TAP: A Token-by-Token Payment Protocol for Streaming LLM Inference Based on Solana State Channels

TAP (Token-by-token Asymmetric Payments) is a token-by-token payment protocol for streaming LLM inference based on Solana state channels, designed using the x402 HTTP payment standard. Its core innovation is supporting consumers and producers to pause generation at any output token boundary, with on-chain settlement only for actually accepted tokens, solving the waste problem of unused tokens caused by the traditional LLM API's 'pay-first-use-later' model. This protocol is an entry for the Solana Frontier 2026 Hackathon, providing a complete technical whitepaper, Python SDK, and demo application.

Current mainstream LLM APIs adopt the 'pay-first-use-later' billing model, where users have to pay for all generated tokens (including discarded content). In agent workflows, a 5% rejection rate of responses could lead to a daily waste of $10 (including retry costs). More seriously, users cannot interrupt billing when they find the output deviates from expectations—this structural flaw gave birth to the TAP protocol.

1. Input Prepayment: When the channel is opened, the consumer prepays input fees. The producer and consumer locally verify the tokenized result, and the locked amount covers the pre-filling cost; 2. Token-by-Token Output Payment: Cumulative payment = prepayment for input + (number of output tokens × output unit price), accurately modeling the cost difference between input processing and output generation; 3. Bidirectional Pause Right: Consumers can pause due to output deviating from the topic/format, etc.; producers can pause if the consumer stops signing commitments. The maximum loss when both parties stop midway is limited to small batches (a few cents).

On-chain Components (Anchor program): State channel management (PDA-hosted USDC), commitment verification (Ed25519 signature), settlement logic (normal/dispute/close), instruction set (open_channel/settle/dispute/close). Python SDK: Modular design, including protocol (commitment schema/signature), chain (PDA/RPC interaction), x402 (wire format), consumer/producer modules, adapters (Gemini and other LLM integrations), evaluators (JSON schema/length evaluation, etc.). Demo Application: FastAPI producer service, CLI consumer, real-time terminal dashboard, Vite+React frontend.

• AI Track: Provides refined cost control for agent workflows;
• Payment Track: Built based on x402, compatible with the v1 specification;
• DePIN Track: Fair per-request settlement for decentralized inference networks;
• Consumer Applications: An effective "stop" button in chat interfaces (stop generation while stopping billing).

Future Expansion: Support audio/video streaming (billing by second/frame), GPU rental (by computation time), metered APIs (interruptible payment for any byte stream). Local Run: Build and deploy using Anchor (anchor build/deploy/test); Install Python SDK (pip install -e '.[anthropic]'); Demo environment configuration requires setting GEMINI_API_KEY, Solana key pair, RPC address, and starting the producer/consumer backend and frontend.

TAP realizes a more fair and efficient payment model for LLM inference services through state channels and token-by-token settlement. Consumers gain control to stop at any time and only pay for the content they have accepted; producers get protection for pre-filling costs. This bidirectional protection mechanism promotes the evolution of LLM API billing from 'pay-first-use-later' to 'pay-as-you-go', laying the foundation for the agent economy infrastructure.