LLM inference over x402: POST a prompt (JSON body), get raw model output. Flat $0.05 per call (or metered pay-actual-usage when the upto scheme is on); every response prints tokens, upstream cost, margin, and the charge. Model: Anthropic Claude Haiku 4.5 via OpenRouter — no API key, bring a wallet. AI lane declared (aiUse: model-output): model output sold as model output, no system prompt injected. We never store prompts or completions; per-call and daily spend caps enforced.