What Is OpenAI Codex?
OpenAI Codex is OpenAI's agentic coding tool, built directly into ChatGPT and accessible via the Codex CLI. It is not a standalone product — it is included in every paid ChatGPT plan: Plus ($20/mo), the new $100/mo Pro tier, and the full Pro at $200/mo. Business and Enterprise teams can add Codex-only pay-as-you-go seats with no fixed seat fee, billed purely on token consumption. As of April 2, 2026, all pricing moved to token-based billing aligned with API rates.
The April 16, 2026 "Codex for (almost) everything" update is the most significant change since launch. It added computer use (Mac control), persistent memory, image generation, background scheduled tasks, and integration with 90+ external tools including GitHub, Slack, and Linear. Codex is now powered by GPT-5.3-Codex and GPT-5.4 models depending on task type, with GPT-5.3-Codex-Spark available as a fast research-preview option for Pro users.
Pricing
| Plan |
Cost |
Codex Access |
| ChatGPT Free / Go |
Free / $8/mo |
Very limited — not viable for daily use |
| ChatGPT Plus |
$20/mo |
Full Codex access, standard usage limits. Best for individual developers. |
| ChatGPT Pro ($100) |
$100/mo |
5x more Codex usage than Plus. New April 2026 tier for heavy coding sessions. |
| ChatGPT Pro ($200) |
$200/mo |
20x Plus usage (ongoing). 25x through May 31, 2026. |
| Business / Enterprise |
$20/seat + usage |
Codex-only pay-as-you-go seats available. SOC 2, SSO, SCIM, audit logs. |
What Changed on April 16, 2026
The "Codex for (almost) everything" update repositioned Codex from a coding assistant into a full developer workstation. Five major additions shipped together:
Computer use: Codex can now control your Mac — open applications, click through interfaces, and automate desktop workflows. The browser automation currently works on localhost only, not the open web.
Memory: Codex now remembers preferences, coding style, and project context across sessions. Rolling out gradually — not all users had it on day one.
Image generation: Integrated directly into Codex sessions. Useful for generating UI mockups and assets without leaving the coding environment.
Scheduled background tasks: Codex can now run automations on a timer — code reviews, test runs, and dependency checks without manual triggers.
90+ plugin integrations: Native GitHub, Slack, and Linear connections plus dozens of third-party tools via MCP. Notion, Ramp, Braintrust, and Wasmer are publicly confirmed enterprise users.
Benchmark Performance
Codex is powered by GPT-5.4 — the same model that leads AIME 2026 (99.2%) and GPQA Diamond (92.8%). On coding-specific benchmarks, GPT-5.4 scores 57.7% on SWE-Bench Pro — second only to Kimi K2.6 (58.6%) among publicly reported models. On OSWorld (desktop GUI task completion), GPT-5.4 leads the field at 75.0%, which is directly relevant to Codex's new computer use capability. Terminal-Bench 2.0: 65.4%.
What Codex Does Well
Zero setup friction: If you have a ChatGPT Plus account, you already have Codex. No separate install, no API key management, no configuration. This is the biggest practical advantage over Claude Code and Kimi Code for non-technical users or teams that need fast onboarding.
GitHub integration: Native — not bolted on. Codex can read your repository, open pull requests, review diffs, and push changes without leaving the interface. The April 16 update added Linear and Slack to the same native integration tier.
Enterprise readiness: SOC 2 Type 2, SSO, SCIM, RBAC, audit logs via Compliance API, and EKM. No training on Business or Enterprise data by default. This is where Codex clearly outperforms Kimi K2.6 and DeepSeek — enterprise compliance is mature and documented.
OSWorld GUI automation: 75.0% on the benchmark that tests desktop interface task completion. If your agent workflows involve clicking through graphical interfaces rather than terminal commands, Codex leads every alternative.
Limitations
Cost at scale: $0.60/M input tokens (Kimi K2.6) versus token-based rates on GPT-5.4 that are significantly higher. For teams running agentic coding pipelines at volume, the cost difference is material. The $20/mo Plus plan is excellent value for moderate use; the API becomes expensive at scale.
Computer use is Mac-only: The April 16 update brought computer use to Mac. Windows support is not yet available. Browser automation is localhost-only — it cannot browse the open web autonomously.
No open weights: Cannot self-host. Every token goes through OpenAI's infrastructure. For teams with data residency requirements outside the US, this matters.
Codex-Spark is Pro-only: The fast GPT-5.3-Codex-Spark model is a research preview available only on the $200/mo Pro plan, not the new $100/mo tier.
Who Should Use It
Use Codex if you are already on ChatGPT Plus, need GitHub-native integration, require enterprise compliance (SOC 2, SSO), need desktop GUI automation, or want zero-friction onboarding for a non-technical team.
Consider Kimi K2.6 or Claude Code if you run agentic coding pipelines at high token volume where API cost is a primary constraint, need self-hosting, or prioritize SWE-Bench Pro performance over ecosystem maturity.
Verdict
Codex is the most complete developer workstation in the AI coding space as of April 2026 — not because it leads every benchmark, but because it covers the most ground in a single product. Computer use, memory, image generation, background tasks, GitHub/Slack/Linear integration, and enterprise compliance all in one tool included from $20/mo. The limitations are real (Mac-only computer use, higher API costs at scale, no self-hosting) but for most individual developers and enterprise teams, Codex is the default starting point before evaluating specialist alternatives.