Anthropic Open-Sources FlashProve, a Cryptographic Tool to Prove Which AI Model Generated a Response

Anthropic has released FlashProve, an open-source cryptographic tool that lets anyone verify a response genuinely came from a specific AI model — without the model's owner revealing its proprietary weights. It is the first publicly available system to make this kind of proof practical at real-world inference speeds, and it matters because it gives regulators, auditors, and enterprises a way to hold AI providers accountable for the outputs they claim to produce.

Why This Is a Hard Problem

When a company says "this response was generated by Model X," you have no independent way to verify that claim today. The company could swap in a cheaper, less safe model and you'd never know. For high-stakes uses — clinical decisions, legal filings, government contracts — that's a real gap. Cryptographic proof of inference closes it.

The existing approaches to solving this used techniques called zero-knowledge proofs (ZKPs) — mathematical methods that let one party prove something is true to another party without revealing the underlying data. The problem: generating ZKPs for large neural networks was brutally slow, taking hours for a single forward pass. FlashProve changes that.

What FlashProve Does

Anthropic, working with researchers at EPFL, MIT, and StarkWare, built a ZKP system optimized specifically for transformer models — the architecture used in virtually every major large language model (LLM).

According to Anthropic's release:

• Speed: ~110 tokens per second on a single NVIDIA H100 GPU
• Scale: Validated on a 400-billion-parameter version of Llama 4 Maverick
• License: Apache 2.0 — free for commercial and research use
• Status: Research-stage; not yet deployed in Claude

The 110 tokens/second figure is significant. Prior systems ran orders of magnitude slower on models this large. At that speed, a typical 500-word article response would proof in under a minute on a single GPU — still slower than unverified inference, but plausible for audit workflows.

The code is live on GitHub.

How It Works (Without the Math)

A zero-knowledge proof lets you prove you performed a computation correctly without showing anyone the inputs or the model weights. For AI inference, it means:

1. The model provider runs the query through the actual model
2. FlashProve generates a cryptographic proof that the computation happened correctly
3. Anyone can verify that proof — fast — without access to the model

What it proves: That a specific, committed model (identified by a cryptographic hash of its weights) produced a specific output for a specific input.

What it does not prove: That the model is safe, accurate, or unbiased. It proves identity and consistency, not quality.

Why It Matters

For enterprises: Companies licensing AI APIs can verify they are actually getting the model they are paying for. A buyer purchasing Claude Opus API access could cryptographically confirm outputs weren't routed through a cheaper model.

For regulators: The EU AI Act and emerging US frameworks increasingly require auditability. FlashProve gives auditors a technical mechanism to verify provider compliance claims — if providers choose to generate proofs.

For security: Any system where AI-generated content needs to be authenticated — court filings, medical records, financial reports — now has a viable path to tamper-evident provenance.

For AI safety research: Knowing which model produced an output is a foundation for accountability chains. If a model causes harm, proof of inference establishes who ran what.

The Catch: It's Research, Not Production

Anthropic is explicit that FlashProve is not deployed in Claude today. The computational overhead — even at 110 tokens/second — still exceeds typical inference costs. Scaling this to billions of daily queries requires hardware and optimization work that hasn't happened yet.

There's also the question of adoption. FlashProve works only if the AI provider chooses to generate proofs. A provider determined to deceive users could simply opt out. The tool is a building block for trust infrastructure, not a mandate.

What to Watch

Watch for whether large AI API providers — OpenAI, Google, Mistral — engage with FlashProve or propose competing standards. Regulatory pressure in the EU and the growing use of AI in legal and medical contexts will push the demand for provable inference from optional to required. If Anthropic deploys this in production, it could become a differentiator and a de facto industry standard. The Apache 2.0 license was clearly designed to encourage that outcome.