CLI agent
Codex
OpenAI's coding agent shipping as an Apache-2.0 CLI, IDE extension, desktop app, and cloud tasks started from web, GitHub, or Slack.
Compare only what the shape makes comparable.
The CLI is Apache-2.0 and "runs locally on your computer". The config reference supports custom model providers—"Built-in provider IDs (openai, ollama, and lmstudio) are reserved and cannot be overridden"—plus model_providers.<id>.base_url for an arbitrary API base URL, so the agent can be pointed at locally served models. The Codex cloud service itself is OpenAI-hosted.
Named: gpt-5.6-sol, gpt-5.6-terra, gpt-5.6-luna, gpt-5.3-codex-spark, gpt-5.5, gpt-5.4, gpt-5.4-mini
- Free$0 per user / month
- Plus$20 per user / month
- Pro$100 per user / month
- Business$25 per user / month
- Enterprise & Educontact
- API key (pay-as-you-go)contact
- MCP support
- Plugins / custom tools
- Subagents / parallel agents
- Hooks / lifecycle events
- Persistent memory
- Background execution
- Sandboxing / permissions
- Git / PR integration
gpt-5.4 and gpt-5.4-mini are "Being retired August 31, 2026". src↗
PR review inside the desktop app requires the GitHub CLI (gh) installed and authenticated via gh auth login; without it, PR details may not load in the sidebar or review pane. src↗
Built-in model provider IDs (openai, ollama, lmstudio) are reserved and cannot be overridden in config. src↗
- first_release
No OpenAI launch-announcement page was fetched with a dated Codex CLI release. 2025-04 is taken from the openai/codex repository creation date (2025-04-13T05:37:54Z via the GitHub API). Note that OpenAI previously used the name "Codex" for a 2021 code model, which is a different product from this agent.
- pricing (Pro 20x)
The pricing page states "From $100/month" for the 5x tier and "5x or 20x more Codex usage than Plus" but does not publish a separate dollar figure for the 20x tier.
- interfaces
Codex also ships a desktop app (codex app / ChatGPT desktop app on Windows, Linux, macOS), which has no matching value in the allowed interfaces enum.
- model_support
Recorded model-agnostic because the models page states Codex supports "any model and provider that supports either the Chat Completions or Responses APIs" and the config reference reserves ollama and lmstudio provider IDs. Every model named by OpenAI as recommended or available is an OpenAI model, so a stricter reading of "multi-vendor" vs "model-agnostic" is arguable.
- capabilities.persistent_memory
The docs feature index lists "Memories" under customization, but the dedicated Memories page was not fetched; the recorded evidence is the AGENTS.md layered-instructions mechanism.
What people say about Codex
58 opinions · 2025-11-19 to 2026-08-19 · medium confidence
Counts are conservative estimates of distinct opinion-holders across the threads read, each tagged by overall polarity toward Codex; not a logged per-comment tally. Score is round(100*(26-17)/58) = 16.
The most comparative sample of the three: a large share of Codex opinion is expressed inside Claude Code threads by people explaining why they switched, which inflates the 'generous limits' and 'drop-in replacement' themes and undercounts complaints from people who only ever used Codex. The GitHub slice is concentrated in one incident (issue #28879, the June 2026 rate-limit burn) and so over-weights a single bad month. Only one blog in the sample. Praise skews toward price/limits rather than output quality, partly because the comparison is usually being made against Anthropic's pricing rather than on the merits.
- hackernews+41
- reddit+29
- github-90
GitHub reads 131 points lower than the highest platform here. Issue trackers collect defects only, so that number measures activity more than opinion. Why we publish it anyway.
Codex is the only one of the three whose sentiment improved over the window, but not monotonically. Through late 2025 and early 2026 it was mostly praised as the generous-limits alternative, with 'slow but never cut off' as the accepted trade. The Codex app launch (Feb 2026) plus doubled limits pushed it further up, and it is the default recommendation in the August 2026 Reddit budget threads. The clear negative spike is June 2026: the rate-limit cost jump (issue #28879, 210 comments) and the July context-window reduction from 372k to 272k both read as the subsidy tightening, and produced the first real wave of switch-away-from-Codex posts, including one user saying Codex now felt more aggressive than Claude. By August 2026 recommendation sentiment had largely recovered. The persistent, non-cyclical weakness is UI and visual design work, which Codex users concede without argument across the entire window. The platforms also disagree sharply: Codex scores +41 on Hacker News and -90 on GitHub, a 131-point split that is mostly an artifact of the GitHub sample being drawn from one incident thread; people open issues when something breaks, not when it works.
Praised for
- Generous rate limits for the price14×
- Thorough context gathering, finds real fixes7×
- Drop-in replacement when Claude limits hit6×
- Async parallel cloud tasks5×
- Codex app UX is thoughtful4×
- Bundled value of the whole ChatGPT plan3×
Complained about
- June 2026 rate-limit burn spike10×
- Weak at UI and visual design5×
- Slow compared to alternatives4×
- Context window cut from 372k to 272k4×
- Agentic scaffolding thinner than Claude Code3×
- Banked limit resets quietly expiring2×
the true magical moment was codex pro letting me run swarms of agents day in day out without any worries about rate limits it truly felt unlimited
ChatGPT 5 agentic functionality just sucks and wont chain workflows anywhere close to as long. The scaffolding is literally missing from Codex.
Because Codex effectively rate limits you by being so slow.
The rate limit for my $20 OpenAI / Codex account feels 10x larger than the $20 claude account.
This week I’ve been hitting CC rate limits, and I switched to Codex with virtually no disruption to my workflow. It’s not good for Anthropic that I was able to do that!
I don't think it's as good at UI, but man is it thorough and able to gather the right context to find solutions.
it's been frustrating how bad it is at UI.
I really like the codex app. just some really thoughtful user experience decisions
Codex is the best at the difficult and intelligent stuff
I'm on the Pro plan and my quota suddenly vanishes in less than 5 hours now. Even with top-ups, it doesn't survive a single intense work session. This rate-limit spike completely breaks the entire workflow.
I had to switch to Claude; now Codex seems like a more aggressive Claude
Before this change, I was never able to exhaust my 5-hour allowance, even during heavy Codex usage sessions involving large codebases and long-running tasks.
Its async cloud mode lets you queue several isolated tasks
I was going to explore paying for Codex since OpenAI seems to be a bit more generous with rate limits, but I'm now not sure - for the stuff I do limited context size would be a dealbreaker
I switched from Claude to Codex recently and I like it. Only one thing I missing from Claude is Design.
Codex - new OpenAI models are really good, you also get ChatGPT plus in the web, image generation, research and whatnot. As a full self contained productivity package this is the best deal.
Right now Codex is the way to go. You have Sol which is a frontier model if you need it and Luna is probably the best model if you want a balance between intelligence and cost.
same split here honestly, opus plans better but 5.2 just executes. only thing is i keep runing into the codex weekly limit mid project and end up back on claude anyway