How usage is counted
At $20 a month most plans look the same, but tools measure usage in three different ways, and that decides how far your money goes.
1. A flat plan with limits that aren’t published
Claude Code, Cursor and Devin sell tiers described as multiples of a base plan (“5x”, “15x”, “20x”) without saying how many prompts or tokens that is.
- Claude Code has rolling 5-hour sessions plus weekly limits. They’re shared with Claude on the web, desktop and mobile, and Anthropic says there’s no fixed message count.
- Cursor has one usage pool for its own models, shared with Grok and refreshed weekly. Its size isn’t published. Claude, GPT and Gemini aren’t in the pool: on Premium and above they’re billed at API list price from prepaid credit.
- Devin, which now includes the Devin Desktop editor (formerly Windsurf), describes its quotas only as “light” on Free, “included” on Pro and “significantly higher” on Max.
Codex sits in between. OpenAI publishes 5-hour estimates per model for Plus and Business Standard, for example 15 to 160 local messages on GPT-6.1 Sol. Pro has no 5-hour limit at present, though weekly limits may apply.
2. Credits with published numbers
- Kiro includes 50 credits on Free, 1,000 on Pro ($20), 2,000 on Pro+ ($40), 5,000 on Pro Max ($100) and 10,000 on Power ($200). Teams can instead buy pooled seats at $20 each, which add 500 credits per seat to a shared pool. Different models use credits at different rates.
- GitHub Copilot uses AI credits worth $0.01 each. Pro includes 1,000 base credits plus a “flex” allotment, 1,500 in total today; Pro+ and Max include 7,000 and 20,000. GitHub says the flex part may change. Code completions and next edit suggestions don’t use credits and are unlimited on paid plans.
- Zed Pro includes $5 of token credit a month for Zed-hosted models.
3. A free tool, and you pay the model provider
Aider, Cline and OpenCode are open source and free, and Zed’s Personal plan costs $0. You pay your model provider per token, or nothing if you run a local model. Cline and OpenCode also sell optional subscriptions for open-weight models: ClinePass ($9.99), OpenCode Go ($10) and Go Plus ($40).
Gemini CLI is free and open source too, but since June 18, 2026 it no longer works on Google’s free tier or Google AI Pro and Ultra. It needs a paid Gemini API key or a Gemini Code Assist Standard or Enterprise license.
What happens when you hit the limit
| Tool | When the included usage runs out |
|---|---|
| Claude Code | Wait for the 5-hour or weekly reset, upgrade, buy usage credits at API rates (with an optional monthly cap), or switch to an API key |
| OpenAI Codex | Wait for the reset, or buy credits on Plus and higher. Credits last 12 months |
| Cursor | Wait for the weekly refresh, pay from prepaid credit at per-model rates, or upgrade |
| GitHub Copilot | Wait for the monthly reset, upgrade, or set a budget for additional usage at $0.01 per credit. GitHub says additional usage may be capped |
| Kiro | Turn on overage at $0.04 per credit (off by default), or upgrade |
| Devin and Devin Desktop | Extra usage on Pro is billed at API pricing |
| Zed Pro | Usage is billed at the provider’s list price plus 10%, up to a spend limit you set (default $10) |
| Gemini CLI | Code Assist licenses have daily caps: 1,500 model requests per user on Standard and 2,000 on Enterprise |
Paying per token instead of a plan
Every tool that takes your own API key lets you skip a plan’s limits and pay per token at the provider’s rates. Whether that’s cheaper depends on how much you use it. Anthropic’s Claude Code docs put the average at about $13 per developer per active day, or $150 to $250 a month, across enterprise deployments. That’s more than Claude Pro and about the same as Claude Ultra. For daily work a plan usually costs less; an API key suits scripts, CI jobs and occasional bursts.
Where your own key works:
- Claude Code: an Anthropic API key or cloud credentials (Amazon Bedrock, Google Cloud, Microsoft Foundry), Claude models only.
- Codex CLI: an OpenAI API key, without the cloud features. It can also run local models through Ollama or LM Studio.
- Cursor: Premium and above only, not Free or Plus. Requests still go through Cursor’s servers.
- GitHub Copilot: in VS Code, JetBrains, Xcode, the Copilot CLI and app, including local models. Business and Enterprise admins can turn it off.
- Cline, OpenCode, Aider, Zed and Gemini CLI: built around your own key. All but Gemini CLI also run local models.
- Claude Code and Cursor don’t officially support local models.
- Kiro and Devin don’t document API-key support. Devin can use an eligible ChatGPT Plus or Pro plan for OpenAI models through “Sign in with ChatGPT”.
Costs that are easy to miss
- Cursor Plus ($8) has no Claude, GPT or Gemini. You need Premium ($30) for those, and they’re billed at API list price from prepaid credit on top of the plan.
- Cursor Teams adds a fee on third-party models: a Cursor Token Rate of $0.25 per million tokens, which also applies to your own API keys on a team.
- Claude Max and Ultra API credits don’t pay for Claude Code. The $100 (Max) or $200 (Ultra) a month covers the Claude API and Agent SDK, not interactive Claude Code sessions.
- An
ANTHROPIC_API_KEYin your environment changes who bills you. If it’s set and you approve it, Claude Code charges that key at API rates instead of using your plan. - Codex fast mode costs 2.5 times the normal included usage.
- Old plans are still around. Existing Cursor Pro, Pro+ and legacy Windsurf subscribers keep their plans, which is why you’ll still see those prices quoted.