A useful free Codex alternative must fit both your workflow and its allowance. GitHub Copilot offers a limited free plan for editor and agent use; Gemini CLI documents free routes whose limits depend on how you sign in; Cursor lists limited free agent and chat use. Claude Code is another coding-agent option, but Claude Pro access should not be confused with a free Claude chat account. We checked the official plan pages on October 25, 2026. This is a comparison of documented access and setup requirements, not a four-product coding benchmark or a guarantee that every account qualifies.
First identify what you need to replace
“Codex alternative” can mean an inline completion tool, an agent that edits a repository, a terminal assistant, or a hosted task runner. Those are different jobs. A free completion allowance may be useful while you type, yet tell you little about the cost of a multi-step repository task.
Write down your minimum requirement before opening pricing pages. For example: “Explain this small Python project, fix one failing test in a practice copy, and show the diff.” Then add where you want to work: your current editor, a new editor, or a terminal.
Also check whether you need a replacement at all. If the problem is a failed Codex command or an unclear usage reset, the troubleshooting guide or usage-limit guide may resolve it without migrating your workflow. For cost structure, see Codex pricing.
| Candidate | A reason to investigate it | Free-access boundary to verify |
|---|---|---|
| GitHub Copilot | You want assistance inside a supported editor or through its CLI | Free completions and limited chat/agent usage; paid features remain separate |
| Gemini CLI | You prefer a terminal and can use a supported Google authentication route | Daily allowance and model access depend on that route |
| Cursor | You want to try an editor-centered workflow | Free agent and chat use is limited; do not assume paid features are included |
| Claude Code | You want to compare another repository coding agent | Pro includes access; this does not establish a free Claude Code entitlement |
These are candidates to evaluate against your requirements, not a ranked performance table. We did not run the same coding task through all four products.
GitHub Copilot: inspect the free feature row
GitHub’s current Copilot plans page lists a $0 Free plan with 2,000 monthly completions, limited chat and agent usage, and CLI access. It separates paid cloud-agent and code-review features from the free tier. A completion allowance is not a promise of unlimited agent work.
The practical question is whether the included features cover your task in your editor. Start with the official free-plan entry, confirm the selected plan in your account, and use a non-sensitive practice project. If the workflow asks you to upgrade, record which feature caused that request instead of assuming that every part of Copilot requires payment.
For a fair personal comparison, keep the starting files and expected result identical. Record how much work the account allows you to complete, but do not turn one account’s observed allowance into a universal quota.
Gemini CLI: the login method changes the allowance
The Gemini CLI quota page documents up to 1,000 model requests per user per day for its individual Google-account route. It lists a different unpaid API-key route with up to 250 daily requests and Flash-model access. These are documented ceilings, not a promise of that many completed tasks; one task can require several model requests.
Check eligibility, authentication type, and the limits shown to your account. A workplace Google identity and a personal identity can have different requirements. The documentation also separates paid API or Vertex usage from the free routes. Do not enable billing simply to finish a free-tier experiment without deciding that you want those charges.
If a terminal is unfamiliar, include setup effort in your decision. You need to recognize the working folder, read a command’s output, and inspect changed files. Free model access does not remove those responsibilities.
Cursor: use the current pricing page
The live Cursor pricing page lists a Free tier with limited agent and chat requests. It shows additional paid capabilities separately. At the time of this check, an older help page used different plan names from the main pricing page, so we used the live pricing page for this comparison rather than repeating the older names.
Check the exact plan and model availability offered in your account before choosing an upgrade. A free editor download is not proof that every model or hosted feature is free. Keep third-party model charges separate from a displayed subscription price when the plan describes them separately.
Cursor also needs a workflow decision: are you evaluating Cursor’s own agent, or using the Codex extension inside Cursor? Installing Codex in an editor does not turn it into that editor’s own agent or transfer one provider’s subscription to another. Record which tool you actually selected during a test.
Claude Code: separate coding access from free chat
Anthropic’s Pro plan documentation lists Claude Code access and a US monthly price of $20. It also says that Console API usage is separate. The existence of a free Claude chat plan therefore should not be treated as proof of free Claude Code use through your intended route.
We verified the installed Claude Code CLI version during this research, but an official-account test failed authentication. We did not obtain a valid comparison run and do not infer a performance result from that failure. Login state is an access issue, not evidence that one model writes better code.
If you already pay for a qualifying plan, check what that account includes before adding another subscription. If you need a strictly $0 trial, establish the available entitlement first. Do not base the choice on an old promotional offer or an unverified claim of permanent free access.
Make your own small comparison reproducible
Use a project without private data, external deployment, or payment actions. The downloadable fixture in our first-task guide is one starting point. Use a fresh copy for each tool so that one tool’s correction does not become the next tool’s starting answer.
Before the first request, record five things: the client version, the selected agent and model, the account route, the starting files, and the test that defines success. Ask each candidate for the same bounded change. Inspect the diff and run that test yourself.
A useful scorecard is practical rather than theatrical:
- Could you reach the intended feature without enabling paid usage?
- Did the tool receive the right project and respect the requested scope?
- Did the independent test pass, and was the changed code understandable?
- Could you identify the allowance used and any upgrade requirement?
- Could you resume, export, or stop the work without losing your changes?
If you time the run, include setup and correction time separately. A fast initial reply that needs several repairs may not save you time. A failed login should be recorded as a setup failure, not scored as coding quality.
Keep migration small and reversible
Begin with one practice task and one tool. Preserve your original repository and commit or copy the starting state. Keep project instructions short enough to review, and adapt them to the destination tool’s documented instruction mechanism. Do not assume every tool automatically reads the same filenames or supports the same plugins.
Move only after the candidate satisfies the requirement that made you look for an alternative. If the need was occasional inline help, a limited free plan may be sufficient. If the need was sustained multi-step agent work, compare the actual allowance and paid path before depending on the service.
The verified result here is a set of current official access rules and a repeatable way to check your own fit. Free allowances can change, eligibility can vary, and this article does not claim that we exhausted or benchmarked any provider’s free tier.