Author: Eren (@G_Gyeomm)

Translated by: Deep Tide TechFlow

Deep Tide Introduction: Since the explosive rise of generative AI, there has always been an insurmountable black-box trust problem between users and model service providers: does the content returned by the terminal token truly come from the promised high-level model, or has it been secretly replaced by a lightweight downgraded model by the cloud service provider to cut costs? Anthropic’s latest proposed FlashProve solution uses cryptographic proof to bind model weights, introducing the zkSNARKs cryptographic system to verify state transitions on Ethereum's Layer 2 network, making authenticity of inference self-evident, lightweight, and efficient. This article delves deeply into the technical context of Anthropic's references to Mantle and Starknet, analyzing how zero-knowledge proofs cross over to become the security trust base for large models.

From Verbal Commitment to Cryptographic Proof

Does every answer generated by AI on your screen really come from that expensive high-level model claimed by the service provider? Since ChatGPT and Claude swept the globe, this core question has lingered over developers and professional users, never truly answered.

As of today, the base commercial operation of mainstream AI models still relies on a fragile foundation of verbal trust. The reason is that all the complete parameter matrices and computational inferences of large models are closed off in the developer's own cloud data center or third-party servers, and the end-users can ultimately only receive a string of plaintext or binary output. Even if API aggregation routing services silently replace originally high-cost flagship models with low-cost small models in the background for profit, users have no way to discern this.

Image: Anthropic FlashProve cryptographic inference self-proving technology architecture

The latest technical research solution from Anthropic, FlashProve, is a cryptographic remedy aimed at this trust ailment. Under the FlashProve architecture, model providers use cryptographic methods to make immutable mathematical commitments to specific neural network weight parameters, and each generated inference response comes with a mathematical proof, providing deterministic evidence to the end-user that the answer indeed comes from the faithful computation of a specific complete model.

The reason why the cryptographic community resonates strongly with this research from Anthropic is that FlashProve directly employs highly mature cryptographic tools from the blockchain field: zkSNARKs. According to official disclosure from Anthropic, zero-knowledge proofs can generate minuscule proof files with very low off-chain verification costs, without disclosing proprietary commercial weights of the models, and are distributed to end users with millisecond-level overhead accompanying each inference response.

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

More notably, Anthropic explicitly states in its research report that the technology stack relied upon by FlashProve and the zkSNARKs used for verifying network state transitions in the Ethereum Layer 2 ecosystem belong to the same underlying cryptographic paradigm, and directly lists 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 are prominently mentioned as co-authors in this authoritative research.

As top AI institutions delve deeply into exploring ZK technology, it is particularly necessary to reassess the engineering maturity of zero-knowledge proofs in the Web3 arena.

In the history of cryptography, zero-knowledge proofs are a theoretical mathematical tool with over 40 years of deep accumulation. However, in the early to mid-stages of the blockchain world, they often became marketing concepts and narrative excitement packaged lavishly by some teams in white papers, hampered by high computational costs and slow proof generation speeds in practical implementation for a long time. Vitalik has repeatedly emphasized that zkSNARKs will hold a strategically important position equivalent to blockchain itself in the next decade, yet for quite a long time, consumer-grade actual deployment seemed far out of reach.

Image: Succinct zkVM and Ethereum Layer 2 network verification integration solution

What truly lifted ZK out of a state of emptiness is the engineering practical transformation of mainstream Layer 2 networks on Ethereum. Taking Mantle as an example, it completed a full architecture upgrade to OP Succinct in September 2025, transforming from traditional optimistic rollups to a true ZK Rollup. In this architecture, the general zero-knowledge virtual machine SP1 developed by Succinct is responsible for generating mathematical proofs for all state transitions in the layer 2 network and submitting them to Ethereum Layer 1 mainnet contracts for rapid verification.

The most immediate change brought about by this leap is the withdrawal confirmation cycle: optimistic rollups, due to the need to leave a window for fraud proof defense, require users to passively wait for as long as 7 days for cross-chain withdrawals; whereas with the full introduction of ZK proofs, Mantle has shortened this state finalization time to about 1 hour, significantly compressing the terminal fund withdrawal cycle to within 6 hours.

At the same time, the long-regarded fatal weakness of ZK in terms of computational cost generation has been thoroughly conquered. With the distributed hardware acceleration of the Succinct prover network, the cost of generating a ZK proof for a single transaction on Mantle has astonishingly been reduced to about $0.002, completely solving the bottleneck for large-scale adoption.

Anthropic's Deep Intention Naming Layer 2 Networks

So, what industry signal does Anthropic send by publicly naming Layer 2 networks in its cutting-edge research? Objectively speaking, Mantle itself did not directly participate in the model training of Anthropic's large models; this is not a direct commercial procurement collaboration.

However, it still releases a crucial milestone significance for the entire Ethereum Layer 2 and decentralized computing ecosystem.

Image: Ethereum Layer 2 network locked value 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 mainstream Layer 2 in the top four that has fully switched to ZK Rollup. When global tech giants need to find a highly available ZK infrastructure that combines low latency, extremely low single-validation costs, and has been validated with billions of dollars in real funds in a large-scale distributed environment, Succinct, which empowers Mantle’s core proof capabilities, naturally becomes the most logical cooperation partner.

This does not mean that a single public chain has technically crushed its peers overall, but it undeniably proves: the Ethereum Layer 2 networks have undergone brutal stress testing over the past few years, and the engineering breakthroughs accumulated in computational acceleration and cost optimization are providing essential nutrients for global leading general artificial intelligence across domains.

As AI large models begin to seriously explore cryptographic transparency, blockchain technology has finally torn off the label of closed narratives, entering the forefront of technological civilization in the form of solid foundational technology.