← All guides

Codex pricing: what OpenAI Codex actually costs

Codex isn't a separate subscription — it comes bundled with ChatGPT paid plans, with usage limits, and heavy use spills over to API billing. Here's the full picture, and how to keep coding-agent costs down.

TL;DR: Codex is included with ChatGPT Plus ($20/mo), Pro ($200/mo), Business and Enterprise plans, each with its own usage allowance. There's no separate 'Codex fee', but once you exceed your plan's limits, additional usage is billed against the OpenAI API by tokens. So the real cost of Codex is your ChatGPT plan plus any API overage — and for heavy coding it adds up fast. If cost is the concern, a model-agnostic agent that lets you plug in cheaper or free models (or a local one) runs the same workflow for a fraction of the price.

Is Codex a separate paid product?

No. Codex isn't sold as its own subscription. It's the coding agent that ships inside ChatGPT and its editor/CLI extensions, and you get access through a paid ChatGPT plan. That means 'Codex pricing' is really 'which ChatGPT plan you're on, plus any usage beyond its limits.'

How Codex pricing actually works

Where the cost comes from

Plan / mechanismRoughlyWhat you get
ChatGPT Plus$20 / monthCodex access with a standard usage allowance — fine for light, occasional use
ChatGPT Pro$200 / monthMuch higher Codex usage limits for heavy, all-day agentic coding
Business / EnterprisePer seatTeam access with pooled/expanded limits and admin controls
API overagePer tokenOnce you pass your plan's limits, extra usage bills against the OpenAI API by input/output tokens

Exact allowances and rates change over time; check OpenAI's current pricing page for the latest numbers. The structure — plan + usage limits + API overage — is what stays constant.

Why the cost adds up for real coding

Agentic work burns tokens

A coding agent reads files, runs tools, and iterates — one task can be many model calls, so usage climbs quickly.

Limits push you to a higher plan

Hit the Plus allowance often and the pressure is to jump to the $200/mo Pro tier or eat API overage.

One model, one vendor

You're paying frontier-model prices for every request, including routine edits that a cheaper model handles fine.

How to run a coding agent for less

The biggest lever on cost is not being locked to a single premium model. A model-agnostic agent lets you route routine work to a cheap model (GLM, DeepSeek, Kimi) and save the expensive one for hard problems — or run a local model for free. You keep the agent workflow; you just stop paying frontier prices for everything.

Rule of thumb: if your coding-agent bill is climbing, it's usually because every request goes to one premium model. Being able to pick a cheaper model per task is where the savings are.

FAQ

How much does Codex cost per month?

Codex itself has no separate fee — you access it through a ChatGPT plan. That's roughly $20/mo for Plus, $200/mo for Pro (with much higher limits), or per-seat Business/Enterprise pricing. Usage beyond your plan's limits is billed against the OpenAI API by tokens.

Is Codex free?

There's no fully-featured free Codex tier for real coding work — access comes with paid ChatGPT plans. If you want a coding agent you can run for free, use a model-agnostic agent with a free or local model plugged in.

Why is my Codex usage so expensive?

Agentic coding makes many model calls per task, so token usage — and any API overage past your plan limit — adds up fast, especially when every request goes to a premium model rather than a cheaper one for routine work.

How do I use Codex-style coding for less?

Use a model-agnostic coding agent that lets you plug in cheaper models (GLM, DeepSeek, Kimi) or a local model and switch per task. Wuwei is a free, open-source agent that does exactly this — bring any model and route routine work to a cheap one.

Codex pricing comes down to your ChatGPT plan plus usage — and for heavy coding, that gets expensive because you're paying frontier prices for every request. If cost is the concern, Wuwei is a free, open-source coding agent that's model-agnostic: plug in a cheap or free model, run a local one, and switch per task. Same workflow, a fraction of the cost. See /en/vs/claude-code-free-alternative.

Windows · macOS · Linux — free, no login