Anthropic FlashProve Brings ZK Proofs to AI Inference

Key insights

• Model identity becomes verifiable. Customers might gain some evidence of which model produced an output based on cryptography.
• Privacy remains central. Providers can protect proprietary weights while allowing independent checks of their computations.
• Efficiency remains the main obstacle. High proving costs and sizable proof files currently limit practical deployment.

Anthropic FlashProve introduces cryptographic verification to AI inference, allowing users to confirm which model generated an output without exposing its weights. The research prototype, released November 2, 2026, is an open-source version that modifies a zero-knowledge proof (zkP) technology developed with blockchain systems to counter the growing worry of authenticity of AI models.

How FlashProve establishes model authenticity

Anthropic created FlashProve to fix a piece in how commercial AI services are checked. Customers normally cannot prove for themselves that the provider actually used the model it said it would. At the time the operators of those services could send user requests to cheaper models and still claim they were using the original model but they do not give clear proof of the switch.

🚨ANTHROPIC IS USING CRYPTO-STYLE ZK PROOFS FOR AI INFERENCE!

Anthropic just open-sourced FlashProve, a system that generates zkSNARK proofs to show a specific model produced a given response without revealing the weights.

The proof system is the same kind used in Ethereum L2s… pic.twitter.com/lrhP3CmMBL

— Crypto Banter (@crypto_banter) November 3, 2026

FlashProve tries to close that gap by using zero‑knowledge non‑interactive arguments of knowledge, which are usually called zkSNARKs. These cryptographic proofs let one party show that a calculation followed the rules that were set out without revealing any data.

The process starts with a set‑up. First a provider locks in a set of neural‑network weights and creates a cryptographic reference that marks which model is being used. Making that commitment takes 39 minutes and produces a digest that is roughly 6.5 megabytes, in size.

During inference, FlashProve generates a proof linking the input prompt, committed model and resulting output tokens. A verifier can then check whether the computation used the specified weights without accessing the model itself.

The design also keeps intermediate activations and private routing decisions confidential. However, it does not aim to conceal the model’s architecture.

Benchmark results reveal significant costs

Anthropic evaluated FlashProve using Meta’s quantized Llama 4 Maverick, a mixture-of-experts model with 400.7 billion parameters. The experiment used a single NVIDIA H100 GPU for proving, alongside 24 CPU cores. Verification took place on a standard CPU.

The reported results demonstrate that cryptographic verification can support large models. However, the extra computational requirements remain substantial.

Key benchmark figures include the following:

• During the test, the throughput of inferences was almost 110 tokens per second.
• The 24 requests were processed in about 223 seconds using batch processing.
• Proof generation produced files of approximately 19 MB for each 1,024-token request.
• Verification time ranged from 17 to 21 seconds on a CPU.
• Computational overhead reached roughly 70 to 95 times the GPU time required for equivalent standard inference using vLLM.

These figures show a difference. Checking a proof can take effort than doing the original computation again but making that proof comes with a big cost, for the provider.

Also, the demo only covers types of computation not every possible use of AI. It relies on integer math, greedy decoding and keeps context sizes at 1,024 tokens. That means the results don’t reflect how things work across all model types or different setup configurations.

Cryptographic research extends beyond blockchain

FlashProve brings techniques associated with blockchain scaling and privacy research into AI infrastructure. Zero-knowledge systems already support verification and privacy applications across cryptocurrency networks, including Ethereum scaling projects.

The project also draws on expertise from applied cryptography researchers. Contributors include Alessandro Chiesa, affiliated with StarkWare and EPFL, alongside researchers from MIT, Succinct and Anthropic. Chiesa co-invented Zerocash and co-founded Zcash.

Anthropic says its researchers used Claude to help build and fine-tune the system. The team also provided security statements and machine-checked proofs using Lean, a tool meant for verifying mathematical logic.

Unlike methods that depend on trusted hardware FlashProve uses math instead of asking users to trust a secure chip. This difference could be important for organizations that need proof their computations are correct without giving away model details.

The project has made its code and technical documents available under the Apache 2.0 license on its public GitHub repository. This lets other researchers look over the implementation and explore ways to make it better.

Commercial Prospects Depend on Efficiency

FlashProve offers a possible foundation for auditing AI services, particularly when customers require evidence that providers served designated models. Enterprise buyers, independent evaluators and developers could eventually use similar systems to check model identity.

However, the existing costs make it hard for many people to use this technology. Companies would have to think about computer costs and decide if the benefit of being able to check the results, on their own is worth it.Large-scale services handling millions of requests would face especially difficult efficiency requirements.

The proof files also create practical challenges. Each request can cost 19 megabytes of data. Saving and sending evidence for every interaction might become pricey. Choosing verification or using smaller proofs could open new research paths.. The current release does not provide a ready-for-production solution.

Anthropic FlashProve shows a way to achieve verifiable AI inference without revealing proprietary model weights. Still, cutting down overhead and proof sizes will decide if the technology moves past research.