Anthropic announced FlashProve publicly, citing efficiency gains over prior methods
Updated todayDay 24 sources
Anthropic Releases FlashProve, Renames Initiative to Inference Verification
The gist
The core security concern Anthropic identified is that an AI model could misrepresent its own weights, for example running as a more capable model while presenting as a weaker one to evade safety classifiers and monitoring. FlashProve offers a technical path toward verifying that declared model weights are the ones actually running on servers.
When stories like this move, I write about them in an email each morning.
One free email each morning, around 8 AM US Eastern.
Check your email.
I have sent you a confirmation link. One click and you are in.
Not in your inbox? Look in spam or promotions. If it's there, mark it "Not spam" so future editions land in your inbox.
The full picture
Anthropic released FlashProve, an open-source cryptographic tool that lets anyone verify which AI model generated a given response. The tool uses zero-knowledge succinct arguments of knowledge, zkSNARKs, in a software-only design that requires no trusted hardware. During an offline phase, a provider publishes a short commitment identifying exact model weights without revealing them; anyone holding that commitment, the prompt, and the reply can then check a proof. Testing on Meta's Llama 4 Maverick achieved proof generation at 110 tokens per second on a single Nvidia H100 GPU, with proofs of roughly 19MB each; verification on a standard CPU takes about 17 seconds. That compares to the prior open-source method VerInf, which produced 93.6GB proofs requiring nearly a day to check. The system's security guarantees are formally verified in Lean 4. The repository is versioned at 0.16.0 and is labeled an unmaintained research artifact not accepting contributions, released under the Apache 2.0 license. Separately, Anthropic changed the name of its broader initiative from 'provable inference' to 'inference verification' as of October 19, 2026, stating the new term reflects a broader approach than cryptographic proofs alone.
How it developed
Anthropic released FlashProve as an open-source research prototype under Apache 2.0
Sources
Rohan Paul on X
- OpenAI's internal Astra model produced Lean 4-verified proofs for ten long-standing open mathematics problems
- DeepSeek's DeepInfer inference engine
- Claude's formal verification of open-source libraries
- Alibaba's Qwen3.8 model releases
BackgroundFrontier models, open models, and how they are compared
Get the daily email
When stories like this move, I write about them in an email each morning.
One free email each morning, around 8 AM US Eastern.
Check your email.
I have sent you a confirmation link. One click and you are in.
Not in your inbox? Look in spam or promotions. If it's there, mark it "Not spam" so future editions land in your inbox.