In an AI coding subscription comparison, the question that decides most choices isn’t the price but which tools the plan works in. A ChatGPT plan runs Codex and also signs in to OpenCode and other apps. A Claude plan runs Claude Code and nothing else that talks to Claude directly. A Cursor plan works only in Cursor. The GLM, Kimi, MiniMax and OpenCode Go plans sell open models that work in Claude Code, Codex, OpenCode and more.
Pick the agent first, then the plan that works in it.
The comparison
| Plan family | Works in | How the limits work | API credits included |
|---|---|---|---|
| ChatGPT (Free, Go, Plus, Pro, Business, Edu, Enterprise) | Codex (CLI, IDE extension, desktop app; web and cloud on paid plans), OpenCode, Contenox | A short recurring window on Plus and Business, a weekly limit on top; shared with ChatGPT Work. Plus and Pro can buy credits | None |
| Claude (Pro, Max, Ultra, Team, Enterprise) | Claude Code, also when another app hosts the unmodified Claude Code | A 5-hour session window plus a weekly limit. Ultra drops the 5-hour window for verified accounts. Extra usage can be bought. Enterprise bills usage at API rates | Monthly on Max, Ultra and Team, for the API only |
| GitHub Copilot (Free, Student, Pro, Pro+, Max, Business, Enterprise) | VS Code, Copilot CLI, the cloud agent, GitHub’s Claude and Codex agents (Pro and up, preview), the Codex VS Code extension (Pro+ and Max), OpenCode, Contenox | Monthly AI Credits, counted by the tokens a model processes. More within a budget you set, or wait for the monthly reset | None |
| Cursor (Free, Plus, Premium, Super, Ultra, Teams, Enterprise) | The Cursor editor, the Cursor CLI, cloud agents (Premium and up) | One weekly pool for Cursor’s own models (Grok and Composer), shared with the Grok apps. Claude, GPT and Gemini are billed at API list price from prepaid credit, Premium and up | None |
| Google AI (Plus, Pro, Ultra) | Antigravity (app and CLI), AI Studio. Not Gemini CLI since 18 June 2026 | Pro and Ultra: a quota that refreshes every five hours up to a weekly limit; other plans: a weekly quota. Pro and Ultra can buy AI credits for overage | Google Cloud credits on Pro and Ultra, through the Google Developer Program premium benefit |
| Z.ai GLM Coding Plan (Lite, Pro, Max, Team) | Claude Code, Codex, OpenCode, Cline, ZCode, OpenClaw and others on Z.ai’s list, and only those | 5-hour and weekly credit limits. Tokens are converted to credits per model; off-peak use counts half | None |
| Kimi Code (Moonshot’s membership tiers) | Kimi Code CLI, desktop app and VS Code extension; Claude Code, OpenCode, Codex and Hermes Agent with a base URL and key | A membership quota; bigger context models use it faster | None |
| MiniMax M Plan (Go, Explore, Build) | MiniMax Code; Claude Code, Codex, Cursor, OpenCode, Pi and Hermes Agent with an M Plan key | A quota per tier, used faster on long context and hard tasks | None |
| OpenCode Go (Go, Go Plus, Go Teams) | OpenCode; also Claude Code, Codex, Hermes, Pi and Kilo Code CLI, which OpenCode lists as validated | A monthly allowance per model, with 5-hour and weekly caps. Extra Usage can draw on your console balance after that | None |
DeepSeek has no subscription. Its API is pay per token, from a balance you top up, with half price outside peak hours. See the DeepSeek pricing page. Its models are also part of OpenCode Go.
Each plan’s name in the table links to its official plans page. Prices and exact limits change often, so read them there.
Can I use my Claude subscription in other tools?
Only through Claude Code. Anthropic’s terms allow you to sign in to the unmodified Claude Code with your own plan, also when another platform hosts it. They forbid other developers’ products from routing requests through Free, Pro, Max or Ultra credentials, and Anthropic says it may enforce this without notice.
The plugin trap: plugins that let OpenCode or another agent sign in with your Claude plan put the account at risk. If you want Claude in OpenCode, use an API key. On Max, Ultra and Team, the monthly API credits pay for that key; they don’t cover Claude Code itself and don’t roll over. Setup is on the Anthropic page.
Can I use my ChatGPT or Copilot subscription in other tools?
Yes, more freely than Claude. OpenAI offers “Sign in with ChatGPT”, which Codex, OpenCode and Contenox use. GitHub lets OpenCode sign in with a Copilot plan, and its own docs list the Codex VS Code extension’s Sign in with Copilot for Pro+ and Max. Contenox offers the Copilot sign-in only after you enable unofficial sign-ins in its settings, because GitHub has not approved it. Steps for each agent: ChatGPT subscription, GitHub Copilot subscription.
On Copilot Business and Enterprise, your organization’s policies decide which agents and models are allowed.
Cursor Ultra or Claude Code Max?
They pay for different models. Cursor’s included pool covers Cursor’s own models, Grok and Composer. Claude models in Cursor are billed per token at API list price from prepaid credit, on top of the plan. Claude Max covers Claude models in Claude Code, within its limits.
So if you want Claude as your daily model, Max is the plan that includes it. If you want Cursor’s editor and are happy with its own models, Cursor’s plan is the one that covers them. Both read best on the official pages: Cursor pricing, Claude pricing.
Does Google AI Pro still work in Gemini CLI?
No. Since 18 June 2026, Gemini CLI and the Gemini Code Assist IDE extensions no longer serve requests for Google AI Pro, Ultra and Code Assist for individuals. Google’s FAQ points those users to Antigravity and the Antigravity CLI, where the plan raises your limits.
Guides that tell you to “sign in with Google” in Gemini CLI to use your subscription are out of date, so following one leaves you locked out. Gemini CLI still works with a Gemini API key, billed per token.
Which is the cheapest AI coding subscription?
For the same amount of agent work, the plans that sell open models (GLM, Kimi, MiniMax, OpenCode Go) usually stretch furthest, because the models behind them cost far less per token than the top closed ones. Our own measurements show the gap: the same kind of session cost about $80 on Claude Fable 5.1 and under $1 on DeepSeek V4 Flash, measured before DeepSeek replaced it with V4.1 Flash in the same price class.
Cheapest isn’t the same as best value. A weaker model that needs three tries costs you more time than it saves. How to choose a model for coding has a method to test that on your own tasks.
Z.ai allows its plan only in the tools it lists, and Kimi treats changing a tool’s client identifier as a violation that can suspend the membership, so wiring the plan into an unlisted tool can cost you the plan. Check the plan’s own tool list before you wire it into a tool that isn’t on it.
Subscription or pay per token: when does a flat plan win?
It depends almost entirely on the model. Our numbers, from the Contenox team’s own use in 2026, with prompt caching on:
| Model, pay per token | What it cost us |
|---|---|
| Claude Fable 5.1 | About $80 per heavy coding session |
| Gemini on Vertex AI | $8 to $20 per developer per working day |
| DeepSeek V4 Flash (now V4.1 Flash) | Under $1 per coding session |
A heavy session here is eight prompts that sent about 300 million input tokens, because every step of an agent resends the whole conversation. AI coding costs explains where those tokens come from.
Read against a plan’s monthly fee, three cases follow:
- Top closed models, daily use: a flat plan wins clearly. At around $80 a session, a few heavy sessions at API rates cost more than most monthly plans. Open the plan pages and compare.
- Mid-priced models: our Vertex numbers come to roughly $160 to $400 per developer for a month of 20 working days. Compare that with the plan for the same model family.
- Cheap open models: pay per token is often cheaper than any plan. Under $1 a session takes a lot of sessions to reach a monthly fee. A plan then buys convenience and a fixed bill, not savings.
A flat fee is only flat until the window closes. A plan that runs out on Tuesday leaves you waiting, buying credits or switching to an API key for the rest of the week. Many people keep both: a plan for daily work and a key for when it runs out.
To decide on your own numbers, run three typical tasks on an API key, note the tokens, and multiply by a month. The steps are in AI coding costs.
Contenox
Contenox signs in with ChatGPT with a one-time code, and with Claude by running Anthropic’s own Claude Code on its Agents page. GitHub Copilot works too, as an unofficial sign-in you switch on in Settings, since GitHub has not approved it. They sit next to API keys for OpenAI, Anthropic, Gemini, DeepSeek and others, so you can work on the plan and switch a session to a key when the window closes. The chat’s permission rules apply to whichever one runs.
Related guides
- Codex API key vs subscription: both ways to pay for Codex, and how to switch
- AI coding costs: why agent sessions use so many tokens
- Which AI provider for coding agents?: providers per agent, including API keys and clouds
- Best AI coding agents: pick the agent before the plan
- Prompt caching: why pay-per-token bills aren’t as high as the token counts suggest