FlashProve: Anthropic releases a prototype to cryptographically verify AI response provenance
Image generated by the Inferama Editorial Team
01

The trust problem in model inference

Current users cannot reliably verify whether a response comes from the model they intended to query. A routing service might substitute the model with a cheaper one for cost reasons, or an attacker could manipulate the token stream. Output quality does not guarantee the identity of the model that generated it, posing risks for independent evaluators, corporate clients, and regulators.

Anthropic identifies this trust gap as an obstacle to the secure adoption of AI. The goal of provable inference is to allow a user with limited resources to verify that a response came from the requested model, without needing access to private weights or running the model locally.

The inability to confirm model identity creates a specific vulnerability in automated workflows where downstream actions depend on the source of the output. If a routing layer silently switches to a different model, the resulting text may still appear coherent and high-quality, masking the substitution from the end user. This gap forces organizations to rely on implicit trust in the infrastructure provider rather than on verifiable cryptographic evidence of the model's origin.

02

Mechanics of zkSNARK proofs in FlashProve

FlashProve is based on zkSNARKs, zero-knowledge succinct arguments of knowledge. The model provider first publishes a cryptographic commitment, a short digital fingerprint of the model's weights. Then, for each response, it generates a proof demonstrating that the committed model produced those specific tokens from the given prompt.

The proof satisfies four key properties: completeness (an honest proof is accepted), soundness (it is impossible to forge a proof for a false statement), zero-knowledge (it reveals no weights or intermediate values), and succinctness (verification is fast and cheap). The system does not hide the model's architecture, only its weights.

The cryptographic commitment acts as a fixed reference point that links the model's weights to a specific digital fingerprint. By publishing this commitment before inference, the provider establishes a baseline against which all subsequent proofs are validated. This mechanism ensures that the verifier can confirm the response originated from the exact model configuration that was initially declared, without requiring the verifier to possess the private weights themselves.

Properties of proofs in FlashProve

PropertyDescription
CompletenessThe verifier accepts the proof if the committed model generated the response.
SoundnessIt is not possible to generate a valid proof for a response that does not come from the committed model.
Zero-knowledgeThe proof reveals no model weights or which experts were activated in each layer.
SuccinctnessVerification requires seconds on a standard CPU, without the need for a GPU.
03

Reported performance and metrics

Anthropic evaluated FlashProve with Llama 4 Maverick, a 400-billion-parameter mixture-of-experts model quantized to 4 bits. On a single H100 GPU, the system generates proofs at a rate of 110 tokens per second. The proof size is approximately 19 MB per 1,024-token request.

Verifying a proof takes about 17 seconds on a 24-core CPU. The cost of generating the proof is 70 to 95 times the GPU time taken by standard inference in vLLM. These metrics are specific to the tested configuration and do not constitute a performance guarantee in other environments.

FlashProve performance metrics

MetricReported valueConditions
Generation speed110 tokens/sH100 GPU, Llama 4 Maverick 4-bit
Proof size~19 MBPer 1,024-token request
Verification time~17 seconds24-core CPU
GPU overhead70-95xCompared to vLLM inference
04

Published artifacts and formal verification

Anthropic has published five artifacts under the Apache 2.0 license: the research post, a technical report, a protocol specification, executable code (prover and verifiers), and formal proofs in Lean. The code includes a fast verifier in C++ and a formally correct verifier in Lean.

To ensure trust in the code, the team used formal verification with the Lean proof assistant. The properties of completeness, soundness, and zero-knowledge were mathematically proven. However, formal verification does not eliminate the need to audit that the executable code matches the formal specification, a task that requires human or agentic review.

The separation between the fast C++ verifier and the formally correct Lean verifier addresses different operational needs. The C++ implementation prioritizes speed for practical use cases, while the Lean code serves as a mathematical anchor that guarantees the underlying logic is sound. This dual approach allows users to choose between performance and formal assurance, depending on the criticality of the verification task and the resources available for auditing.

05

Current limitations and scope of the prototype

FlashProve is a research prototype, not a deployed product. The system is limited to mixture-of-experts models and assumes greedy decoding (selection of the most probable token). The maximum tested context is 1,024 tokens, which restricts its applicability to long conversations.

The proof is generated post-hoc, after inference, meaning it does not prevent manipulation in real time but allows detection afterwards. Additionally, the weight commitment is valid for approximately 16 million requests, after which the provider must generate a new commitment. There is no confirmed date for a commercial launch.

The restriction to greedy decoding simplifies the proof generation process by eliminating the complexity associated with sampling multiple possible tokens. This assumption limits the system's applicability to scenarios where deterministic output is required or where the most probable token is sufficient for the task. Users requiring stochastic sampling or diverse generation strategies cannot currently rely on this prototype for provable inference.

06

Keep exploring

Sources consulted

03

Corrections and transparency

If you spot incorrect or outdated information, send us a correction with the page and source we should review.

Submit a correction