Introduction
OpenAI Codex is OpenAI's coding agent. OpenAI describes it as an agent that writes code, runs tasks and builds multi-step workflows that automate coding work. You can't buy it on its own. It comes bundled with ChatGPT plans, from Free up to Enterprise, and it also runs on an OpenAI API key with per-token billing.
This review uses OpenAI's published pricing and documentation plus independent benchmark reporting. scored.tools did not test Codex hands-on, so nothing here describes how it feels to use. The question we can answer is whether the documented features and prices justify a subscription. Our score is 6 out of 10. The underlying model benchmarks well. How much Codex you actually get for your money is still unclear.
Key Features
An agent that writes code and runs tasks
OpenAI positions Codex as an agent that carries out tasks, going further than autocomplete. Per OpenAI's description, it writes code, runs tasks and chains steps into automated workflows. The workflow claims come from OpenAI alone. We found no independent evidence that the multi-step automation holds up on real codebases, which is why problem fit scores lowest in our criteria.
Where it runs
According to OpenAI's Codex pricing docs, Plus and higher plans get Codex on the web, the CLI, the IDE extension, iOS, the desktop app and cloud tasks. Free and Go users get the desktop app only, running GPT-6 Luna at Standard speed, and OpenAI says this is subject to rollout. If you use an API key you get the CLI, the SDK and the IDE extension, but no cloud features. That last point matters if you planned to hand off long background jobs while paying per token.
Model options and context
The tool data lists GPT-6 Astra with a 1.05M token context window and 128K max output tokens. A window that size can hold a large part of a mid-sized repository in one session. Codex plans also include access to GPT-6.1 Sol, GPT-6 Sol and GPT-6 Luna, and each model draws down your usage at a different rate (see the pricing section).
Benchmark performance
Independent benchmark reporting from CodeAnt AI puts the underlying model at 64.6% on SWE-bench Pro and 88.8% on Terminal-Bench. SWE-bench Pro measures fixes to real GitHub issues, and Terminal-Bench measures how well an agent works in a shell. Both scores are strong. They describe the model, though, and they can't tell you how Codex's agent loop, sandboxing or review flow will work on your repository.
API access
The OpenAI API is well documented and its per-token prices are public. The tool data lists no Codex-specific SDK or integration details beyond that, so our API score of 7 reflects the general OpenAI platform rather than anything built specially for Codex.
Pricing Breakdown
You can pay for Codex in two ways: a ChatGPT subscription with included usage, or an API key billed by the token. The figures below come from OpenAI's Codex pricing documentation and the tool data.
| Plan | Price | Codex access (per OpenAI docs) |
|---|---|---|
| Free | $0/month | Desktop app, GPT-6 Luna at Standard speed, subject to rollout |
| Go | $8/month | Desktop app, GPT-6 Luna at Standard speed, subject to rollout |
| Plus | $20/month | All surfaces including cloud; estimated 5 to 45 GPT-6 Astra or 15 to 160 GPT-6.1 Sol local messages per five hours |
| Pro | $100, $200 or $500/month | Same model estimates as Plus; OpenAI says there is no five-hour limit currently; Ultrafast mode on the $500 tier |
| Business | $20/user/month annual, $25 monthly (2+ users) | Same model estimates as Plus; workspace credits available |
| Enterprise / Edu | Contact sales | Flexible pricing scales with credits; admin and security controls |
| API key | Per token | CLI, SDK, IDE extension; no cloud features |
API token prices
| Model | Input per 1M tokens | Output per 1M tokens |
|---|---|---|
| GPT-6.1 Sol | $2.00 | $10.00 |
| GPT-6 Luna | $0.10 | $0.50 |
What the pricing actually means
The API side is easy to model. You know the rate, you can meter your own usage, and Luna is twenty times cheaper than Sol on both input and output, which makes it the sensible pick for routine edits.
The subscription side is harder to plan around. OpenAI labels the Plus limits as estimates, and the ranges are wide: 5 to 45 Astra messages per five hours is a ninefold spread. Actual use depends on task size, model choice and context length. Once you hit a limit you can buy extra ChatGPT credits, and the docs say a typical task may use 2 to 15 credits. That range is useful as a rough guide, but you can't build a team budget on it. This is why pricing value scores 5. The prices are public, but the cost of a month of real Codex use is not.
Plan pricing and limits on AI coding tools have shifted often through 2026, and third-party guides already disagree with each other on Codex figures. Check OpenAI's page before you commit.
Pros & Cons
Pros
- Strong model results from independent reporting: 64.6% on SWE-bench Pro and 88.8% on Terminal-Bench, per CodeAnt AI.
- Published per-token API prices, with a very cheap Luna tier for high-volume, low-stakes work.
- A 1.05M token context window and 128K max output on GPT-6 Astra.
- Broad surface coverage on Plus and above: CLI, IDE, web, desktop, iOS and cloud tasks.
- A $20 Plus plan gets you in, so you can try it without a large commitment.
Cons
- Usage limits inside ChatGPT plans are published as wide estimates, so monthly cost is hard to predict.
- API key users lose cloud features, so the predictable billing path is also the less capable one.
- The multi-step workflow claims rest on OpenAI's own description, with no independent confirmation found.
- No Codex-specific SDK or integration details were listed beyond the general OpenAI API.
- No evidence covered setup or day-to-day ergonomics, so we give no ease-of-use score.
Scores
| Criterion | Score | Basis |
|---|---|---|
| Output quality | 7/10 | Independent SWE-bench Pro and Terminal-Bench results |
| Ease of use | N/A | No evidence found |
| Pricing value | 5/10 | API prices published; plan usage not itemized |
| API integration | 7/10 | Documented OpenAI API; no Codex-specific SDK facts |
| Problem fit | 5/10 | Workflow claims are vendor-only |
| Overall | 6/10 |
Who Is It For
Codex fits developers who already pay for ChatGPT Plus or Pro. You get the agent with no new vendor, no new contract and no new bill. If your team already runs on the OpenAI API, the CLI and IDE extension on an API key give you metered, auditable spend, and Luna keeps the cost of bulk work low.
It fits less well for teams that need a firm monthly number before signing off on a tool. Estimated message ranges and add-on credits make finance teams nervous. It also fits less well if you need proof that autonomous multi-step workflows work before you rely on them, because that proof is missing from the public record for now.
If you mainly want an AI-first editor, Cursor and Windsurf are the obvious alternatives. If your team lives in GitHub, GitHub Copilot is the default to compare against. We haven't scored those tools here, so treat them as alternatives to look at, not as ranked picks.
Verdict
Codex gets a 6 out of 10. The model underneath is strong, with independent benchmark results that put it among the better coding agents, and the API prices are clear. What holds the score down is the packaging. Plan limits are estimates, credits cover the overflow, and cloud features are missing on the one billing path you can predict.
Our recommendation: if you already pay for ChatGPT Plus, turn on OpenAI Codex and use it, because it costs you nothing extra. If you're on the OpenAI API, try the CLI with Luna on a narrow task type and meter what it spends. If you're choosing a coding agent from scratch and need predictable spend or verified workflow automation, compare it against the alternatives before you buy Pro or Business seats. Once OpenAI publishes firmer limits, or independent reports confirm the workflow claims, this score has room to go up.
