Bring your own subscription

The double-charge problem

A developer in 2026 can easily end up paying three separate AI bills: a chat subscription for thinking and writing, an editor subscription for coding, and metered model usage on top of the editor when the included pool runs dry. The three overlap heavily. All of them are ultimately renting time on the same handful of frontier models.

The overlap is worth noticing because two of the biggest coding agents are already included in subscriptions that a lot of people hold for unrelated reasons.

What Anthropic and OpenAI actually say

On Claude Code, Anthropic's documentation is direct about how you authenticate. Its surfaces — terminal, IDE extensions, desktop app, web — "require a Claude subscription or Anthropic Console account." A Console account is the pay-per-token API route. A Claude subscription is the other, equally supported route.

On Codex, OpenAI's pricing documentation is more explicit still:

"ChatGPT Work and Codex are included in your ChatGPT Free, Go, Plus, Pro, Business, Edu, or Enterprise plan."

learn.chatgpt.com/docs/pricing, checked 9 September 2026

Using an API key instead is a separate arrangement: "Pay for Codex usage based on API pricing." That is a genuinely different billing relationship, metered per token, and for interactive coding it is usually the more expensive of the two.

Subscription or API key: which should you use?

Both work. They suit different shapes of work.

A subscription suits interactive work

Interactive coding is bursty. You think for ten minutes, then fire off a dense stretch of agent work, then read. A flat monthly fee with generous limits fits that rhythm, and — more importantly — it removes the small ongoing tax of wondering what a request costs. That hesitation is a real productivity cost, and it is the main argument for a subscription over a meter.

An API key suits automation

Programmatic, high-volume or unattended work is different. If you are running agents in CI, processing a large batch of files, or building a product on top of a model, per-token billing is predictable, auditable per request, and scales without a seat limit. Subscriptions are designed around a human at a keyboard and are usually the wrong tool for a cron job.

What a subscription does not get you

Being honest about the limits matters more than the pitch.

• Limits still exist. A subscription is not unlimited. Every plan has usage limits, and heavy agent work reaches them faster than chat does.
• Parallelism multiplies consumption. Four agents working at once draw roughly four times the allowance of one. Running a team is a real cost, just denominated in allowance rather than dollars.
• Terms are set by the provider, not by any tool built on top. Availability, rate limits and model access are Anthropic's and OpenAI's to change.
• Plans differ. Higher tiers exist precisely because the lower ones run out. Match the tier to how hard you actually work the agents.

Where IDE4 fits

IDE4 is a Windows desktop workspace built on this premise. You sign in with the Codex or Claude Code account you already have, and IDE4 adds no model billing of its own — there is nothing for it to meter, because it is not reselling model access.

Practically, that means:

• No API key to create, paste or rotate. You sign in with the coding account.
• Different agents in one project can use different providers — a Claude agent and a Codex agent side by side, each drawing on its own plan.
• Because allowance is the scarce resource rather than dollars, IDE4 keeps the remaining allowance for each connected account visible while you work.
• Your projects, drafts and workspace state stay on your machine. The provider you choose processes the requests you send it.

If you do not already hold one of those subscriptions, this advantage does not apply to you — starting one purely to use a workspace on top of it just moves the cost. The argument only works when the spend is already there.

Read the full comparison against Cursor and GitHub Copilot →