OpenAI Codex — a terminal coding agent from OpenAI. Reads your repo, edits files, runs commands, works in a loop.
Alternative to: Cursor Pro $20/mo, Pro+ $60/mo, Ultra $200/mo, Teams Standard $40/user/mo. GitHub Copilot Pro $10/mo, Pro+ $39/mo, Max $100/mo, Business $19/seat/mo. Claude Code, included in Claude Pro at $20/mo monthly or $17/mo billed annually, and Max from $100/mo.
Cost to run: not free. The client is free; running it needs a ChatGPT plan — Go $8/mo, Plus $20/mo, Pro $100 or $200/mo, Business $20/user/mo annual or $25 monthly — or API credits. There is a genuinely free local path against Ollama or LM Studio, documented by OpenAI, which costs hardware instead.
Licence: Apache-2.0. Free for commercial use.
Runs on: macOS, Linux, Windows. Terminal, VS Code extension, and desktop app.
Requires: an OpenAI account and a paid plan or API credits, unless you point it at a local model.
Traction: 120,183 stars · 18,361 forks · 10,021 commits · 474 committers in the last year.
Liveness: very live — 1,217 commits in 30 days; stable rust-v0.151.0 on 2026-08-29, with alpha prereleases shipping most days.
Maturity: rust-v0.151.0 stable. Alpha tags cut daily.
Links: repo · pricing and rate limits · config docs
Figures as of 2026-08-31.
What it does
It is a coding agent that lives in your terminal. You give it a task, it reads the repository, plans, edits files, runs commands, and iterates. It also ships as a VS Code extension and a desktop app, and the same account works across all three.
It is written largely in Rust, released under Apache-2.0, and developed in the open at a pace that is genuinely unusual: 1,217 commits in the last thirty days from a contributor list 474 people long.
What it replaces, and what that costs
This is the entry's whole point, so it gets stated flatly: "free and open source" describes the client, not the cost of using it.
The competition is priced per month and priced honestly. Cursor is $20 for Pro, $60 for Pro+, $200 for Ultra, $40 per user for Teams Standard. GitHub Copilot is $10 for Pro, $39 for Pro+, $100 for Max, $19 per seat for Business. Claude Code comes with Claude Pro at $20/mo billed monthly or $17/mo billed annually, and with Max from $100/mo.
Codex is on every ChatGPT tier including Free — OpenAI's help centre says "Codex is included across ChatGPT plans, including Free and Go" — but the feature table grades Free and Go as "Limited," Plus as "Yes," and Pro as "Expanded." So the practical price of using Codex seriously is a ChatGPT Plus subscription at $20/mo, the same number Cursor and Claude Pro charge. Annual billing is not offered on Plus or Pro; OpenAI states so directly.
Nobody in this category publishes an absolute usage number, and Codex is no exception. OpenAI does publish per-plan Codex rate limits, which is more than Cursor does — but as ten-times-wide bands. GPT-5.6 Sol on Plus: "10–100" local messages per five-hour window. On Pro 20x: "200–2,000." Cursor expresses its tiers as multiples of an undisclosed base ("3x Pro limits," "20x Pro limits"). Copilot has moved to dollar-denominated AI Credits — $15/mo on Pro, $70 on Pro+, $200 on Max — which is the most legible scheme of the three. Codex Business and Enterprise also run on credits, and OpenAI's rate card does not publish a dollar value per credit. It does say Codex "costs ~$100-$200/developer per month" on average.
On API pricing, two OpenAI pages disagree, and this entry reports both rather than picking. openai.com/api/pricing lists GPT-5.6 Sol at $5.00 in / $30.00 out per MTok; developers.openai.com lists the same model at $2.00 / $10.00. The likely reconciliation is on the developers page itself, which notes GPT-5.6 Sol has "promotional pricing available at least through November 21, 2026" — but OpenAI does not label either page as the promotional one, so both numbers are printed here as found.
The free path is real and OpenAI documents it. Codex accepts arbitrary OpenAI-compatible base URLs, and `ollama` and `lmstudio` are reserved built-in provider IDs, not community workarounds. You can run the whole agent against a local model for nothing. What that costs instead is hardware, and the frontier models the entire rate-limit table is about.
Is it still alive?
This is the least interesting liveness question in the batch, and the answer is yes, overwhelmingly. 1,217 commits in 30 days, 2,977 in 90, 474 committers in the past year, 10,021 commits total since the first one on 2025-04-16. Stable rust-v0.151.0 shipped 2026-08-29; alpha tags rust-v0.152.0-alpha.4 through alpha.6 all landed within 36 hours of this being written.
The issue tracker is past #41,700, and the open issues read like a shipping product's: high RAM usage and PC crashes on launch, Deep Research disappearing on Windows, Android remote voice failing after working on 23 August, deleted conversations stuck in Recents on macOS, VS Code chat blocks losing syntax highlighting.
The more useful liveness question for a vendor repo is not whether the code moves but whether the terms under it hold still. They do not.
GitHub's signed-out "Issue creation is restricted" banner appears here too, on a repo taking dozens of outside issues a day. Disregard it.
What people say
There is no single large Hacker News thread for Codex, which is surprising for a 120,000-star repository and is worth saying rather than padding around. The closest is a February 2026 "Tell HN: OpenAI's Codex CLI is currently free to use" at 6 points, and an ANSI escape-code injection report at 2 points. Reddit is blocked from this column's research sandbox, so no Reddit vote counts are claimed.
What exists instead of discourse is OpenAI's own changelog, and it is more informative than a comment thread would have been. From the ChatGPT release notes, 9 April 2026: "As the temporary Codex promotion on Plus ends, we're rebalancing Plus usage to support more sessions across the week." That is OpenAI, in its own words, confirming it reduced what a $20 subscriber gets. The same entry defines Pro at $100/month with "up to 10x more Codex usage than Plus."
11 June 2026: "Eligible Plus and Pro users now have rate-limit reset banking, including one free reset at launch" — a concession in the other direction.
And today, 31 August 2026, GPT-5.4 and GPT-5.4 mini stop being available in Codex, per OpenAI's own plan documentation. If you pinned that model, this is the day it stops.
Community threads speculating about exact Codex limits are user testing, not documentation, and one of the most-cited says so itself: "Everything above is based on practical testing and approximation, not on official OpenAI documentation." This entry does not cite those as fact.
The catch
The economics are identical to every competitor. The client is free; the tokens are not. That is true of Codex, and equally true of Cline, Aider, OpenHands and Goose. Open-sourcing the client changes who can audit and fork the harness. It does not change the bill.
Codex is not the only open-source option, and the alternatives are not fringe. Cline is Apache-2.0 with 63,900 stars. Aider is Apache-2.0 with 48,600 stars. OpenHands is MIT with 85,700 stars. Goose is Apache-2.0 with 53,600 stars — and it moved, from Block to the Agentic AI Foundation at the Linux Foundation, so `block/goose` is no longer the address. All four run against whatever provider you point them at.
Two corrections for anyone comparing: Continue.dev is archived. Its repository banner reads "no longer actively maintained and is read-only for all users," and its 35,700 stars are a vanity number now. And there is an open issue on Goose arguing its Acceptable Use Policy may impose Apache-2.0-incompatible restrictions, so "cleanly open source" deserves a hedge there.
Windsurf has effectively disappeared as a comparable. windsurf.com/pricing now redirects to devin.ai/pricing under Cognition. This entry quotes no Windsurf price, because there is no vendor page left to quote.
The vendor sets the limits and can move them, which it has. This is the structural difference between Codex and everything else in this directory. A self-hosted tool's capabilities are whatever you deployed. Codex's capabilities on your machine are whatever OpenAI's rate limits say this week, and April's release note is the receipt.
The published limits are bands, not numbers. "10–100 messages" is a ten-fold spread. Business gets the same per-window band as Plus at the same or higher price.
Credits
Maintained by OpenAI. The last year's commits are led by jif-oai (1,013), Michael Bolin (695), Eric Traut (568), pakrym-oai (548), jif (524) and Ahmed Ibrahim (518), among 474 committers in total. @OpenAI on X.
Prices and limits verified against learn.chatgpt.com/docs/pricing, help.openai.com articles 11369540, 6950777 and 9793128, chatgpt.com/pricing, the Codex rate card, openai.com/api/pricing, developers.openai.com/api/docs/pricing, cursor.com/help/account-and-billing/pricing, docs.github.com Copilot plans, claude.com/pricing and code.claude.com/docs. Limit-change history from OpenAI's own ChatGPT release notes. Competitor licences read from each project's LICENSE file. Repository figures from a blobless clone and ungh.cc, both read 2026-08-31.