Selective Model Routing Enables Subscriptions

The key advantage is cost smoothing. Cursor and Copilot do not send every keystroke, edit, and code action to the most expensive model. They reserve frontier models for hard reasoning and bigger generation jobs, then use cheaper in house or lighter models for fast tasks like autocomplete, edit prediction, and applying diffs. That lets them sell a predictable seat price, while Bolt.new passes model usage through more directly with token based pricing.

• In practice, many coding actions are tiny. Completing the next few lines, suggesting a rename, or applying a requested patch needs low latency more than deep reasoning. Cursor explicitly uses Anthropic and OpenAI for core generation, plus proprietary models for edit prediction and code changes, which lowers average cost per user action.
  2 sacra
• Copilot and Cursor both present pricing as a subscription. GitHub offers Copilot plans with fixed monthly seat pricing, and Cursor lists fixed monthly plans with bundled model access. That only works if the product can keep average inference cost below the subscription price through routing, limits, and internal optimization.
  4 github 5 cursor
• This creates a product split in AI coding. Tools like Bolt.new and Lovable are closer to metered compute products because they generate larger chunks of app code per request. IDE tools like Cursor sit inside a developer workflow with many small actions, which makes selective model routing much more valuable economically.
  1 sacra 3 sacra

Going forward, more AI coding products will converge on this routing layer as the core business lever. The winners will not just have the best model, they will be the ones that know exactly when to spend for deep reasoning and when to use a cheaper model fast enough that the user never notices the handoff.