Gemini CLI free tier limits: how much work 1,000 requests a day buys

Google's Gemini CLI free tier now serves Flash models only as of 25 March 2026, with a daily limit of 1,000 model requests per user on a personal Google account and 250 for unpaid API keys, according to the official quota page. A single prompt can trigger dozens of model requests, so the daily quota translates to roughly 83 twelve-call agent tasks per day, not 1,000 prompts. The per-minute cap of 60 requests means the daily allowance could be exhausted in under 17 minutes of sustained agent work.

Gemini CLI free tier limits: how much work 1,000 requests a day buys Google documents the free tier as 1,000 model requests per user per day and 60 per minute, on a personal Google account. The unit is a model request , which is one call to a model. A prompt is not one call. The project's own maintainer put it in writing: "A single prompt could actually trigger dozens of model requests API calls ." On the twelve-call agent task this blog has been pricing all summer, a day's quota is 83 finished tasks at the ceiling. Two further things have moved and are missing from most of the articles about this number: the free tier serves Flash models only since 25 March 2026, and the reset window is not documented anywhere. What Google publishes Start with the primary source rather than with the summaries of it. The Gemini CLI quota page lists two unpaid routes, and they are not the same tier under two names. | How you sign in | Per day | Per minute | Models | |---|---|---|---| | Personal Google account | 1,000 model requests | 60 model requests | "across the Gemini model family as determined by Gemini CLI" | | Gemini API key, unpaid | 250 model requests | 10 model requests | "Model requests to Flash model only" | The model column on the first row is worth sitting with. Google does not say which model your quota is spent on, because Gemini CLI decides. That is a reasonable engineering choice and it also means the quota cannot be converted into a known amount of capability, only into a known number of calls to something. The two per-minute ceilings do a second thing that is easy to miss. Sixty requests a minute against a thousand a day is arithmetic anyone can finish: the day's allowance is under 17 minutes of sustained agent work at the rate Google itself permits. Nobody codes at a sustained ceiling for 17 minutes, so that is a bound rather than a prediction. It is still the clearest statement of how small a per-day request count is once an agent, rather than a person, is the thing making the requests. A model request is not a prompt This is the whole article, and it is not our observation. It is in the project's own discussion thread about the daily limit, from a maintainer, in July 2025: "A single prompt could actually trigger dozens of model requests API calls ." The same thread clarifies that the thousand is a total "across model calls including Pro when applicable , Flash and Flash lite", so it is not a per-model allowance either. One counter, every call, whichever model served it. Why an agent behaves that way is the subject of an earlier post here /blog/free-ai-coding-tier-limits/ : you type one instruction, the agent reads a file, edits it, runs the tests, reads the failure, edits again, and each of those steps is a fresh call carrying the entire conversation so far. That post declared a worked task shape in its markup, twelve calls with the context resent every round, and this page reads the same shape out of it rather than inventing a friendlier one. | How you sign in | Model requests a day | Twelve-call tasks a day | |---|---|---| | Personal Google account | 1,000 | 83 | | Gemini API key, unpaid | 250 | 20 | Eighty-three tasks a day is a real allowance and worth saying plainly before anything else in this post is read as criticism. What it is not is a thousand things. The gap between what the number sounds like and what it is comes entirely from the unit, and it is the same gap that turns 50 chat requests a month into about four agent tasks over at GitHub. What changed on 25 March 2026 The figure most people are working from is older than the tier it describes. Google's launch announcement in June 2025 offered "access to Gemini 2.5 Pro and its massive 1 million token context window" at "60 model requests per minute and 1,000 requests per day", and called it the industry's largest free allowance. It was. A service update on the Gemini CLI repository, effective 25 March 2026, states that "Free tier users will be limited to Gemini Flash models" and that "Gemini Pro models will only be accessible via paid subscriptions". The request count survived the change. The capability behind it did not, and a Flash request and a Pro request were never the same thing to begin with. This is the failure mode that makes a published number worse than no number. The quota page and the launch post are both live, both from Google, and a reader who finds the older one first walks away believing the free tier includes a frontier model. Check the date on any allowance you are planning around, including the ones in this post: every figure in the sources below carries the date it was read, for exactly this reason. What the meter cannot tell you A limit you can see coming is a different product from a limit that arrives. On this one, several things that ought to be legible are not. The reset window is undocumented. The quota page publishes the per-day and per-minute ceilings and says nothing about when the day turns over or in which time zone. The Google Cloud quota page it points to says only that quotas reset in alignment with resets for Google Cloud quotas. An issue opened on the repository in March 2026, titled "Quota never resets after hitting daily limit Gemini Code Assist for individuals ", reports the practical version of that: "Hit the daily quota on Friday. Error shows a reset time, but 3 days later the quota is still blocked." It was closed as not planned. The reset stamp that user was shown was in their own local offset, which is a hint that the window is per account rather than a common midnight, and a hint is all anyone has. The CLI's own counter and the enforced quota can disagree. An issue from January 2026 reports the CLI refusing prompts with a usage limit error while its /stats command showed quota remaining, and concludes that "The discrepancy between local stats reporting and API response suggests a sync issue or hidden 'preview' constraints not exposed via the /stats command." Whether that specific case is fixed matters less than what it implies about the design: the meter lives on the server, the display is a local estimate of it, and the two are not the same object. None of this is unique to Google, and none of it is evidence of bad faith. It is what happens when the thing being rationed is counted in a unit the user cannot observe, which is the argument this blog has been making about request counts since its second post. The same day, metered in tokens Here is our own side of it, and the comparison has to be made carefully, because a request quota and a token balance do not convert into each other . There is no exchange rate between 1,000 model requests and any number of credits, and anyone offering you one has made it up. What can be compared is how many of the same twelve-call task each side finishes in a day, with the assumption written down. Clixad meters real tokens. One task costs what its tokens cost, and a free account earns up to 500,000 credits a day by completing advertiser actions in a browser. That is the ceiling below, not a daily expectation: credits arrive on completion , a screenout part way through a survey pays nothing while still taking the minutes, and that is the most common outcome of starting one across the whole offerwall industry. | Model | One task | Tasks in a 500,000-credit day | |---|---|---| | GPT-5 Nano | 3,051 credits | 163 | | Gemini 2.5 Flash-Lite | 5,664 credits | 88 | | Gemini 3.6 Flash | 86,538 credits | 5 | | Claude Opus 5 | 288,450 credits | 1 | Read the third row before the first one. On Gemini 3.6 Flash , the class of model Google's free tier actually serves, a full day of earning here is 5 tasks and Google's free tier is 83. That is not close. If Flash-class models do your work, Google's free tier is more generous than ours by more than an order of magnitude, it costs no minutes on a survey wall, and the correct advice is to use it. The rows above and below it are where the difference sits. A token meter prices whatever you point it at, so the same balance reaches Claude Opus 5 at one task a day and GPT-5 Nano at 163, and the choice is yours rather than the CLI's. A request quota cannot express that: every call costs one request whether it was Flash-Lite answering or a frontier model, which is exactly why the free tier had to be restricted to Flash rather than repriced. The second difference shows up on the task that fails. A run abandoned at step 8 of 12 has spent eight requests on Google's meter and produced nothing, and the counter cannot tell you that, because it counts arrivals rather than value. What a stopped task has already cost /blog/ai-coding-credits-run-out/ is measurable in tokens and is not measurable in requests. How to pick Use Gemini CLI if Flash-class models finish your work. Eighty-three twelve-call tasks a day, no account beyond a Google login, no earning step and no credit card. On that workload nothing here beats it, and the rest of this post is about the cases where the unit stops describing what you get. Use Clixad if the model matters more than the volume. Fourteen models on one balance, frontier models included and not behind a tier, priced per token so a cheap model genuinely costs less rather than costing the same one request. A day at the cap is 163 tasks on the cheapest model and one on Claude Opus 5, and both of those are true at once. Use both, which is what most people reading this already do. They do not share a ceiling, so a blocked quota on one is not a blocked afternoon. Keeping a second agent installed costs nothing until the day the first one stops, and neither of them can tell you when that will be. Use neither if your work needs a Pro-class model every day. Google moved those behind a subscription in March 2026, and our free ceiling reaches Claude Opus 5 exactly once a day. At that point you are looking at a paid plan or your own API key, and the shortlist of tools that include inference /blog/free-ai-coding-agent-no-api-key/ is the wrong list to be reading. FAQ How many requests does the Gemini CLI free tier give you? Google's quota documentation gives two free routes. Signing in with a personal Google account, which is Gemini Code Assist for individuals, is documented as 1000 model requests / user / day and 60 model requests / user / minute. Using an unpaid Gemini API key instead is documented as 250 model requests / user / day and 10 model requests / user / minute, restricted to Flash. The larger number is the one every article quotes, and it is the one worth reading twice, because the unit is a model request rather than a prompt. Is 1,000 requests a day the same as 1,000 prompts? No, and the project says so itself. In the Gemini CLI discussion about the daily limit, a maintainer wrote that a single prompt could actually trigger dozens of model requests API calls . An agent has to read files, run commands, re-read the results and try again, and every one of those steps is another call carrying the whole conversation again. On the twelve-call task this blog has been pricing all summer, 1,000 model requests is 83 finished tasks a day, and that is a ceiling rather than an expectation: any step that needs more than one call lowers it. Can you still use Gemini Pro models on the free tier? No. A service update on the Gemini CLI repository, effective 25 March 2026, states that free tier users will be limited to Gemini Flash models and that Gemini Pro models will only be accessible via paid subscriptions. The launch announcement of June 2025 offered access to Gemini 2.5 Pro and its massive 1 million token context window on the same free tier, and that sentence is still being quoted by review sites more than a year later. It no longer describes what a free account gets. When does the Gemini CLI daily limit reset? Google's quota page for Gemini CLI does not say. It publishes the per-day and per-minute ceilings without stating a reset window or a time zone, and the Google Cloud quota page it links to says only that quotas reset in alignment with resets for Google Cloud quotas. Users report reset stamps in their own local offsets rather than a common midnight, and an issue opened in March 2026 reported a quota that was still blocked three days after the error said it would reset. So the honest answer is that the reset time is not documented and is not reliably observable either. What can you use when the Gemini CLI quota runs out? Either you wait for the window or you stop sharing a ceiling with it. Paying for a subscription lifts the limit, and running a second agent on a different funding model sidesteps it. Clixad is the second kind: credits are metered against real tokens rather than counted in requests, every model in the catalogue is available on the same balance, and the free balance is earned by completing an advertiser action in a browser. It is not a drop-in replacement for a Flash-model allowance, because a day of earning at our cap finishes 5 twelve-call tasks on Gemini 3.6 Flash and Google's free tier finishes 83. The companion to this post is when free AI coding tiers reset, and what carries over /blog/free-ai-coding-tier-reset/ , which puts the undocumented window above next to the five other free tiers and finds that four of six publish nothing at all. If you are still working out where the free credits come from in the first place, the four sources /blog/free-ai-coding-credits/ is the place to start. Sources - Google, Gemini CLI: Quotas and pricing https://google-gemini.github.io/gemini-cli/docs/quota-and-pricing.html , for "1000 model requests / user / day", "60 model requests / user / minute" and "Model requests will be made across the Gemini model family as determined by Gemini CLI" on Gemini Code Assist for individuals, and "250 model requests / user / day", "10 model requests / user / minute" and "Model requests to Flash model only" on an unpaid Gemini API key. The page states no reset window. Read 20 August 2026. - Gemini CLI maintainer, Clarification on Daily Request Limit https://github.com/google-gemini/gemini-cli/discussions/4122 , discussion 4122, 14 July 2025, for "A single prompt could actually trigger dozens of model requests API calls ." and for the daily figure being a total across model calls including Pro, Flash and Flash Lite. Read 20 August 2026. - Google, Service update: mitigating abuse and prioritizing traffic https://github.com/google-gemini/gemini-cli/discussions/22970 , discussion 22970, effective 25 March 2026, for "Free tier users will be limited to Gemini Flash models" and "Gemini Pro models will only be accessible via paid subscriptions". Read 20 August 2026. - Google, Gemini CLI: your open-source AI agent https://blog.google/technology/developers/introducing-gemini-cli-open-source-ai-agent/ , 25 June 2025, for "60 model requests per minute and 1,000 requests per day", "access to Gemini 2.5 Pro and its massive 1 million token context window" and "Simply login with a personal Google account". Quoted as the launch position, which the March 2026 update supersedes. Read 20 August 2026. - User report, Quota never resets after hitting daily limit Gemini Code Assist for individuals https://github.com/google-gemini/gemini-cli/issues/22643 , issue 22643, opened 16 March 2026 and closed as not planned, for "Hit the daily quota on Friday. Error shows a reset time, but 3 days later the quota is still blocked." A user's account of their own experience rather than a statement by Google. Read 20 August 2026. - User report, Incorrect "Usage limit reached" error while /stats shows remaining quota https://github.com/google-gemini/gemini-cli/issues/17081 , issue 17081, opened 20 January 2026, for "The discrepancy between local stats reporting and API response suggests a sync issue or hidden 'preview' constraints not exposed via the /stats command." Also a user report. Read 20 August 2026. - Credit figures and task costs: Clixad's own pricing engine, packages/pricing , over the model catalogue read from OpenRouter on 26 July 2026, recomputed by the test beside this post. The twelve-call task shape is read from the markup of the earlier post /blog/free-ai-coding-tier-limits/ that declared it. - Offerwall payouts: amounts observed on Clixad's own CPX wall in Germany, credited on completion, July 2026. One completed action has credited between 40,000 and 300,000 credits, and payouts vary by more than 2x at the same duration.