A new VS Code extension called LLM OneStop Code is pitching itself as the pay-as-you-go alternative to Cursor, GitHub Copilot, and Claude Code. Instead of a monthly subscription, it bills in credits at the underlying model cost plus a 5% markup. Use it for two hours one month and forty the next - you only pay for what you actually consume.
The pitch addresses a real friction point. Cursor's Pro plan runs $20/month whether you touch it daily or once. GitHub Copilot starts at $10/month. Claude Code ties usage to your Anthropic API tier with monthly caps that can feel arbitrary. For developers who code in bursts - freelancers between projects, hobbyists, or anyone who doesn't write code eight hours a day - subscriptions are a lousy deal. You're paying for idle capacity.
That said, the tool is brand new and unproven. The AI coding agent space is brutal right now, with Cursor, Copilot, Continue, Cody, and Aider all competing aggressively on features, model access, and IDE integration. Pricing alone rarely wins in developer tools. The question is whether LLM OneStop Code can match the autocomplete quality, codebase awareness, and multi-file editing that established players have spent years refining. A 5% markup on raw API costs sounds cheap until you factor in the context window usage that good code agents require - long conversations with large codebases burn through tokens fast, and the per-session cost can surprise you.
The pay-per-use model itself is worth watching, though. If subscription fatigue in AI tools keeps growing (and it will - most people now pay for three or four AI subscriptions they don't fully use), more tools will move toward consumption-based pricing. LLM OneStop Code is an early experiment in that direction.