Codex Credits: About $0.04 Each, Billed Per Token

In brief

Codex credits work out at about $0.04 each as of November 2026, since OpenAI's student offer values 2,500 credits at $100. OpenAI bills them per token: GPT-6.1 Sol output costs 250 credits per million tokens. Credits only spend after your plan's included usage runs out, and they expire 12 months from purchase.

What is a Codex credit, and what does it cost?

A Codex credit is the unit OpenAI charges once your ChatGPT plan's included Codex usage runs out, and it works out at about $0.04 each. OpenAI prints no list price on its pricing page or credits guide. Its credits article says amounts and prices "can vary by account, region, and plan." But the Codex for students page values the grant at $100 and 2,500 credits. That is 25 credits a dollar. We read all three pages on 21 November 2026.

Five chrome discs in falling sizes on a long black plinth, a terracotta OpenAI knot on a round plinth beside them, and a rainbow glass hourglass with its sand half run on a third plinth, in a charcoal studio with rim light.

What a credit buys depends on the model that you pick. OpenAI charges per million tokens, and it sets three rates for each model: one for fresh input, one for cached input and one for output. Here are the output rates, since that is the dear part of most coding tasks:

One model costs 100 times more per output token than another
10 credits100 credits1,000 creditsGPT-6 LunaGPT-6.1 TerraGPT-6.1 SolGPT-6.2 AstraGPT-6 Astra

Those are credits per million output tokens at Standard speed, from OpenAI's Codex pricing page. The rate card is below. The last column shows the per message averages OpenAI still prints, which it labels as its legacy Codex rate card.

Model, as of November 2026Input per 1M tokensCached input per 1MOutput per 1MLegacy average credits per local messageAbout, in dollars
GPT-6 Luna2.50.2512.5about 1$0.04
GPT-6.1 Sol502.5250about 4$0.16
GPT-6.2 Sol502.5250about 4$0.16
GPT-6 Sol505250about 5$0.20
GPT-6.2 Astra200101,000about 9$0.36
GPT-6 Astra250251,250about 16$0.64

The rates and the averages come from OpenAI's ChatGPT rate card. OpenAI says most customers have moved to token based pricing and that the per message averages apply to the small set of Enterprise workspaces not yet migrated. Treat them as a rough guide to relative cost, not a bill. We converted them to dollars at 25 credits a dollar. Long or tool heavy turns land well above an average.

How many credits does a Codex task use?

A Codex task uses credits in proportion to its tokens, so the cost runs from about 1 credit for a short GPT-6 Luna message to far more for a long Astra session. OpenAI's legacy averages are about 4 credits a local message on GPT-6.2 Sol, roughly $0.16. It prints no credit figure for cloud tasks, only a note on the Codex pricing page that they "may use more" of your allowance than local messages. Treat cloud runs as dearer than the figures in the table above.

You can price a task yourself. Credits equal input tokens over a million times the input rate, plus cached tokens over a million times the cached rate, plus output tokens over a million times the output rate. That formula is OpenAI's. Here's our own example on GPT-6.1 Sol: 100,000 fresh input tokens cost 5 credits, 20,000 output tokens cost 5, and 400,000 cached tokens add 1. That turn is 11 credits, about $0.44. Paste a prompt into the token counter to get the input side.

Speed changes the bill. On OpenAI's legacy rate card, Fast mode spends at 2 times the Standard rate and Ultrafast at 6 times on GPT-6.2 Astra and GPT-6.2 Sol. OpenAI also runs a 3 times Ultrafast promotion on token based pay as you go, and its rate card lists no end date, so check it before relying on either figure. Voice in ChatGPT costs 1.25 credits a minute.

How do Codex credits work with your plan limits?

Codex credits work as overflow: OpenAI uses your plan's included usage first, and only then draws from your credit balance. The credits guide says it directly. Included plan usage goes first, and after plan limits are reached, usage draws from the credits. So credits never make your plan allowance smaller, and they never refill the 5 hour or weekly window. What they do is let you keep working once you are past it.

Plus has a 5 hour window and a weekly cap. The Codex pricing page estimates 15 to 160 local messages on GPT-6.1 Sol per window on Plus. It also says Pro plans currently have no five hour limit, so Pro users meet the weekly cap and then credits. Codex usage limit covers the windows in full. If you run out mid turn, OpenAI says Codex can finish that turn, subject to fair use.

One pool backs several products. The Codex help page says Codex and several other features "use a shared allowance and credit pool when those features are available on your plan." So a long Work task spends the same balance a Codex session would. Support can't reset your limits, per the same page.

Who can buy Codex credits?

Eligible Plus and Pro users can buy Codex credits without changing plans, and business workspaces have their own Codex credit plans. The credits guide says "Credit purchases are currently available to a limited group of Free and Go users." If you see no option to add credits, purchases aren't open to your account yet. The guide says Plus and Pro users may see the add credits option after they hit a Codex usage limit.

The price is set at checkout. Auto reload lets you set a minimum and a target, and OpenAI buys just enough to reach the target. OpenAI's example: minimum 100, target 500, balance 40 gives a 460 credit purchase. A monthly spend limit you set can reduce that amount.

Business and Enterprise billing differs by contract, so a team should ask its admin or OpenAI Sales rather than rely on the personal plan rules above.

Do Codex credits expire, and are they refundable?

OpenAI says "Credits expire 12 months from purchase" and "Unused credits do not roll over." Credits purchased in Japan after 13 November 2026 expire 180 days from purchase instead. So buy only what you expect to use in the next year, and don't buy a bulk pack just for comfort.

Refunds are narrow. OpenAI says credits are "non-refundable unless required by law or contract, or an approved exception applies." Unused credits may qualify after a verified billing error, confirmed unauthorized activity, or a confirmed OpenAI service failure.

A bad habit to avoid is topping up before the plan allowance is gone. Included usage runs first anyway, so early credits just sit on a 12 month clock.

Can students get free Codex credits?

Yes. Verified students at degree granting universities in the United States or Canada can claim $100 of Codex credits, which OpenAI says is 2,500 credits. The student page requires verification through SheerID and residence in the US or Canada when you claim. It allows one offer per student.

The credits "expire 12 months after the grant date." OpenAI lists no claim window on the page, as read on 21 November 2026. At 4 credits a message, OpenAI's legacy average for GPT-6.1 Sol, 2,500 credits pay for about 625 messages, by our math.

Outside the US and Canada there is no student grant on that page. The cheapest route there is the one everyone has: pick the cheaper model.

Are Codex credits cheaper than upgrading your plan?

For a bad week now and then, credits are the cheaper fix; for every week, a bigger plan may be. At $0.04 a credit, each $80 of credits is 2,000 credits, about 500 legacy average GPT-6.1 Sol messages. If you would spend that every month, compare it with the price gap to the next plan before you top up. Codex pricing lays out every plan.

Four habits stretch a balance. Use GPT-6 Luna for reading and small edits, since its output rate is 20 times lower than GPT-6.1 Sol's. Turn off Fast mode, which doubles the rate. Keep your prompts cached by not rewriting the start of a long session, because that is what the cheap rate pays for. And buy small, because of the 12 month expiry. An API key is the other route. OpenAI says API key sessions "use separate API pricing, not your ChatGPT allowance."

Where does Whizi fit if you pay for AI tools?

Whizi is a chat app, so it can't replace Codex or edit your repo. It's the place to ask Claude, GPT and Gemini the same coding question side by side while your Codex balance lasts. Whizi Pro costs $29.99 a month, its first week is $1.99, and side by side Compare comes with Pro. Web search is on by default in every chat, and you can upload PDF, Word and Excel files.

The free plan gives 10 messages a day on Whizi AI, Auto and Claude Haiku, with no card. It's a fair place to test a question before you spend credits on an agent run. Whizi loses on anything agentic: it won't run tests, open pull requests or work in your terminal. For that, pay OpenAI.

Workflow checklist
  • Check that your plan allowance is empty before you buy credits, since included usage is spent first
  • Price a task with the rate card: input, cached and output tokens each at their own rate
  • Switch to GPT-6 Luna for file reading and small edits, at about 1 credit a message on the legacy card
  • Turn off Fast mode, which costs 2 times the Standard rate on OpenAI's legacy card
  • Buy small packs, because credits expire 12 months after purchase and don't roll over
  • If you're a US or Canadian student, claim the 2,500 credit offer before you pay for anything
Common questions

Frequently asked questions

How many Codex credits is $20 worth?

About 500, at the $0.04 a credit that OpenAI's student offer implies, since it values 2,500 credits at $100. OpenAI prints no list price on its help pages, and it says prices can vary by account, region and plan. Your checkout shows the real figure. At OpenAI's legacy average of about 4 credits a GPT-6.1 Sol message, 500 credits is roughly 125 messages.

Can you buy Codex credits on a free plan?

Only for a limited group, so most Free and Go users can't. OpenAI says credit purchases are open to a limited group of Free and Go users. Eligible Plus and Pro users can buy credits without changing plans.

How many tokens is one Codex credit?

It depends on the model and the kind of token. On GPT-6.1 Sol, one credit buys 20,000 fresh input tokens, 400,000 cached input tokens or 4,000 output tokens, from OpenAI's rates of 50, 2.5 and 250 credits per million. On GPT-6 Luna the same credit buys 400,000 input or 80,000 output tokens. The token counter counts the tokens in your prompt.

Do Codex credits work for ChatGPT Work and Excel?

Yes, one balance covers them. OpenAI says Codex and several other features use a shared allowance and credit pool when they are available on your plan. A big Work task can therefore leave fewer credits for Codex.