Why does Anthropic's FlashProve mention Mantle? How do zero-knowledge proofs verify that an AI model hasn't been swapped out?

Author: Eren (@G_Gyeomm)

Compiled by: TechFlow

TechFlow Summary: Since the surge in popularity of Generative AI, an unbreakable black-box trust issue has persisted between users and model service providers: Is the content returned by terminal prompts actually from the promised advanced models, or has the cloud provider secretly swapped them for lightweight, downgraded models to cut costs?

Anthropic's newly proposed FlashProve solution uses cryptographic proofs to bind model weights. By introducing the zkSNARKS cryptographic system used by Ethereum Layer 2 networks to verify state transitions, it makes proving the authenticity of inference lightweight and efficient.

This article deeply dissects the underlying technical context behind Anthropic's references to Mantle and Starknet, analyzing how zero-knowledge proofs are crossing over to become the security and trust foundation for large models.

From Verbal Promises to Cryptographic Proofs

Is every answer just generated by AI on your screen truly from the expensive advanced large model claimed by the provider? Since ChatGPT and Claude swept across the globe, this core question has loomed over developers and professional users, never receiving a definitive answer.

To date, the basic commercial operation of mainstream AI models still rests on fragile verbal trust. This is because the complete parameter matrices and computational reasoning of all large models are enclosed within developers' own cloud data centers or third-party servers, and terminal users ultimately only receive a string of plain text or binary output. Even if an API routing service quietly replaces the originally high-cost flagship model with a low-cost small model in the background to earn massive profit margins, users have absolutely no way to distinguish it.

Figure: Anthropic FlashProve Cryptographic Inference Self-Proving Technical Architecture

Anthropic's recently published technical research proposal, FlashProve, is precisely a cryptographic antidote prescribed for this chronic trust ailment. Under the FlashProve architecture, the model provider uses cryptographic methods to make an immutable mathematical commitment to specific neural network weight parameters. Every generated inference response comes with a mathematical proof, deterministically demonstrating to the terminal that the answer was faithfully computed by the specific complete model.

The reason the crypto community resonates so strongly with Anthropic's research is that FlashProve directly adopts an extremely mature cryptographic tool from the blockchain field in its engineering implementation: zkSNARKs. According to official disclosures from Anthropic, zero-knowledge proofs can generate proof files that are extremely small in size and have minimal off-chain verification overhead without completely leaking the model's proprietary commercial weights, distributing them to terminal users with millisecond-level overhead alongside each inference response.

The Engineering Breakthrough of Zero-Knowledge Proofs on Ethereum Layer 2

More notably, Anthropic explicitly points out in its research report that the technology stack relied upon by FlashProve belongs to the exact same underlying cryptographic paradigm as the zkSNARKs used to verify network state transitions in the Ethereum Layer 2 ecosystem, directly listing Starknet and Mantle as industry representative cases. Meanwhile, the StarkWare team that developed Starknet, and the Succinct team that provides the underlying zero-knowledge proof network for Mantle, both appear prominently in the author list of this authoritative research as co-authors.

At a time when top AI institutions are stepping in to deeply explore ZK technology, it is particularly necessary to re-examine the engineering maturity of zero-knowledge proofs in the Web3 field.

In the history of cryptography, zero-knowledge proofs are a theoretical mathematical tool with deep roots spanning more than 40 years. However, in the early and middle stages of the blockchain world, it often became a marketing concept and narrative frenzy heavily packaged by certain teams in whitepapers. Its practical implementation has long been dragged down by high computational costs and slow proof generation speeds. Vitalik has repeatedly emphasized that zkSNARKs will hold a strategic status equally important to blockchain itself in the coming decade, but for a considerable period, real-world consumer-grade deployment has always seemed out of reach.

Figure: Succinct zkVM and Ethereum Layer 2 Network Verification Integration Solution

What truly freed ZK from being castles in the sky is the engineering and practical transformation of Ethereum's mainstream Layer 2 networks. Taking Mantle as an example, it completed a full architectural upgrade to OP Succinct in September 2025, officially transforming from traditional Optimistic Rollups into a genuine ZK Rollup. In this architecture, SP1, the general-purpose zero-knowledge virtual machine developed by Succinct, is responsible for generating mathematical proofs for all state transitions on the Layer 2 network and submitting them to Ethereum Layer 1 mainnet contracts for rapid verification.

The most immediate change brought by this leap is the withdrawal confirmation cycle: Optimistic Rollups must reserve a fraud-proof defense window, requiring users to passively wait up to a full 7 days for cross-chain withdrawals; whereas after fully integrating ZK proofs, Mantle shortened the finality time to approximately 1 hour, drastically compressing the terminal fund withdrawal cycle to under 6 hours.

At the same time, the computational generation cost, long regarded as ZK's fatal weakness, has been completely conquered. Leveraging the distributed hardware acceleration of the Succinct prover network, Mantle has astonishingly driven the ZK proof generation cost for a single transaction down to about $0.002, completely breaking through the commercial bottleneck for large-scale adoption.

Anthropic's Deeper Intent in Naming Layer 2 Networks

So, what industry signals does Anthropic's public naming of Layer 2 networks in its cutting-edge research actually convey? Objectively speaking, Mantle itself did not directly participate in Anthropic's internal large model training, which means this is not a direct commercial procurement partnership.

Nevertheless, it still holds crucial milestone significance for the entire Ethereum Layer 2 and decentralized compute ecosystem.

Figure: Ethereum Layer 2 Network TVL Scale and ZK Architecture Evolution Distribution

According to on-chain statistics from the professional Layer 2 data monitoring platform ETHBEAT, Mantle is currently the only major Layer 2 among the top four globally that has fully transitioned and implemented ZK Rollups. When global tech giants need to find highly available ZK infrastructure that combines low latency, extremely low single-verification costs, and has been stress-tested by hundreds of millions of dollars in real funds in large-scale distributed environments, Succinct, which empowers Mantle with core proving capabilities, naturally becomes the most logical partner.

This does not mean that a single public chain has overwhelmingly surpassed its peers in overall technology, but it irrefutably proves that the brutal stress tests endured by Ethereum Layer 2 networks over the past few years, along with the accumulated engineering breakthroughs in compute acceleration and cost optimization, are providing ready-made, battle-tested infrastructure nourishment for crossing-industry, top-tier global artificial general intelligence.

As AI large models begin to seriously explore cryptographic transparency, blockchain technology has finally shed the label of closed narratives, stepping onto the forefront of technological civilization in the form of a solid foundational technical base.