Codex CLI Pricing: Every Plan, and Which One You Need for Parallel Agents

The plans at a glance

As of October 31, 2026, Codex access maps onto the ChatGPT plans like this. There is no plan you can buy that gives you Codex and nothing else.

The Codex pricing page does not publish a usage multiplier for each Pro tier. It sends you to OpenAI's Pro help article to compare limits. What the pricing page does state: Pro plans currently have no 5 hour limit, weekly limits may still apply, and only Pro $500 gets Ultrafast mode. If Codex is the only reason you are paying, start at $100 and move up only when the weekly limit stops you.

Go at $8/month is the cheapest paid door, but OpenAI pitches it at lightweight coding tasks and lists GPT-6 Luna for it, not GPT-6.1 Sol. It is fine for small jobs, not for a working day.

The trap: one quota across the CLI, the web and the IDE

This is the detail that catches people out, and it is worth putting above the mechanics of the windows themselves: Codex CLI usage draws from the same plan allowance as Codex on the web and Codex in your IDE. OpenAI's pricing page says local messages and cloud chats share your plan's usage allowance, and that ChatGPT Work and Codex share usage too.

In practice that means a morning spent driving Codex from the web interface directly reduces what your terminal can do that afternoon, and it means the IDE extension quietly running in the background counts against the agent you thought had the quota to itself. If you use more than one surface, your effective allowance is smaller than any single-surface estimate suggests.

Add parallel terminals on top and the arithmetic gets unforgiving fast. Four Codex sessions plus an IDE extension is five consumers on one meter.

How the quota actually works: a 5 hour window plus a weekly cap

On Plus and Business, Codex meters you on two layers. The first is a rolling 5 hour window: an intense session can exhaust it while your weekly allowance is barely touched, and the fix is to wait for the window to slide forward. Pro plans currently skip this layer.

The second is a weekly cap on top, which OpenAI says "may also apply" on every plan. When both apply, being blocked tells you nothing by itself about which ceiling you hit.

What makes Codex unusual is that the published figures are given per model, as ranges, rather than as a single number. On Plus, OpenAI estimates 15 to 160 local messages per 5 hours with GPT-6.1 Sol, 35 to 285 with GPT-6.1 Terra and 350 to 3,000 with GPT-6 Luna, because a model that thinks longer consumes more of the same allowance. The practical consequence is that your model choice changes your effective quota more than it changes your bill, which is the opposite of how pay-per-token works.

Which plan you need when several Codex sessions run at once

Because the CLI, the web version and the IDE extension share one allowance, the question that decides your tier is not how many hours you code, it is how many Codex sessions are consuming at the same time. Every extra terminal is another consumer on the same meter, and so is the IDE extension you forgot was running.

Before you upgrade, check the two cheaper levers this guide already covers: drop to a faster model for routine work, which buys more requests inside the same window, and keep credits for the occasional deadline. The mechanics of driving several terminals are in run multiple Codex sessions, and the wider workflow in the Codex agent swarm guide.

Credits: the pressure valve on top of the plan

When you run past your included allowance, Codex does not simply stop if you have credits on the account. Credits are a top-up balance that maps onto token usage, priced separately for input, cached input and output tokens: GPT-6.1 Sol costs 50 credits per million input tokens and 250 per million output tokens, and Fast mode doubles the rate. OpenAI says the purchase price of a credit depends on your plan or agreement, so check it in your account.

This is a genuinely useful design: on Codex you can buy your way through a deadline. Claude paid plans can now do something similar with usage credits, so the real difference is how each provider meters and prices overflow. Treat credits as an occasional rescue rather than part of normal capacity, or a $20 subscription stops being a $20 budget.

The honest rule of thumb: if you buy credits more than once or twice a month, the next tier up is cheaper than the habit.

The pay-per-token alternative: an OpenAI API key

Codex accepts an OpenAI API key instead of a ChatGPT login, which swaps the whole quota system for per-token billing. As of October 31, 2026, OpenAI's API pricing lists GPT-6.1 Sol at $2.00 per million input tokens, $0.10 for cached input and $10.00 per million output tokens, and GPT-6 Luna at $0.10 input and $0.50 output.

For interactive work this is usually the worse deal: coding agents are output-heavy and output is where per-token billing bites. Where it wins is automation. A CI job, a scheduled refactor, a batch of repository migrations, anything that runs unattended and must not fail because a human's personal quota happened to be exhausted, belongs on an API key with a spending cap set in the OpenAI console.

How Codex pricing compares with the other agent CLIs

A price only means something next to the alternatives. The $8 Go tier puts Codex at the low end of the category, next to the $8 Plus plans from Cursor and xAI and Google AI Plus at $4.99, and its allowance is shared with ChatGPT on the web and in your IDE. Here is the whole category, one line each, checked on the same day.

Each row has its own guide with the full ladder and the quota mechanics behind it: Claude Code, Kimi Code, OpenCode, Antigravity, Grok Build and Cursor Agent.

Is Codex free? What the free tier actually allows

Partly. The Codex CLI is a free, open-source download, and ChatGPT Free does include some Codex access. Unlike Claude Code, which is absent from its free tier entirely, you can genuinely try Codex without paying.

What you cannot do is work on it. The free allowance is sized to let you form an opinion, not to carry a project. Expect to hit the ceiling on your first real task.

The realistic budget framing: evaluating Codex costs $0, using it as a daily single-agent assistant costs $20/month, with Go at $8 for light work, and running it as a parallel agent workflow costs $100/month or more.

Watching Codex quota while agents run

A shared quota across three surfaces, metered on two windows, with per-model ranges rather than fixed counts, is not something you can track in your head. And the failure mode is the usual one: you find out when the agent stops.

CodeAgentSwarm, the desktop workspace for running several AI CLI agents in parallel, shows your Codex usage alongside your other agents' quotas, so you can see the wall approaching while the agents work rather than discovering it from a stalled terminal.

It pairs with the pattern from the Codex agent swarm guide: spread work across agents on different meters, so that exhausting one provider does not stop the whole session. Codex and Claude bill from entirely separate pools, which is a better reason to run both than any benchmark.