1. Home
2. Guides
3. Anthropic

Claude Code vs Codex vs Gemini CLI: which plan you need and which models you get

One of the three stopped serving individual subscribers in June, but its own docs still advertise the free tier it closed.

By Think Facility · September 23, 2026 · Claude Code Codex Gemini CLI usage limits

I set out to compare three terminal coding agents on price, and ended up comparing two, because Gemini CLI stopped serving Google's individual subscribers on June 18, 2026. If you pay for Google AI Pro or Ultra, or used Gemini CLI on the free tier, Google has moved you to a different tool called Antigravity CLI.

So this piece covers four products. All of them are coding agents, and the differences that matter are what you pay before one will start and which models it hands you once it does (plus the clock your usage runs against, which turned out to be the most interesting part). I read each vendor's own docs and help pages on September 23, 2026.

The short version

From Anthropic's, OpenAI's and Google's own docs and help centers, read September 23, 2026. Details and links below.

Claude Code needs a paid plan

There's no free way into Claude Code. Anthropic's setup page is direct about it: you need a Pro, Max, Team, Enterprise or Console account, and "the free claude.ai plan does not include Claude Code access." Pro is $20 a month in the US. Max comes in two tiers, $100 for five times Pro's per-session allowance and $200 for twenty times.

Since version 2.1.280 the default model is Opus 5.5 on every one of those plans, according to the model configuration docs. Before that, Pro users started on Sonnet 5. Fable 5.1 is never the default. You pick it with /model fable, and on Pro it doesn't come out of your plan at all: Anthropic's Fable help page says it runs on pay-as-you-go usage credits from the first message, while Max can spend up to half its weekly limit on it.

The limit itself works on two clocks. Your session allowance resets every five hours, and a weekly limit across all models resets at a fixed time Anthropic assigns to your account (you'll find it under Settings, then Usage). Anthropic raised the five-hour limits on Pro and Max, and on Team, when it released Opus 5.5 on September 22.

It's also one pool. Anthropic's Claude Code help page says usage is "shared across Claude and Claude Code", so an afternoon of chatting in the Claude app comes out of the same allowance your terminal session needs. And if ANTHROPIC_API_KEY is set anywhere in your environment, the same page says, Claude Code bills that key at API rates instead of using your subscription.

Codex has a free tier, with a catch

Codex is the only one of the three with a free door, though it's narrow. OpenAI's Codex pricing page gives the Free plan and the $8 Go plan GPT-6 Luna, which OpenAI calls its most efficient model, at standard speed in the ChatGPT desktop app, "subject to rollout". That's all. The CLI and the IDE extension start at Plus, $20 a month, which adds GPT-6 Sol and GPT-6 Astra. Pro starts at $100 and buys five or twenty times Plus's usage.

Codex meters usage differently from Claude Code. OpenAI publishes an estimated number of local messages per five-hour window for each model, and the spread between models is enormous. On Plus it's 5 to 45 Astra messages, 15 to 150 on Sol, and 350 to 3,000 on Luna, and the table's footer adds that "weekly limits may also apply". OpenAI calls every one of these an estimate.

Like Claude, it's a shared pool: the pricing page says ChatGPT Work "uses the same pricing, credits, and usage limits as Codex." And there's a detail about the cloud side I didn't expect. Codex cloud tasks need a ChatGPT sign-in, and on ChatGPT plans they run on GPT-5.6 Sol, the previous generation. The models page says you can't change that. The Codex models guide goes through each model and plan in more detail.

Gemini CLI moved most of its users out

This is the section I had to rewrite. Google announced on May 19, 2026, in a developer blog post, that on June 18 Gemini CLI would "stop serving requests for Google AI Pro and Ultra, as well as those using it free of charge". The Gemini CLI plans page now carries a banner saying Gemini CLI "was replaced by Antigravity CLI on June 18th, 2026" for unpaid and Google One users.

Two groups kept it. Organizations with a Gemini Code Assist Standard or Enterprise license still get Gemini CLI, and so does anyone running it on a paid Gemini API key. Google's Code Assist quotas page, updated September 22, gives Standard 1,500 requests per user per day and Enterprise 2,000. One prompt can turn into several requests, and every model draws on the same count.

The project's own docs haven't caught up. The quota page in the Gemini CLI repository still lists 1,000 free requests a day for individuals, which is the tier that ended in June. The code's behind the models, too. The stable release, v0.60.0 from September 15, lists Gemini 3.5 Flash and 3.1 Pro as its newest models, and support for Gemini 3.8 Flash only merged on September 22, three weeks after Google released that model.

Antigravity CLI is where Google's individual users went

Antigravity CLI runs as agy, is built in Go, and shares its agent with the Antigravity 2.0 desktop app, according to Google's announcement. It imports Gemini CLI extensions as plugins and still reads your GEMINI.md files, per its migration guide.

Antigravity's models page offers Gemini 3.8 Flash, 3.7 Flash, 3.6 Flash and 3.1 Pro on every plan. It surprised me by also listing Claude Sonnet 4.6, Claude Opus 4.6 and OpenAI's open-weight GPT-OSS-120b, ticked for everyone from the free tier up (only Enterprise is left out). Its plans page doesn't quite agree, listing "access to third-party models" as an Ultra benefit. I couldn't settle which one is current.

Limits follow the plan. On Google AI Pro and Ultra, quota refreshes every five hours until a weekly cap; everyone else gets a quota that refreshes weekly. Google says the limits are "correlated with the amount of work done by the agent", so a quick task costs less than a long one, much as Anthropic and OpenAI say about theirs.

Where the limits come from

Put side by side, the three companies meter three different things. Anthropic sells time windows (five hours and a week) and doesn't publish message counts. OpenAI sells the same kind of window but prints a range per model, which makes it the only one where you can see what switching model buys you: Luna gets you roughly twenty times as many messages as Sol on the same plan. Gemini CLI counts requests per day, a hard number you can plan a week around.

If you're choosing on price alone, Codex's Free plan and Antigravity's free tier are the only two you can try without paying anything, and Claude Code's cheapest door is Pro at $20. How many hours you'll get out of each depends on the model you pick far more than on the plan, and there's more on the Claude side in what Max 5x and 20x actually get you.

Gemini CLI's own quota page on GitHub still lists the free individual tier that closed on June 18, 2026.