Anthropic released FlashProve on November 2, 2026 — an open-source (Apache 2.0) zero-knowledge proof system for verifiable large language model inference. The system generates cryptographic proofs that a specific model produced a specific output, without revealing the model's weights. Tested on M...
"Claude produced improvements on zkSNARKs for inference, building a system substantially faster than prior open-source systems." — Anthropic, FlashProve Technical Documentation (November 2, 2026)
Anthropic released FlashProve on November 2, 2026 — an open-source (Apache 2.0) zero-knowledge proof system for verifiable large language model inference. The system generates cryptographic proofs that a specific model produced a specific output, without revealing the model's weights. Tested on Meta's Llama 4 Maverick (400B parameters, mixture-of-experts, ~17B active per token), FlashProve achieves approximately 110 tokens per second on a single NVIDIA H100 GPU, representing a 40-5,500x throughput improvement over prior ZKML systems. The release arrived days after Anthropic filed its S-1 with the SEC in October 2026, planning to list on Nasdaq under the ticker "ANT."
The system addresses a real and growing commercial problem: enterprise AI buyers cannot currently verify which model served a given response, and providers have economic incentives to swap cheaper models behind opaque APIs. Data from Vercel's AI Gateway shows 48% of development teams switched their flagship AI model in September 2026 alone. SAP operates 70+ AI models and actively substitutes for cost. Replit moved more than $100M per year in spend from Anthropic to OpenAI after a 30%+ discount. FlashProve proposes a cryptographic solution to this trust deficit. The question is whether anyone will pay for it.
At 66-95x compute overhead relative to standard inference, FlashProve transforms a $0.40 per million token inference cost into $28-38 per million tokens. The AI Verification Engine Market stood at $150M in 2026, projected to reach $1.36B by 2031 at a 55.4% CAGR, according to Knowledge Sourcing Intelligence. But FlashProve is a research prototype — limited to 1,024 tokens of context, greedy decoding only, and mixture-of-experts architectures. Production LLMs operate at 128K+ token context windows with stochastic sampling. The gap between research demonstration and deployable product remains wide.
FlashProve is a zero-knowledge succinct non-interactive argument of knowledge (zkSNARK) system that generates cryptographic proofs binding a model's weights, an input prompt, and the resulting output. If a verifier holds a commitment to a model's weights, they can confirm — without GPU hardware — that the output was produced by that specific model on that specific input.
The system uses the BLS12-381 elliptic curve, the same curve underpinning Zcash's cryptographic primitives. The research team bridges cryptography and AI: Roy Rinberg and Alex Allain from Anthropic; Alessandro Chiesa from EPFL and StarkWare, a co-founder of Zcash; Miranda Christ and Ben Mayer from MIT; Rahul Dalal from Succinct; and Giacomo Fenzi and Guy Weissenberg from EPFL. The system's correctness properties — completeness, soundness, and zero-knowledge — have been formally verified in the Lean proof assistant.
Hardware requirements are non-trivial but contained: one H100 80GB GPU, 24 CPU cores, and approximately 290GB of host memory. Proof generation for a 1,024-token request produces an 18.78 MB proof. The weight commitment is approximately 6.5 MB, takes roughly 39 minutes to create, and remains valid for approximately 16 million requests. Verification takes about 17 seconds on 24 CPU threads, requiring no GPU.
FlashProve is not deployed for Claude production inference. Anthropic explicitly designates it as a research prototype.
FlashProve's primary claim is throughput. Prior ZKML systems operated at speeds that made LLM verification impractical for any interactive use case. The following table summarizes the state of the field.
| System | Model | Throughput | |---|---|---| | zkLLM (2024) | LLaMA-2-13B | ~2.5 tok/s (A100) | | zkGPT (2025) | GPT-2 | 21.8 sec total | | zkPyTorch (2025) | Llama-3 8B | ~150 sec/token (CPU) | | VerInf | Llama 4 Maverick | ~0.02 tok/s | | DeepProve | GPT-2, Gemma 3 | 2.9-31 tok/s (distributed) | | Attestable (commercial) | Gemma 4 31B | 53-77 tok/s (H100) | | FlashProve | Llama 4 Maverick | ~110 tok/s (H100) |
Several caveats apply. Model sizes vary significantly across these benchmarks — comparing throughput on GPT-2 (124M parameters) to Llama 4 Maverick (400B parameters, ~17B active) is not apples-to-apples. Attestable, the closest commercial competitor, achieves 53-77 tokens per second on a smaller model (Gemma 4 31B) on the same H100 hardware. FlashProve's throughput advantage over Attestable narrows when normalized for model complexity, though direct parameter-adjusted comparisons are not available in the published data.
The academic survey literature places ZK proof overhead for LLMs at 1,000-10,000x compute relative to unproved inference. FlashProve reports 66-77x overhead relative to vLLM at standard loads, rising to 94x at highest throughput. This represents a one-to-two order-of-magnitude improvement over the academic baseline but remains far from overhead-free.
This is the central question. FlashProve's technical achievement is real, but the economics of verifiable inference remain punishing.
The cost arithmetic. LLM inference costs have dropped from approximately $20 per million tokens in 2022 to roughly $0.40 per million tokens in 2026, driven by hardware improvements, quantization, and competition. FlashProve's 66-95x GPU overhead implies that proving the same inference costs $28-38 per million tokens — returning the effective price to 2022 levels or above.
The global AI inference market reached $132.7B in 2026, according to Fortune Business Insights. Verifiable inference would need to capture only a thin slice of this to be commercially relevant. But the AI Verification Engine Market — the closest proxy — sits at $150M in 2026, according to Knowledge Sourcing Intelligence, with projections reaching $1.36B by 2031 (55.4% CAGR). Even the optimistic 2031 projection represents roughly 0.5% of the current inference market. Verification is, at present, a rounding error on inference spending.
Who has willingness to pay. Three potential buyer categories exist:
The subsidy question. FlashProve is Apache 2.0 open-source, produced by a team spanning Anthropic, EPFL, MIT, and Succinct. The research costs are borne by these institutions. If FlashProve or its descendants reach production, the proving costs shift to whoever operates the prover — likely the inference provider, who would pass costs to customers, or a third-party verification service. No published pricing model exists. The current economic model is pure research subsidy.
Marginal vs. average cost. One potential path to economic viability: selective verification. Rather than proving every inference call, a buyer could prove a random sample — say, 1 in 1,000 requests — reducing amortized overhead to roughly 0.07-0.1x per request. This provides probabilistic rather than deterministic guarantees but may be sufficient for audit and compliance use cases. FlashProve's architecture does not preclude this approach, though the team has not published economic modeling around sampling strategies.
The commercial motivation for verifiable inference is straightforward: API providers have economic incentives to serve cheaper models behind premium pricing tiers, and buyers have limited ability to detect this.
The evidence is circumstantial but persistent. Vercel's AI Gateway data shows 48% of development teams switched their flagship AI model in September 2026 alone, indicating intense cost-driven churn. Replit reportedly moved more than $100M per year in spend from Anthropic to OpenAI after securing a 30%+ discount. SAP operates 70+ AI models and actively swaps for cost, with its CFO fielding customer questions about model provenance, according to SAP Insider reporting.
None of this proves that providers serve misrepresented models. But the incentive structure is clear: if a provider can serve a $0.10/M-token model while billing at $0.40/M-token rates, the margin incentive is significant. FlashProve's weight commitment scheme would make such substitution cryptographically detectable. A buyer holding a commitment to Model A's weights could verify that any given response was produced by Model A, not a cheaper substitute.
The practical obstacle is that this requires the provider to publish a weight commitment in the first place — a voluntary act. A provider engaged in model swapping has no incentive to publish commitments. The verification scheme works only if providers opt in, likely under market or regulatory pressure.
Two Ethereum standards are building the on-chain infrastructure for verified AI inference.
ERC-7992 standardizes ML model registration and ZK inference proof verification on Ethereum. It defines a contract interface for registering model commitments and verifying proofs against them — effectively a decentralized registry of model identities and a verification endpoint. This is the natural on-chain counterpart to FlashProve's off-chain proving system.
ERC-7007 addresses a narrower use case: verifiable AI-generated content NFTs. It provides a standard for minting tokens that carry cryptographic proof that a specific model generated the associated content.
Both standards remain in draft or early adoption. Their relevance to FlashProve is architectural: if ZKML systems like FlashProve mature, on-chain verification registries become the settlement layer for model identity claims. The ZKML startup ecosystem — Modulus Labs (EZKL), which raised $6.3M in seed funding for SNARK-based proofs handling ResNet-class models in 30-60 seconds, and Giza, which raised $5.2M total for Cairo-based proving targeting DeFi applications — is building toward this stack.
FlashProve's limitations are substantial and clearly documented by the team.
Context window: 1,024 tokens. Production LLMs routinely operate at 128K+ token context windows. FlashProve's limit is 125x below production requirements. Extending proof systems to longer contexts is not a linear scaling problem — proving costs grow with context length, and the memory requirements (already ~290GB for 1,024 tokens) would scale accordingly.
Greedy decoding only. Production LLM systems use stochastic sampling — temperature scaling, top-p (nucleus) sampling, and other techniques that introduce controlled randomness. FlashProve supports only greedy decoding (argmax token selection), which produces deterministic outputs but is not how most production systems operate. Extending ZK proofs to cover stochastic sampling adds significant cryptographic complexity, as the randomness source itself must be verifiable.
MoE architectures only. FlashProve is tested on Llama 4 Maverick, a mixture-of-experts model. Dense models (GPT-4, Claude) would present different proving characteristics. The system's applicability to the broader model landscape is undemonstrated.
Architecture disclosure. While FlashProve hides model weights, the model architecture is exposed. A verifier learns the structural design of the model — layer counts, attention patterns, routing mechanisms — even if they cannot extract the trained parameters. For providers who consider architecture proprietary, this is a meaningful limitation.
Proof size. At 18.78 MB per 1,024-token request, proof storage and transmission costs are non-trivial at scale. A service handling 1 million requests per day would generate approximately 18.78 TB of proofs daily.
Not deployed. FlashProve is not running on any production Anthropic system. It has not been integrated with Claude. The release is a research contribution, published open-source, with no announced productization timeline.
FlashProve represents a genuine technical advance in zero-knowledge proof systems for LLM inference. The throughput numbers are real and substantially ahead of the prior state of the art. The formal verification in Lean is rigorous. The team is credentialed. The open-source release under Apache 2.0 is a legitimate public good contribution.
None of this changes the fundamental economics. At 66-95x overhead, verifiable inference is a luxury good in a market defined by relentless cost compression. The model-swapping problem is real, but the solution requires voluntary provider opt-in, and the cost premium dwarfs the economic value of the assurance for most buyers. The path to commercial viability likely runs through selective sampling (proving 1-in-N requests), regulatory mandates, or decentralized protocols where trustlessness is a design requirement rather than an optional feature.
Anthropic's timing — publishing days after its S-1 filing — suggests the primary audience for FlashProve is not today's inference buyers but tomorrow's public market investors. The message is: Anthropic does fundamental research that others cannot, and it gives it away. That is a defensible narrative for a company seeking a premium valuation in an increasingly commoditized inference market. Whether FlashProve's descendants will generate revenue is a question for 2028 or beyond. For now, it is a research subsidy producing a public good, with the costs borne by Anthropic and its academic collaborators, and the economic returns deferred indefinitely.