Anthropic Gives Buyers Proof A Claude Reply Was Not Swapped
Anthropic, the company behind Claude, launched FlashProve, a tool that gives customers a digital receipt to prove an AI answer came from the model they paid for.
Key Takeaways
- Anthropic unveiled FlashProve, which can generate cryptographic proofs at about 110 tokens per second on a single H100
- FlashProve generated proofs for a 1,024 token request with a proof size of about 19 MB
- FlashProve uses about 70x to 95x the GPU time of running the same model on vLLM
- FlashProve emits the answer and a compact proof tied to the committed model that a verifier can check faster than rerunning the full model
Georgios Konstantopoulos posted about the work. Anthropic builds the Claude model family, and its own research laid out the numbers and system design.
Anthropic said users still cannot easily check whether an answer Came From the Model named, a cheaper substitute, or a tampered token stream, and its provable inference system fixes that by having the model provider commit to a specific model and then produce a cryptographic proof that the response Came From the Model with those committed weights.
The research focuses on zkSNARKs, a proof system that lets one party prove a computation was done correctly without exposing all the underlying data. In this case, provable inference confirms that text was generated and that a specific model, with specific weights, produced it.
Anthropic said FlashProve generated proofs for a 1,024 token request with a proof size of about 19 MB, and for Llama 4 Maverick, a 400 billion parameter mixture‑of‑experts model, the company said a single H100 could generate those proofs at roughly 110 tokens per second.
Also Read: Anthropic Pays Professors To Test Claude In Path To Jobs
The Speed Gain Still Comes With A Cost
Those numbers still carry a steep penalty.
FlashProve uses about 70x to 95x the GPU time of running the same model on vLLM, the open‑source inference engine many developers use in production, which means provable inference is not near default for all chatbot traffic but is an option for high‑trust workloads, enterprise contracts, regulated systems and security‑sensitive AI services where model identity matters as much as speed.
Anthropic framed the work as part of a broader compliance and safety toolkit, arguing that if AI labs face rules about which systems can serve which users, proof of inference gives auditors more than a dashboard promise and provides a cryptographic receipt that the output Came From the Model the provider committed to.
From Trust Me APIs To Verifiable Outputs
AI distribution has grown more layered, enterprises often access models through gateways, brokers and orchestration layers instead of a direct lab endpoint, making substitution easier for cost control, latency or attackers, and Anthropic has spent this cycle adding enterprise and security features around Claude.
Tuesday’s research moves that effort down a layer, from who can call a model to whether the output can be proven after the fact. There is also a crypto angle. zkSNARKs emerged from the same family powering privacy‑preserving blockchain systems, and Anthropic applies that machinery to AI inference rather than on‑chain transfers.
How FlashProve Changes The AI Audit Trail
FlashProve sits between normal inference and formal verification.
A customer sends a prompt, the model runs, and the system emits both the answer and a compact proof tied to the committed model, and a verifier can then check the proof far faster than rerunning the full model, confirming the response Came From the Model without every customer hosting a giant model.
For cloud buyers, provable inference matters when Anthropic is not the only party in the request path. A bank using Claude through a broker or a government team relying on an internal gateway may care less about style than whether the named model actually handled the request, and FlashProve gives those buyers a technical way to test that claim, though the cost remains the key limit to wider adoption.
Read Next: Anthropic Lands $23.5B Cash Haul Before IPO
