What if running inference cost nothing to start and nothing to maintain?
In Brief
Every LLM API today demands an account, an API key, and pre-funded credits. For AI agents and onchain applications, this friction is a hard blocker — you cannot onboard an agent that needs to provision an account before it can reason. A model marketplace built on x402 changes this: zero signup, zero API keys, zero wallet to pre-fund. Just HTTP 402 and the model responds.
The Landscape
Seven inference access models in active use, compared across the dimensions that matter for agent-native and developer-first deployment.
x402 (proposed)
- Signup Required: No
- API Key: No
- Pre-Fund Required: No
- Per-Inference Cost: $0.001-0.01
- Models Available: 5-10 (MVP)
x402.press (production)
- Signup Required: No
- API Key: No
- Pre-Fund Required: No
- Per-Inference Cost: ~$0.001
- Models Available: 43 models
OpenRouter
- Signup Required: Yes
- API Key: Yes
- Pre-Fund Required: Yes
- Per-Inference Cost: $0.0001-0.01
- Models Available: 321 models, 69 providers
Together AI
- Signup Required: Yes
- API Key: Yes
- Pre-Fund Required: Yes
- Per-Inference Cost: $0.0002-0.01
- Models Available: 200+ models
Replicate
- Signup Required: Yes
- API Key: Yes
- Pre-Fund Required: Yes
- Per-Inference Cost: $0.0002-0.01
- Models Available: 100+ models
HuggingFace Inference
- Signup Required: Yes (free tier)
- API Key: Yes
- Pre-Fund Required: No (rate-limited)
- Per-Inference Cost: Free (throttled)
- Models Available: 200k+ models
Direct provider APIs
- Signup Required: Yes
- API Key: Yes
- Pre-Fund Required: Yes
- Per-Inference Cost: $0.00015-0.15
- Models Available: Provider-specific
The Trade-Offs That Matter
x402.press already works in production — 43 models live, serving prompts at DeepSeek subdomains. The infrastructure is proven. The remaining gap is breadth: 43 models is useful but narrow compared to OpenRouter’s 321. An MVP of 5-10 carefully chosen open models at $0.001-0.01 per prompt would cover the most common inference workloads — Llama, Mistral, DeepSeek, Phi — while keeping the cost structure simple enough that an agent running 10,000 inferences pays $10-100 with no accounts to manage.
Zero signup is the actual differentiator, not price. OpenRouter, Together, and Replicate all compete on per-token cost, and they are all within striking distance of each other. What none of them offers is the ability for an AI agent to discover a URL, hit it with a 402, and get a response without any human in the loop holding an API key. For onchain dApps and autonomous agents, this is the difference between deployable and not.
The $8B LLM inference market is dominated by pay-later billing, which is exactly wrong for the agent-native future. Every major inference provider bills monthly or by pre-funded credit. An agent that runs for 30 minutes and completes its task should not need a monthly subscription. x402 shifts inference billing from subscription economics to metered economics — the model fits the workload, not the other way around.
The Verdict: Extend x402.press, Do Not Build From Scratch
The infrastructure exists. The extension path is clear.
Layer one — the x402.press core. 43 models already serving via HTTP 402. The routing, payment, and model-serving pipeline is production-proven. Any extension starts here.
Layer two — a curated open-model tier. 5-10 models at $0.001-0.01 per prompt, selected for agent workloads (instruction-tuned, function-calling capable, small enough for low-latency inference). This is the MVP for the x402 model marketplace.
Layer three — public discovery. A simple registry where an agent can fetch available models, their prices, and their capabilities. No registration, no auth — just a GET request that returns the menu.
The addressable market is the set of developers and AI agents who will never sign up for an API account but will pay per prompt. That is a large and growing segment, and no existing provider serves it.
N.P. Vincent
x402 Review 