Codex Pricing: Included Usage by Plan, Credits, and the API Route (2026)

Three ways to pay for Codex

Codex — OpenAI's coding agent, available as a CLI, in the IDE, on the web and inside ChatGPT — does not have its own price. You pay for it in one of three ways:

1. Included in a ChatGPT plan. Per OpenAI's Codex documentation, Codex is included in Free, Go, Plus, Pro, Business, Edu and Enterprise. Each plan has an allowance that refills on a rolling five-hour window.
2. Credits on top of a plan. When the allowance runs out, credits extend it. Credits are a unit that converts token usage into a simpler number, and their rate depends on which model Codex is running.
3. The API. Use Codex with an API key and pay per token at standard API rates, with no plan allowance involved.

The right one depends almost entirely on how much you use it, so the numbers below are the decision.

What each plan includes

OpenAI publishes the allowance as an estimated number of local Codex messages per five-hour window, which varies by model because the models cost different amounts to run. The figures OpenAI lists, from cheapest model to most capable:

• Plus: roughly 250–2,000 messages on the light model, 25–200 on the mid model, 10–100 on the flagship.
• Pro (5× tier): 1,250–10,000 / 125–1,000 / 50–500.
• Pro (20× tier): 5,000–40,000 / 500–4,000 / 200–2,000.
• Business: the same allowance as Plus, per seat.
• Enterprise and Edu: no fixed rate limit; usage scales with the workspace's credits.
• Free and Go: described as basic exploration and lightweight tasks respectively, with no published range.

The ranges are wide because a "message" can be a one-line edit or a multi-file refactor; the low end is heavy agentic work, the high end is small tasks. Two things to take from the table: the 5× Pro tier is where a full-time developer stops hitting the wall, and Business gives a team no more per person than Plus does.

How credits are priced

OpenAI's documentation gives the credit cost per million tokens for each model Codex can use — the spread is what makes model choice a cost decision:

• Flagship reasoning model: 100 credits per million input tokens, 500 per million output.
• Mid-tier model: 50 input, 300 output.
• Light model: 5 input, 30 output.

That is a twenty-fold difference between the light and flagship models on input, and roughly seventeen-fold on output. Most everyday Codex work — tests, small fixes, scaffolding — runs fine on the light model, and pointing Codex at the flagship by default is the single most common way to burn through an allowance. Credits can be bought on Plus and Pro to extend usage; Business and Enterprise buy them at the workspace level.

When the API beats a subscription

With an API key, Codex bills per token at the model's API rate and never touches a plan allowance. That is the better route in three situations:

• Automation. Codex running in CI, on a schedule, or inside another tool has no "five-hour window" — it needs metered billing.
• Very bursty use. A week of heavy refactoring followed by a month of nothing costs less by the token than a Pro subscription you keep paying for.
• Many light users. A team where ten people each run Codex a few times a week can spend less on a shared API key than on ten seats.

The subscription wins for the opposite profile: one person, steady daily use, who values not thinking about the meter. The API vs subscription guide walks through the break-even arithmetic for AI plans generally; for Codex specifically, the rule of thumb is that if you regularly exhaust the Plus allowance, price the 5× Pro tier against your actual token spend before buying credits every month.

Frequently Asked Questions