Codex On-Demand Credits: Pricing, Draining, and When to Skip Them

Codex on-demand credits are OpenAI’s metered escape hatch: when a Plus or Pro account hits a usage limit, you can buy credits inside ChatGPT and keep working instead of waiting for the window to reset. They drain per usage, at rates published on developers.openai.com/codex/pricing, and they suit rare overflow far better than routine use. Credits are a relief valve, not a pricing plan: open-ended metering bolted onto a flat subscription. Where they sit in the wider limit system is mapped in Codex usage limits, explained.

How credits work

The flow is deliberately simple. You hit a limit, ChatGPT offers the option to continue with credits, you buy a chunk, and Codex resumes immediately. Consumption is metered against the credit balance until your normal window recovers or the balance runs out. There is no subscription change, no proration, no commitment.

Three properties define the experience:

• Manual. Buying credits is an in-product action by a human. Nothing about it is automatable, which matters the moment Codex is part of a pipeline rather than a pairing session.
• Metered. Credits bill by work performed, the same way limits meter: reasoning time, tool calls, and retries all count. The flat-plan economics you bought the subscription for are suspended while you are on credits.
• Open-ended. A second account costs $20 or $100 a month, full stop. A credit habit costs whatever your agents decide to spend.

The draining problem

The loudest complaint in community threads is speed: people buy credits expecting an evening of runway and report the balance gone within the hour. A second pattern shows up too: users insisting credits drained while they “weren’t doing anything,” then discovering a cloud task, a retry loop, or a parallel session was metering the whole time.

We treat these as anecdotes, not measurements, but the mechanism behind them is real and predictable. Agent work spends continuously without keystrokes, deep reasoning settings multiply the cost of identical prompts, and failed attempts meter as fully as successes. A credit balance is a bucket with several taps open at once. If you do run on credits, run /status first, close the sessions you are not using, and drop reasoning effort for routine tasks.

Credits vs the alternatives

The honest comparison is between three relief valves, and the right one depends on how often overflow happens:

Situation	The sane option
Overflow a few times a year, deadline-shaped	On-demand credits
Overflow most months, predictable	Second owned account, flat cost
Overflow that must never block automation	API-key fallback lane, switched automatically

The arithmetic for the middle row: a second Plus account costs $20 a month and, by our planning estimate, absorbs roughly $700 of API-equivalent work (an estimate, never a guarantee). For credits to beat that, your overflow has to be genuinely tiny. If overflow happens every month, a second account caps the cost; credits never do. Each account must be your own, with its own subscription, since OpenAI’s terms prohibit sharing accounts between people; the mechanics of running two are in Codex with multiple ChatGPT accounts.

The bottom row is the operational answer. An API key has no windows at all, just per-token billing, which makes it the natural lane of last resort for production work. Switching lanes by hand defeats the point, so we built the ordering into Codex Hosted: subscription lanes first, your API key last, every request logged with the lane that served it. The full behavior is in what happens when you hit your Codex usage limit.

When to skip credits entirely

Skip them when overflow is a pattern rather than an event. Buying credits most months means your sustained load has outgrown the plan, and you are paying metered rates on top of a flat subscription, the worst of both models. Compare a typical month of credit spending against the price gap to the next tier; the upgrade usually wins, and it comes with bigger windows everywhere, not just during the crunch that prompted the purchase.

Skip them for automation, always. A pipeline that stalls until a human buys credits in a web UI is not a pipeline. Give unattended work a fallback lane instead.

And skip them when the real problem is visibility. Plenty of credit purchases are panic buys that a glance at the meter would have prevented; pacing against a known reset time is free.

Credits have a job: absorbing the rare deadline that lands in the wrong window. For everything past that, size the plan to the workload and give overflow somewhere to go. The calculator maps your actual monthly spend to a tier in thirty seconds.