OpenAI · Alternatives
Codex alternatives.
People leave Codex for four reasons, and they lead to four different shortlists: the billing unit is opaque, the personal plans train on your work by default, the terminal is the wrong shape, or the bill is simply too large. This page is organised by which of those is yours. Every price below was read off the vendor's own page on the date printed beside it.
Why people look.
Worth being precise about, because a generic "top 10 AI coding tools" list serves none of these four and the shortlists barely overlap.
You cannot see what you are spending
Codex meters in credits, and OpenAI publishes no price for one. You can work out what a credit buys — we do, on the pricing page — but a finance conversation that starts with a derivation is a harder conversation than one that starts with a rate. Tools billed straight in tokens remove the step.
The personal tiers train on your work
On Plus and Pro, conversations may be used to improve models unless you turn training off yourself. That is an opt-out, and an opt-out is not a policy your legal team can point at. Where the requirement is contractual, the answer is either a business tier or a tool you host.
The shape is wrong
A terminal agent that edits files is not what everyone wants. Some people want the agent inside the editor they already have open, with a file tree and a visual diff to review against. That is a different product category, not a cheaper tier of this one.
The bill is too big
OpenAI's own rate card puts Codex at a per-developer figure that dwarfs the subscription price. Two tools below are free at the point of use, and one of them documents a daily quota generous enough to work against, which changes the calculation entirely for solo developers and students.
If you only read one line.
Claude Code
If you like the shape and want different models underneath. Same terminal-agent idea from the other frontier lab, and the one tool on this list we have actually run ourselves — 27 capability tests, published with transcripts.
Gemini CLI
If the bill is the problem. A published daily quota on a personal Google account makes it the only entry here you can run all day without paying anything.
OpenHands
If the blocker is policy rather than price. Permissive licence, runs on your own machine, and the hosted version is optional rather than the product.
Every alternative we checked.
Ordered by how many of the four reasons above each tool answers, not by any score — we run no benchmark that ranks these against one another, and a number here would be invented. Where a full head-to-head against Codex exists on this site, the row links to it.
The direct counterpart: a terminal coding agent from the other frontier lab.
If the shape of Codex suits you and the model or the billing does not, this is the one-for-one swap. It runs in the terminal, in VS Code and JetBrains, in a desktop app and the browser on one engine, and it is billed in dollars rather than in a proprietary unit — a subscription, a seat or a per-token key, with the rates published on the vendor's own page. It is also the tool we have tested most heavily, so the claims here are checkable against our own transcripts rather than against a marketing page.
There is no free tier at all, which is exactly where Codex is stronger: no free plan, no trial, and the cheapest way in is a paid subscription. Its allowance is also shared with Claude chat, so a heavy chat habit shortens the terminal session you were counting on.
Best for developers who want the same workflow with different models and legible billing.
An editor with the agent built in, rather than an agent that drives your editor.
The swap for people whose complaint is the terminal itself. The agent works against a file tree and a visual diff instead of a transcript, which makes supervising a long autonomous run considerably easier. Our own head-to-head found the harness difference is real in both directions: Codex scores higher on a neutral coding-agent index, and Cursor finishes faster and far cheaper per task, with the gap surviving even when both run the same model.
You are adopting a whole editor, not adding a tool. If your work lives on remote machines, in a terminal multiplexer, or in an IDE you are not giving up, that migration cost is real and rarely appears in the comparisons.
Best for developers who want the agent where they already read code.
A terminal agent with a published free quota on a personal account.
The only tool here you can run all day without paying. Google documents the free allowance rather than leaving it vague, which is what makes it plannable: sign in with a personal Google account and the repository states 60 requests a minute and 1,000 requests a day. Apache-2.0 licensed, so the client is auditable — and unlike Codex the number defining the free tier is actually published.
A daily request cap is not a daily token cap, and agentic work spends requests quickly — one long task can be dozens. Treat the quota as generous for exploration and unreliable for anything with a deadline attached.
Best for students, solo developers and anyone evaluating before spending.
MIT-licensed agent that runs on your own machine, with the hosted version optional.
The answer when the blocker is policy rather than price — including the training question, which no consumer tier of a hosted tool resolves. It is open source under a permissive licence, the local tier is the real product rather than a crippled demo, and the enterprise option deploys inside your own VPC. Its hosted model provider is offered at cost with no markup, which is an unusually legible arrangement in this market.
Self-hosting moves the work to you: sandboxing, updates and model access all become your problem, and the polish gap against a commercial agent is real. Choose it for the licence and the deployment, not for a smoother first hour.
Best for teams whose code cannot leave their infrastructure.
Open-source agent that lives in VS Code and bills nothing of its own.
A middle path between the editor and the terminal, and the cleanest answer to the credit problem: the extension is open source and free for individuals, and you attach whichever provider key you already hold. Its own framing states the model plainly — pay for inference, not for a subscription — which means every line of your bill is denominated in something you can look up.
No allowance means no ceiling. Per-token billing on an agent that re-sends context every turn punishes long sessions, and there is no plan absorbing the spike the way a subscription does.
Best for developers who already pay for API access and want no second bill.
Provider-agnostic terminal agent that signs in with subscriptions you already have.
Its distinguishing feature is refusing to pick a model vendor: it connects to a large catalogue of providers including local models, and it can authenticate against a subscription you already pay for rather than demanding a new key. For anyone whose objection to Codex is being tied to one lab's account system, that is the argument.
Breadth costs depth. A tool that supports everything is tuned for nothing, and the model-specific behaviours a first-party agent relies on — caching defaults, reasoning budgets, tool conventions — are yours to configure.
Best for developers who switch models often or run local ones.
The original terminal pair-programmer, Apache-2.0, deliberately small.
It predates most of this list and still does one thing well: edit files in a git repository from the command line, with commits as the unit of work. Apache-2.0 licensed, connects to almost any provider including local models, and has no service behind it that can change its terms on you — which is the structural version of the training-policy complaint.
It is a pair-programmer rather than an autonomous agent — less initiative, fewer integrations, no delegated cloud runs. Also worth checking before committing: the model recommendations on its own front page are visibly behind the current generation.
Best for people who want a small, auditable tool and tight git discipline.
Completion first, agent second — and the one your organisation probably already buys.
Included because it is the incumbent, not because it is the same product. It has a genuine free tier, the lowest paid entry price on this list, and it is already approved in most enterprises, which removes the procurement step that kills evaluations of everything else. For pull-request review specifically it competes directly with the Codex feature that an API key does not include.
The paid tiers meter agent work with monthly credit allowances rather than selling it flat, so you have traded one proprietary unit for another. Teams frequently end up paying for Copilot and an agent, which is worth planning for rather than discovering.
Best where procurement has already said yes and completion is most of the value.
Pick by what you're doing.
Finance wants a number, not a derivation
Somebody has to forecast next quarter's spend and credits do not convert cleanly into a line item.
Cline
Pay-per-inference bills in the same unit your provider publishes, so the forecast is a multiplication rather than an argument. Check the direction first with our subscription-versus-API calculator — for steady daily use a flat fee still usually wins on price, just not on legibility.
Your code cannot leave the building
A policy, a client contract or a regulator says source stays inside your infrastructure.
OpenHands
Permissive licence, runs locally, and deploys into your own VPC on the enterprise tier. This is the only requirement on the page that rules out most of the list outright — no subscription tier of a hosted tool solves it, and the training opt-out on a personal plan is not the same thing as data never leaving.
You want the agent in your editor
The terminal is the part you dislike, not the agent.
Cursor
A file tree and a visual diff change how supervision feels on long autonomous runs — and our head-to-head found the cheaper, faster side of that trade is real rather than marketing. If you would rather not change editors at all, the VS Code route is the softer version of the same move.
You are learning and cannot spend
A student, a side project, or an evaluation with no budget attached.
Gemini CLI
A published daily allowance on a personal account beats an unpublished one. Codex's free tier is real and does not expire, but OpenAI does not say how much of it you get — which makes it fine for trying and impossible to plan against.
You want the same thing with better models
Nothing about the workflow is wrong; the output quality on long refactors is.
Claude Code
The like-for-like swap, and the one where our own comparison has numbers rather than impressions. Read the head-to-head before moving: the two tie on some neutral benchmarks and diverge sharply on others, so which one wins depends on the work you actually do.
When to stay put.
The honest counterweight. In at least three common cases switching costs more than the difference it saves.
Frequently asked.
What people search next.