GitHub Copilot
Copilot is moving to usage-based billing. The seat tiers, from github.com/features/copilot/plans:
- Free
- Pro: 10ドル per month, with 15ドル in monthly AI credits
- Pro+: 39ドル per month, with 70ドル in monthly AI credits
- Max: 100ドル per month
Chat, agent mode, code review, the cloud agent, and the CLI all draw down GitHub AI Credits. When the included credits run out, you pay for more. GitHub even paused new Pro+ and Max sign-ups while it sorts out the billing change. The seat is cheap. The credits are the meter.
Cursor
Cursor keeps a simpler face on it, from cursor.com/pricing:
- Hobby: Free
- Pro: 20ドル per month
- Teams: 40ドル per user per month
- Enterprise: Custom
Every plan includes a set amount of model usage. Cursor calls the overflow on-demand usage, billed in arrears after you burn through what is included. Same shape as Copilot: a seat price plus a usage tail.
Claude Code
Claude Code rides on the Claude subscription, from anthropic.com/claude-code:
- Pro: 17ドル per month on the annual plan (200ドル up front), or 20ドル billed monthly. Good for short sprints in small codebases.
- Max 5x: 100ドル per month
- Max 20x: 200ドル per month
You can also run it on a Team or Enterprise seat, or against an Anthropic Console account where you pay API rates per token. Usage limits apply on the subscription plans, so heavy days can still hit a wall.
Side by side
AI coding assistant pricing 2026: monthly seat floors by tool, each sitting on a usage tail
| Tool |
Free |
Entry |
Power user |
Team |
| GitHub Copilot |
Yes |
Pro 10ドル/mo |
Max 100ドル/mo |
usage-based |
| Cursor |
Hobby |
Pro 20ドル/mo |
on-demand |
40ドル/user/mo |
| Claude Code |
No |
Pro 20ドル/mo |
Max 5x 100ドル/mo |
seat or Console |
Every cell above is a starting point. The real number rides on usage underneath it.
The pattern: seat price is a floor
Line them up and the story is the same across all three.
- Entry tier: Copilot 10,ドル Cursor 20,ドル Claude Code 17ドル on annual or 20ドル monthly
- Power tier: Copilot 100,ドル Cursor on-demand, Claude Code 100ドル on Max 5x or 200ドル on Max 20x
Every plan now has a usage component underneath the flat fee. Credits, on-demand billing, token rates, usage caps. The flat monthly number is what you pay to walk in the door. What you actually owe depends on how hard your agents run.
That is the part teams miss. You budget for ten seats at 20ドル and forget that each seat has a usage tail that scales with how much the tool is used. A few heavy users on agent mode can move the bill more than adding five seats would.
What to do about it
Pick the tool on fit, not on the headline price, because the headline price is not the bill. Then watch usage like it is real money, because it is.
Two concrete moves:
- Track usage per developer, not just total seats. The spread between a light user and a heavy agent user is wide. You want to see it before the invoice does.
- Put a hard cap on anything that calls a model in a loop. Agent mode, cloud agents, and CLI runs are where a normal month turns into a surprise.
That second point is why I built AgentGuard. It is a runtime budget, token, and rate limiter for AI agents. You set a ceiling, and the agent stops at the ceiling instead of running your bill into a number you did not approve. The vendors are happy to meter you. AgentGuard is how you meter yourself first.
Prices change, and all three vendors say so plainly on their pages. The structure will not. Usage-based billing is the model now. Budget for the floor, plan for the tail, and cap the loops.
Originally published on bmdpat.com. I run a one-person AI agent company and write about what actually works.
Want these in your inbox? Subscribe to the newsletter - no spam, unsubscribe anytime.