Is Gemini CLI Free?

The Short Answer: Free, But the Fine Print Keeps Moving

Yes — Gemini CLI is free. But "free" now depends on two moving targets: how you sign in, and when you last checked the rules.

When Google launched Gemini CLI on June 25, 2025, the deal was remarkably generous. According to the official launch announcement, the open-source (Apache 2.0) tool offered 60 model requests per minute and 1,000 requests per day at no charge — a tier Google called "the industry's largest allowance." Setup required nothing more than a personal Google account, and the whole thing took under a minute.

Then the fine print started moving. On March 25, 2026, a policy change announced on GitHub restricted free-tier users to Gemini Flash models only, pushing Pro models behind paid subscriptions. The community response was blunt: 1,040 downvotes against 137 upvotes, with the top comment advising users to "Switch to Codex." Users also reported free Flash quotas dropping to roughly 150–250 requests per day — figures that conflict with the official documentation and remain unconfirmed.

The other catch is that your free experience depends entirely on your authentication method:

• Personal Google account login: 1,000 requests/day with full model access, per the official quota documentation
• Unpaid Gemini API key: only 250 requests/day, Flash models only — a tier one analysis describes as a "downgrade"
• Vertex AI Express Mode: free for 90 days before billing kicks in

So the honest answer is: free, yes — but the boundaries keep shifting, and agentic workflows burn through quota faster than most people expect. Third-party analysis notes that tool loops consume multiple requests per task, and sub-agent calls count toward your quota even when you didn't explicitly select the model that made them.

That shifting-ground problem is exactly why verification matters in the broader agent ecosystem too. If you're extending Gemini CLI with community-built agent skills — instruction packages that change how the tool behaves — you want the same discipline applied there. SkillzBench is building a free, independent directory that pins every skill to a verified commit, security-scans it, and benchmarks it in a sandbox, precisely because published research found roughly 26% of skills carry at least one vulnerability. Free tools are worth having; verified free tools are worth trusting.

Your Free Tier Is Determined by Your Login, Not Google's Generosity

Here's the counterintuitive part: your Gemini CLI free tier isn't one thing — it's determined entirely by how you log in, and some login methods are dramatically worse than others.

Sign in with a personal Google account and you get what Google's launch announcement billed as "the industry's largest allowance": 60 requests per minute and 1,000 requests per day, with full Gemini 3 model access. But authenticate with an unpaid Gemini API key instead, and you drop to just 250 requests per day on Flash models only — a setup one third-party analysis bluntly calls a "downgrade."

The quota table, in short:

• Personal Google account: 1,000 requests/day with Gemini 3 access
• Unpaid Gemini API key: 250 requests/day, Flash models only
• Google AI Pro subscription: 1,500 requests/day
• Google AI Ultra subscription: 2,000 requests/day

One trap catches business users especially: if your company pays for Gemini through Google Workspace, that subscription doesn't apply to Gemini CLI at all. Google's quota documentation states Workspace plans "currently apply only to the use of Gemini web-based products," with CLI support only "under active consideration." You can pay for Workspace AI and still hit free-tier walls in your terminal.

Even paying doesn't remove the ceiling. Google AI Pro moves you to 1,500 requests/day and Ultra to 2,000 — but as the inventivehq analysis puts it, "AI Pro and AI Ultra move the wall; only metered billing removes it." Pay-as-you-go billing through AI Studio or Vertex AI is Google's own "recommended path for uninterrupted access," charging per token instead of enforcing a daily cap.

And there's a quieter trade-off most people miss: on the free tier, your usage may be used to improve Google's products — paid tier usage is not, per Google's API pricing terms. If you're running proprietary code or client work through the CLI, that distinction matters.

This matters for agent builders for a practical reason: agentic workflows burn quota fast, since tool loops and sub-agent calls consume multiple requests per task. When every request counts, the skills and instructions you load into your agent directly affect how much work each request accomplishes. That's exactly the problem SkillzBench is being built to solve — a free, independent directory of agent skills for Gemini CLI and other agents, with security scanning and sandbox benchmarking so you can see which skills actually earn their quota before you install them. Join the waitlist to get access when it launches.

The takeaway: before assuming "Gemini CLI is free," check which door you logged in through. The answer changes everything.

Why Your Quota Disappears Faster Than Expected

The 1,000 requests a day Google promises at sign-in sounds generous — until you realize an agentic coding session can burn through dozens of them in a single afternoon. The math of how agents actually consume quota surprises almost every new user.

Here's the core issue: an agent doesn't make one request per task. It runs tool loops — reading files, running commands, checking results, revising — and each step counts as a separate model request. A task that feels like "one prompt" can quietly consume 10 or 20 requests behind the scenes, according to third-party analysis of the free tier.

Two hidden behaviors make it worse:

• Sub-agent calls count toward your quota — and may use models you didn't explicitly pick.
• Requests aggregate across model families, so switching to Flash saves latency, not quota.
• Hitting the cap returns a hard rate-limit error with no auto-retry, halting your workflow mid-task.

That last point matters most. When the limit hits, work simply stops — and since daily limits reset on a rolling ~24-hour basis, you may be waiting until the next cycle before you can continue.

Monitoring helps. The official documentation notes that you can check usage with the /stats model command, and a usage summary appears on exit at the end of each session. Checking before starting a long agentic run is cheap insurance against a mid-task shutdown.

If you consistently hit the wall, pay-as-you-go billing is Google's recommended path for uninterrupted access. Estimated costs from third-party analysis run roughly $5–15/month for light usage (around 100 requests/day) up to $75–150/month for heavy usage (1,000+ requests/day). That's a meaningful jump from "free," so it pays to know your actual consumption before committing.

Reducing request waste matters too. Well-chosen agent skills can tighten tool loops by giving an agent clearer instructions up front — SkillzBench is building a free, verified directory of such skills, complete with security scanning and benchmarking, so you're not guessing which ones actually help.

What to Do Now: A Decision Guide for Free and Paid Paths

You've mapped out the quotas, weighed the March 2026 policy shift, and now you're staring at a terminal wondering what to actually type first. The good news: the right starting move is simple, and the escape hatches are clear if you outgrow the free tier.

Skip the API key path entirely for now. Signing in with a personal Google account unlocks 1,000 requests per day with full Gemini 3 model access, compared to just 250 requests per day on Flash models only through an unpaid API key. Google's official quota documentation confirms the Google account method is the most generous free experience available.

If your work depends on Gemini Pro models, the clock is ticking. As of March 25, 2026, free tier users are limited to Gemini Flash models, and Pro access requires a paid subscription. Community reports suggest free Flash quotas may have dropped from the originally advertised 1,000 requests per day to roughly 150–250 requests per day, though these figures remain unverified against current official documentation.

Google explicitly recommends pay-as-you-go billing through AI Studio or Vertex AI as "the recommended path for uninterrupted access" once daily limits become a bottleneck. Estimated costs range from about $5–15 per month for light usage to $75–150 per month for heavy usage. Pay-as-you-go removes the daily cap entirely, billing per token instead.

Free tier usage may be used to improve Google's products, while paid tier usage is not. Users handling sensitive code or proprietary data should factor this into their decision before committing to the free path.

Agent skills let you automate workflows inside Gemini CLI, but they still consume requests per task. The smart play is choosing skills that have been independently benchmarked rather than guessing which ones actually deliver. SkillzBench is building exactly that: a free directory where every skill is security-scanned, quality-judged, and tested in a sandbox with and without the skill, publishing the performance Lift in percentage points. Once it launches, you can browse and test free skills on SkillzBench to extend Gemini CLI's usefulness without trial-and-error quota drain.

A quick decision framework:

• Just exploring? Google account login, monitor usage with /stats model
• Need Pro models daily? Plan for a paid subscription or pay-as-you-go before March 2026 restrictions bite
• Hitting caps mid-project? Switch to AI Studio or Vertex AI pay-as-you-go for uninterrupted work
• Handling sensitive data? Paid tier avoids the product-improvement data usage clause
• Want to automate more? Wait for SkillzBench's free skills directory to pick vetted agent skills without burning quota on trial runs