On this page (7)
Every AI coding tool advertises a monthly price. Almost none of them cost that. Cursor meters usage beyond an allowance, Claude applies weekly caps, Copilot sells credits on top of the seat price, and Windsurf sits somewhere in the middle.
If you are choosing for yourself, this barely matters. If you are choosing for a team of ten, it is the whole decision.
The four pricing shapes
| Tool | Shape | What bites |
|---|---|---|
| Cursor | Monthly fee + metered overage | Long agent sessions |
| Claude Code | Weekly caps tied to subscription | Heavy days on Pro |
| GitHub Copilot | Cheap seat + credit pool | Agent usage draws down credits |
| Windsurf | Monthly fee + tiers | Plan structure in transition |
Cursor: the overage problem
Cursor includes an allowance and meters beyond it. The allowance is comfortable for normal work and inadequate for a week of aggressive refactoring, because a single long agent session can consume a noticeable chunk of it.
What to do: set a spend cap before your first month, and check the usage dashboard in week two. Most billing surprises happen in month one, when nobody has calibrated yet.
Claude Code: the weekly cap problem
Claude Code is included with a Claude subscription, which makes it the best value in this category — with one condition: you already pay for Claude.
The constraint is weekly rather than monthly. That is fine for steady use and painful for crunch periods, because you cannot borrow from next week. The published information about what the caps actually are is thin, which is a legitimate criticism of the product.
Copilot: the credit pool problem
At $10 per user with a real free tier, Copilot is the cheapest way to put AI in front of a whole team. Agent usage, however, draws from a credit pool that is separate from completions. A team that leans on agent features will exhaust it and face a steep jump to the next tier.
Windsurf: the transition problem
Windsurf is capable and has a usable free tier, but it is mid-transition into Cognition's Devin product line. Buying into a tool whose name, packaging and pricing are all in motion is a risk that has nothing to do with code quality.
What we actually recommend
Three things that control your bill regardless of tool
- Scope the task. One sweeping instruction across a large repo burns far more than five scoped ones, and produces worse results.
- Review before you re-run. Re-running an agent on a bad plan is the most expensive habit in this category.
- Watch the first month. Every pricing surprise we have seen happened before anyone had usage data.