All three are the same shape: a terminal agent that reads files, runs commands and loops until it stops. They differ on two things that matter — what you pay and how much context you get.
| Harness | Maker | Entry price | Context | Free tier |
|---|---|---|---|---|
| Claude Code | Anthropic | Included in paid Claude plans; Pro is $20 a month | Up to 1M, varies by model | No |
| Codex | OpenAI | Included in every ChatGPT plan, Free included | 1,050,000 | Yes |
| Gemini CLI | Free with a personal Google account | 1M on Gemini 3 | 1,000 requests a day |
Claude Code is the subscription one
Anthropic includes Claude Code in every paid plan and shares the plan's usage limits with it, so a long terminal session spends the same allowance as a long chat. Pro is $20 a month billed monthly, $17 on the annual plan. Max starts at $100 and doubles to $200 for 20 times the usage.
Codex is the bundled one
OpenAI puts Codex inside every ChatGPT plan, including Free, so the marginal cost of trying it is zero. The models it runs list a 1,050,000-token context and a 128,000-token maximum output.
Gemini CLI is the free one
Google's CLI gives 60 requests a minute and 1,000 requests a day with a personal Google account, mixing Flash and Pro models. An unpaid API key gets 250 requests a day on Flash only. For a solo developer, that is a working agent for nothing.
What the table cannot tell you
These are three different bets on the same loop. Claude Code and Codex assume you already pay for the assistant and want the terminal included. Gemini CLI assumes you will not pay, and prices the ceiling accordingly.
The way to choose is to run the same task in all three on your own repository. The harness, not the model, decides how many attempts it takes.