OpenAI’s Codex pricing changed in a meaningful way on April 2, 2026. Instead of leaning on rough per-message estimates, OpenAI shifted Codex toward a token-based pricing model built around credits, model selection, and workspace billing. For teams, that makes Codex easier to reason about in one way and more confusing in another.
The short version: there is no single Codex price. What you pay depends on whether you use Codex through included ChatGPT access, add Codex-only seats, buy workspace credits, run a heavier model like GPT-5.5, or use faster premium execution modes.
What changed in Codex pricing
OpenAI says the April 2 update moved Codex from message-based estimates to token-based metering for new and existing Plus, Pro, and ChatGPT Business customers, plus new ChatGPT Enterprise plans. On April 23, 2026, OpenAI extended that update to existing Enterprise plans as well, including Edu, Health, Gov, and ChatGPT for Teachers.
That matters because Codex now maps more directly to model usage. Instead of treating every task as roughly the same, OpenAI prices Codex around three token categories:
- Input tokens
- Cached input tokens
- Output tokens
That is better for serious budgeting, but it also means teams need to understand where their token mix is actually going.
Seats versus credits: the first pricing distinction that matters
Codex pricing is now split between access and usage.
Standard ChatGPT Business seats
In ChatGPT Business, standard seats include ChatGPT access and Codex access. OpenAI’s current self-serve pricing is:
- Monthly: $25 per user per month
- Annual: $20 per user per month, billed annually
Those seats require a minimum purchase of two.
Codex-only seats
Codex-only seats are different. They have a fixed seat price of $0 per user per month, require no minimum seat count, and rely on workspace credits for actual activity. That makes them attractive for small pilots, contractor access, or teams that want deep coding-agent usage without full ChatGPT seats.
In other words, the seat itself may be free, but the work is not. The real billing happens through credits consumed by model usage.
The current Codex rate card
OpenAI’s current token-based Codex rate card shows how quickly costs can diverge by model:
| Model | Input | Cached input | Output |
|---|---|---|---|
| GPT-5.5 | 125 credits / 1M | 12.5 credits / 1M | 750 credits / 1M |
| GPT-5.4 | 62.5 credits / 1M | 6.25 credits / 1M | 375 credits / 1M |
| GPT-5.4 Mini | 18.75 credits / 1M | 1.875 credits / 1M | 113 credits / 1M |
| GPT-5.3-Codex | 43.75 credits / 1M | 4.375 credits / 1M | 350 credits / 1M |
OpenAI also notes that code review uses GPT-5.3-Codex, and that Fast mode consumes credits at a higher rate for supported models.
This is the key budgeting reality: moving from GPT-5.4 to GPT-5.5 in Codex is not a small step. It roughly doubles credit consumption across input, cached input, and output.
What a Codex workload actually costs in credits
Because Codex is now token-metered, a team should estimate usage from real workflows rather than a flat “per developer” mental model.
For example:
| Workload | Estimated credit use |
|---|---|
| GPT-5.5 with 10M input + 2M output | 2,750 credits |
| GPT-5.4 with 10M input + 2M output | 1,375 credits |
That gap matters. If you move a team’s default Codex workflow from GPT-5.4 to GPT-5.5 without changing behavior, your credit burn can jump sharply.
OpenAI also says that, on average, Codex costs around $100 to $200 per developer per month, but it warns that variance is large depending on model choice, number of instances running, automations, and use of Fast mode. That is best read as a rough midpoint, not a reliable cap.
Where teams overspend on Codex
1. Confusing free seats with free usage
Codex-only seats remove the fixed seat fee, but they do not remove metered spend. If usage grows quickly, credits become the real budget line.
2. Defaulting everyone to GPT-5.5
GPT-5.5 is powerful, but the rate card makes clear that it is expensive relative to GPT-5.4, GPT-5.4 Mini, and GPT-5.3-Codex. Not every developer workflow needs the most expensive model.
3. Ignoring output-heavy behavior
Large diffs, long code explanations, detailed agent notes, and verbose review output can turn output tokens into the biggest cost driver.
4. Letting Fast mode become the norm
Fast mode can be worth it for latency-sensitive work, but it is a premium path. Teams that leave it on broadly may trade convenience for a surprisingly higher burn rate.
5. Running too many concurrent agent instances
Codex becomes more useful as teams parallelize work. It also becomes easier to lose visibility when many agents, local tasks, reviews, and automations are running at once.
How to budget Codex more intelligently
The smartest Codex budget is not “buy credits and hope.” It is a model-routing and workflow-design problem.
- Use GPT-5.5 for harder implementation, debugging, and multi-step engineering tasks.
- Use GPT-5.4 or GPT-5.3-Codex for cheaper review, iteration, and mid-complexity work.
- Take advantage of cached input wherever repository context or standing instructions repeat.
- Reserve Fast mode for work where latency really matters.
- Watch automations and concurrency, because parallel usage compounds fast.
What teams should budget for Codex in 2026
If your team mainly wants lightweight experimentation, Codex-only seats plus a modest credit budget can be a very efficient starting point. If your team wants broad AI access, standard ChatGPT Business seats may be simpler because they bundle baseline Codex access and broader workspace features.
But the real Codex budget question is no longer “how many seats do we need?” It is “which model is running which class of work, how often, and with how much output?”
That is why the best way to control Codex spend is not just finance approval. It is workflow design. Teams that separate expensive agent work from cheaper repetitive work will get a much cleaner return from Codex than teams that let every task hit the same premium model path.