Codex pricing explained, with what a real task costs

AutoRouter · Nov 11, 2026 · 11 min read

As of November 9, 2026, Codex comes with a ChatGPT plan: Plus at $20 a month, Pro at $100, $200 or $500, and Business at $20 per user a month billed annually. With an API key you pay per token instead, and GPT-6.1 Sol, which OpenAI describes as near-Astra performance at a lower cost, costs $2 per million input tokens and $10 per million output tokens. OpenAI publishes estimates instead of fixed limits, such as 15 to 160 local GPT-6.1 Sol messages per five hours on Plus. In a test on real bug reports from an open-source project, a Codex task on GPT-6.1 Sol cost $0.085 at API list prices, about 2.1 ChatGPT credits. AutoRouter, the Mac app for Codex and Claude Code, ran that test and wrote this post.

How much does Codex cost on each ChatGPT plan?

Codex is included in ChatGPT Plus at $20 a month and every plan above it, as of November 9, 2026 (OpenAI). Free and Go get a limited version: GPT-6 Luna at Standard speed in the ChatGPT desktop app, subject to rollout.

OpenAI recommends the API key for automation in shared environments like CI. The pricing page links OpenAI's comparison of the Pro plans, and How do Codex and Claude Code compare? sets ChatGPT's plans beside Claude's. For Anthropic's side alone, see Claude Code pricing explained, with what a real task costs.

What do Codex usage limits mean in practice?

Codex usage limits are estimates that depend on the model and the task: OpenAI says small scripts may use a fraction of the allowance while long tasks that hold more context use significantly more per message. For Plus and standard Business seats it estimates these local message counts per five hours, as of November 9, 2026:

• Weekly limits may also apply, and local messages and cloud tasks share one allowance; cloud tasks may use more of it (OpenAI).
• Pro has no five-hour limit for now. OpenAI publishes no message counts for it.
• Speed costs usage. Fast mode uses included limits at 2.5 times the Standard rate, and Ultrafast at 8 times (OpenAI).
• ChatGPT Work shares the pool. Work in ChatGPT uses the same pricing, credits and usage limits as Codex (OpenAI).
• At a limit, Codex can finish the task it is working on; then you can buy ChatGPT credits on Plus or Pro, or run local sessions with an API key at API rates.

Your current limits and reset times are in the usage dashboard, or /status in a Codex CLI session. Codex usage limits explained, and what to do when you run out covers resets and each option at the limit.

What do ChatGPT credits and API tokens cost for Codex?

GPT-6.1 Sol costs $2 per million input tokens, $0.10 per million cached input tokens and $10 per million output tokens on the API, and 50, 2.5 and 250 ChatGPT credits for the same tokens, as of November 9, 2026 (API, credits). Credits pay for usage past a plan's included limits. For these four models, every credit rate is 25 times the API price in dollars.

The API prices are for prompts up to 272,000 tokens; longer prompts cost more, though GPT-6.1 Sol drawn from a subscription has no long-context surcharge. The API also lists cache-write prices, $2.50 per million tokens for a 30-minute write on GPT-6.1 Sol, while Codex credit billing has no separate cache-write charge. OpenAI doesn't publish one dollar price for a credit: purchase prices and discounts depend on your plan or agreement. It also says API token prices shouldn't be used to estimate included tasks on a plan.

What does a real Codex task cost?

A Codex task on GPT-6.1 Sol cost $0.066 to $0.125 at API list prices in AutoRouter's November 9, 2026 test, or 1.7 to 3.1 ChatGPT credits. Four tasks were real bugs from the click library's issue tracker, given to codex exec as the original bug report, and one was writing tests for a module. A run passed when the maintainers' own fix tests passed and the rest of click's suite still did. Each figure is the mean of three fresh runs with no router in between (method).

The cost is Codex's reported input, cached input and output tokens times GPT-6.1 Sol's list prices; Codex reported no cache writes. Credits use the same tokens at the credit rates. Most of each run's input came from the cache, about 90% of it, which is why a task that read 150,000 to 400,000 tokens stayed under 13 cents. Bigger repositories and longer sessions cost more.

At $0.085 a task, $20 pays for about 235 tasks like these on the API. On a plan that arithmetic doesn't apply, because OpenAI meters included usage separately; the next section has outside measurements of what a plan holds.

In the same test, Claude Code on Claude Opus 5.5 passed all 15 runs and cost $0.252 per task as Claude Code priced it on a Claude Max plan, about three times Codex's cost; Codex passed 14 and finished faster, at a median of 57 seconds against 69. Each tool ran its default or recommended model, so the result compares the tools as shipped, on five tasks in one Python library. Codex vs Claude Code has the full results.

GPT-6 Luna costs a twentieth of GPT-6.1 Sol per input and output token, and OpenAI describes it as suited to focused, repeatable tasks. The test didn't measure Luna. Give it trivial jobs such as commit messages and one-line commands, and keep bug fixes and multi-file changes on GPT-6.1 Sol or GPT-6 Astra; How to send quick coding tasks to a cheaper model shows how in Codex.

Are ChatGPT plans subsidized for Codex?

Measured at API list prices, a ChatGPT plan holds several times its price in Codex usage, though less than a Claude plan by outside measurements. OpenAI publishes message estimates instead of dollar allowances, so the figures come from outside tests that watch how far a plan's usage meter moves for a known number of tokens, then price those tokens at API rates:

These are measurements by Wokey, which runs a gateway, of real subscription accounts, not figures OpenAI publishes. On the mid-tier models, SemiAnalysis measured about 5 times the API-equivalent value on Claude plans as on ChatGPT plans, comparing Claude Opus 5.5 with GPT-6.1 Sol. It also found that OpenAI halved the tokens each model tier gets on the $200 Pro plan, and that keeping that plan's Sol limits unchanged when GPT-6.1 Sol launched lowered its API-equivalent value by about 30%. Claude Code pricing explained has Anthropic's side.

OpenAI has said a plan can cost it more than it earns. In January 2025, Sam Altman wrote that OpenAI was losing money on the $200 ChatGPT Pro plan because people used it more than expected; that was the earlier Pro plan, before today's Codex limits. Value at API prices also isn't the same as a loss: SemiAnalysis estimates that subscriptions bring in about 10% of revenue but can take over 40% of inference compute, and that a lab's margin on a plan depends on how much of the limit each subscriber uses.

For a developer, the practical point is the break-even: once a month of your Codex work at API prices would cost more than a plan, the plan is the cheaper way to pay, and from then on its limits are what you manage.

How does AutoRouter change what Codex costs?

AutoRouter, the Mac app for Codex and Claude Code, adds 20+ models to Codex and picks the model for each task, so the ChatGPT plan you pay for goes to the work that needs it. In Codex, OpenAI models run on your ChatGPT plan. AutoRouter works out how hard each task is: hard work goes to the most capable model your plans cover, and to credits only when none is strong enough, while everything else goes to the model that costs you least, which is often your plan or a low-cost open model. A model has to clear a quality bar for the task before price counts, and you can see each request's model, what it cost and why.

AutoRouter Free costs $0 a month, and routing across your ChatGPT and Claude plans is free. Pro costs $50 a month with $20 of included usage and adds every other model at the provider's price plus 10%; when your ChatGPT plan hits its limit, Pro keeps that work going from the included usage until the plan resets. New accounts get Pro free for their first month, or until $20 of usage is spent. AutoRouter earns the 10% on paid models and nothing on plans. It works with the Codex app and CLI on macOS 14 or later and is in beta; you pick AutoRouter from the model menu in Codex. More ways to bring the bill down are in How to cut Claude Code and Codex costs.

Questions

Is Codex free?

Partly. As of November 9, 2026, ChatGPT Free and Go ($8 a month) include GPT-6 Luna at Standard speed in the ChatGPT desktop app, subject to rollout (OpenAI). Codex in the CLI, the IDE extension and the cloud, with GPT-6.1 Sol, starts at Plus for $20 a month.

Is Codex included in ChatGPT Plus?

Yes. Plus at $20 a month includes Codex on the web, in the CLI, in the IDE extension and on iOS, with GPT-6.1 Sol and GPT-6 Luna, cloud code review and the Slack integration (OpenAI). Its usage is estimated per five hours, and weekly limits may also apply.

How much does a ChatGPT credit cost?

OpenAI doesn't publish a single dollar price per credit; it says purchase prices and discounts depend on your plan or agreement. What it does publish is the credit rate per million tokens for each model, and a GPT-6.1 Sol task in AutoRouter's test used 1.7 to 3.1 credits.

Is an API key cheaper than ChatGPT Plus for Codex?

Only for light use. At the $0.085 measured per task, $20 of API usage covers about 235 tasks like those in AutoRouter's test, and Plus costs $20 for an allowance OpenAI estimates in messages per five hours. An API key also drops the cloud features, so OpenAI positions it for automation like CI. Subscription or API: which is cheaper for coding agents? works through the break-even.

Is Codex cheaper than Claude Code?

In AutoRouter's November 9, 2026 test, Codex on GPT-6.1 Sol cost $0.085 per task at API list prices against $0.252 for Claude Code on Claude Opus 5.5, and Claude Code passed all 15 runs to Codex's 14. Both start at $20 a month on a plan; Claude Code pricing explained covers Anthropic's plans.

Sources

• Pricing, Codex docs, OpenAI, checked November 9, 2026
• API pricing, OpenAI, checked November 9, 2026
• Models, Codex docs, OpenAI
• pallets/click, GitHub, issues #2819, #3125 and #3360 and pull request #3884
• Anthropic Subscriptions Offer 5x+ More Value Than OpenAI, SemiAnalysis, October 5, 2026 (partly paywalled)
• Codex usage limits, measured, Wokey, measured October 19 to November 2, 2026
• OpenAI is losing money on its pricey ChatGPT Pro plan, CEO Sam Altman says, TechCrunch, January 5, 2025