OpenAI's Codex is an AI coding agent integrated into ChatGPT that runs tasks in a cloud sandbox — reading repositories, writing code, running tests, and producing pull requests autonomously. Pricing changed in April 2026 when OpenAI switched from per-message billing to token-based billing.
What is included in each ChatGPT plan?
Codex agent access comes with existing ChatGPT subscriptions. You do not need a separate product or credit purchase to use it.
| Plan | Monthly Cost | Codex Access |
|---|---|---|
| Plus | $20 | Included; usage-limited |
| Pro | $200 | Included; higher limits |
| Business | $20/user (annual) or $25/month | All users, admin controls, data not used for training |
| Enterprise | Custom | All users, compliance, SSO |
Starting April 2, 2026, Codex usage bills based on API token consumption rather than message counts. Additional credits can be purchased if you exhaust plan limits.
What are the Codex API token rates?
API access bills per million tokens with separate input and output pricing:
{{CODEX_API_PRICES}}
For most coding tasks, GPT-5.1 Codex mini offers the best cost-to-performance ratio. The higher-tier models are useful for complex multi-file reasoning and tasks that require the most capable model available.
Cached input tokens — context that repeats across turns in a session — cost approximately 10% of the standard input rate. In agentic coding sessions where repository context is carried across many turns, caching is the most impactful cost optimization available.
How much does Codex cost in practice?
Published data from enterprise deployments shows average spend of $100–$200 per developer per month, with significant variance based on workflow intensity.
In ChatGPT auth mode (the default), Codex usage draws from your plan's included limits at no extra charge. If you switch to API key mode, you are billed at standard token rates.
The Codex Pro 2x promotional pricing ended May 31, 2026. Users on the Pro plan at $200/month saw effective capacity halve from June 1 at the same price.
What are common mistakes to avoid?
- Using GPT-5.3 Codex via the API for tasks that GPT-5.1 Codex mini handles adequately
- Not enabling prompt caching for sessions where repository context repeats across many turns
- Switching to API key mode without tracking per-session token costs
- Running Codex in a loop without iteration limits, which can exhaust plan credits rapidly
- Not purchasing additional credits before hitting hard limits in production workflows
