which AI model wrote the answer
Anthropic published FlashProve, a research prototype, on November 2, 2026. It is designed to prove mathematically that a response really came from a specific AI model, though TechCrunch reports there are no plans to deploy it in products.
My tool tip · AdWispr Flow – AI dictationTry it free
Key facts
- Anthropic released FlashProve on November 2, 2026, as an open research prototype (Apache 2.0).
- In a test with Llama 4 Maverick, one H100 GPU produced proofs at about 110 tokens per second, at roughly 72 times the usual compute.
- According to TechCrunch, Anthropic has no current plans to deploy it in its own products or at other labs.
What FlashProve does
When you use AI through a provider today, you have to trust that the model you pay for is the one doing the work. FlashProve tackles that. The provider publishes a short "fingerprint" of its model once. For every response, it then generates a proof that this exact model turned this exact prompt into this exact text. Anyone can check the proof on an ordinary computer without seeing the secret model weights. Under the hood, it uses a zero-knowledge proof.
Key numbers (Anthropic's test)
- Test model: Meta's open Llama 4 Maverick, not Claude.
- Speed: one H100 GPU produces proofs at around 110 tokens per second.
- Overhead: roughly 72 times the compute of a normal response.
- Verification: about 17 seconds on a CPU.
- Limit: context currently capped at 1,024 tokens.
Code, specification and proofs are available under the open Apache 2.0 license.
Why it matters
According to TechCrunch, a system like this could reveal whether a provider quietly swaps in a cheaper model to cut costs. Vendors who advertise "powered by the best model" could be checked too. Anthropic is clear about the limits: it is a research prototype, the overhead makes it too expensive for everyday use, and there is currently no deployment at Anthropic or other labs.
What this means for you
- Today you have to trust providers about which model is behind an answer – in the future, it could be verifiable.
- FlashProve is still too slow and costly for daily use: about 72 times the compute, short texts only.
- The code is open, so other companies and researchers can build on it.
How to use this
Add one line to your AI contracts and vendor requests: "Which model and version do you use, and will you announce any change in advance?" That gets you a written answer long before technical proofs exist.
Luigi's take
This is one of those stories that changes nothing today and may matter a lot in two or three years. Many companies don't buy AI straight from the lab but through agencies, platforms and resellers, where it's often unclear which model is really running. A verifiable proof would be real progress for procurement and quality – much like signatures for email.
My advice: don't wait for the technology. Ask every vendor in writing which model and version they use for your work, and whether they may switch without notice. What's in the contract, you can enforce – no cryptography needed.
Source
Anthropic: Provable inference (FlashProve)
More sources
- TechCrunch – Einordnung: kein Einsatz bei Anthropic oder anderen Laboren geplant
This article was researched and written with AI assistance. Responsible: Luigi Lugmayr.
