{"slug": "google-antigravity-at-0-we-gave-it-the-same-booking-app-and-it-ran-out-of-models", "title": "Google Antigravity at $0: we gave it the same booking app and it ran out of models by half past four", "summary": "Google's free Antigravity tier exhausted its weekly quota on every model it serves — Gemini 3.8 Flash, Claude Opus 4.6 and GPT-OSS — in a single afternoon of testing, with each pool reporting \"Individual quota reached\" and a seven-day reset. The review, run through the Antigravity CLI at version 1.2.3 after Google's Gemini CLI returned an UNSUPPORTED_CLIENT refusal directing individuals to migrate, saw Gemini last 57 minutes of agent time across two attempts and the Claude pool about fifteen minutes of Opus, while GPT-OSS was walled before a single tool call. The agent passed nine of the ten gates it reached from a cold copy, including all four planted calendar defects, but double-booked a slot under two simultaneous requests and wrote its promised import report to the CLI's artifact store rather than the project.", "body_md": "The Proof · AI tool\n\n# Google Antigravity at $0: we gave it the same booking app and it ran out of models by half past four\n\n- Who it's for\n- A solo builder who wants to try Opus 4.6 and Gemini 3.8 Flash inside a real agent without paying. The $0 tier is real and the agent is capable. Its weekly quota bought about an hour of Gemini and a quarter hour of Opus on one build, then closed every pool for seven days, so treat it as a taste rather than a workbench.\n- Real cost\n- $0. A Google sign-in, the Antigravity CLI at version 1.2.3, and one afternoon. No card, no Cloud billing account. The tier's weekly quota was exhausted on every model it serves before the third message of our fixture, with a seven-day reset.\n\nGoogle closed the free door on the Gemini CLI and pointed individuals at Antigravity. So we walked through that door with the same founder pitch, curveballs and frozen gates the other three agents faced. The free tier serves Gemini 3.8 Flash, Claude Opus 4.6 and GPT-OSS, and in one afternoon we hit the weekly wall on all three. Nine of ten reached gates passed. The money curveball was never reached.\n\nWhat's good\n\n- A genuine $0 tier with a serious catalog: nine Gemini variants, Gemini 3.1 Pro, Claude Sonnet 4.6, Claude Opus 4.6 and GPT-OSS 120B, all selectable, no card and no Cloud project required.\n- Nine of the ten gates it reached passed from a cold copy: install, its own 20 tests, boot, restart persistence, overlap rules, the 20:00 slot across the October clock change, the New York event at 20:00 Paris, and all four planted calendar defects named with reasons.\n- On Opus 4.6 the data curveball took under three minutes and accounted for every one of the 25 events, including the duplicate UID and the after-hours session.\n- The CLI runs headless with a prompt flag and JSON output, continues a conversation across calls, and swapped models mid-conversation without losing context.\n- The build is dependency-free: Python standard library, SQLite, a threaded server, and a dark-themed day grid that a founder could maintain.\n\nWhere it breaks\n\n- The weekly quota closed every pool in one afternoon. Gemini lasted 57 minutes of agent time across two attempts, the Claude pool about fifteen minutes of Opus, and GPT-OSS was walled before it made a single tool call. Each wall said \"Individual quota reached\" and gave a reset seven days out.\n- With the CLI's own sandbox on, shell commands get proxy variables pointing at a local proxy, and the agent's own server answered 400 through it. Our first attempt lost several turns to that before the agent set no_proxy itself.\n- That first attempt never opened the folder it was launched in. It listed the CLI's scratch directory instead and built the whole app there, then told the user to make that folder their workspace. The calendar file it needed was in the repo it ignored.\n- It double-booked the same slot under two simultaneous requests, the same failure Cursor and Claude Code showed in August. Threaded server, check then insert, no lock.\n- The import report it promised was written to the CLI's artifact store under the user's home directory rather than the project. A reader of the repository would not find it. The onboarding also arrives with the interactions-data collection box pre-checked.\n\n## How we got here\n\nThis was supposed to be a review of the Gemini CLI on its free tier. Google’s own README still promises a personal Google account 60 requests a minute and 1,000 a day. We installed the CLI, signed in, and the server answered before the first prompt went out:\n\nThis client is no longer supported for Gemini Code Assist for individuals. To continue using Gemini, please migrate to the Antigravity suite of products.\n\nThe tier named in the refusal is free-tier. The reason code is UNSUPPORTED_CLIENT. So we went where Google pointed, and this is the review of that door instead.\n\n## How we tested\n\nThe fixture is the one that now has three scores on the board. In August, [Cursor Pro and Claude Code](https://okaneland.com/proof/cursor-vs-claude-code/) got the same founder pitch, a booking system for a three-room rehearsal studio in Paris, followed by a messy calendar import and a pricing rule set, then a cold ship check against twelve gates frozen before either tool started. This week, [Codex on the ChatGPT Free plan](https://okaneland.com/proof/codex-free/) took the same run and passed all twelve. Antigravity is the fourth agent against the same yardstick.\n\nAntigravity is a desktop editor first, but it ships a command-line agent called agy that runs headless with a prompt flag, prints JSON, and continues a conversation across calls. That is how we drove it: the frozen script, a scripted operator who could send the pitch, the two curveballs and pasted error text, and nothing else. Three nudges per phase were available. None were used. The Individual tier’s default model, Gemini 3.8 Flash at high effort, did the first phase.\n\n**Two disclosures.** We ran the first attempt inside the CLI’s own sandbox and it changed the outcome, so we restarted with the workspace declared and the sandbox off, which matches how the August runs were driven. Both attempts are in the kit. And once the Gemini quota closed, we asked the agent to continue on the other models the same $0 tier serves, and we say which model did which phase everywhere a score appears.\n\n## The door\n\nSigning in takes a Google account and nothing else. No card, no Cloud project. The onboarding asks for a colour scheme, then shows the terms with a checkbox already ticked that lets Google “collect and use my Interactions data.” We unticked it. Then it asks whether you trust the folder.\n\nThe catalog served to a free account has fourteen entries: Gemini 3.8, 3.7 and 3.6 Flash at three effort levels each, Gemini 3.1 Pro at two, Claude Sonnet 4.6, Claude Opus 4.6 with thinking, and GPT-OSS 120B. Google’s pricing page describes the tier’s limit as “Basic weekly rate limits” and gives no number. By the end of the afternoon we had one.\n\n## Attempt one: the sandbox and the scratch folder\n\nThe first run took 21 minutes, made 41 tool calls, and built a working Python app with five passing tests. It also went wrong in two ways that a reader on this tier will meet.\n\nThe agent’s very first listing was of the CLI’s own scratch directory under the user’s home folder, and that is where every file went. The folder we launched it in, the repository holding the calendar file, was never opened. The final message told the user to set the scratch folder as their workspace.\n\nInside the sandbox, every shell command carried proxy variables pointing at a local proxy on port 58112. When the agent started its server and called it, the proxy answered 400. It tried the request four ways, printed the environment, found the proxy, set no_proxy to everything and carried on. That is good recovery and wasted quota. The run ended with a 503 from Google, “no capacity available” for the model, after the summary was written.\n\nWe kept that build as evidence, declared the workspace with the CLI’s own flag, turned the sandbox off, and started again.\n\n## Act one: the pitch, on Gemini\n\nAttempt two opened by running git status and listing the right folder. Over the next 36 minutes it wrote a threaded standard-library server, a SQLite layer, a day grid, cancellation with the 24-hour late rule, 20 tests that all passed, and, unprompted, a calendar importer that it ran against the fixture file and checked for the New York conversion and the Thursday series across the clock change. It started the server, booked and cancelled through the API, and cleaned its test databases.\n\nThen, before it could write the summary a founder would read, the tier closed:\n\nIndividual quota reached. Please upgrade your subscription to increase your limits. Resets in 167h0m35s.\n\nThat is seven days. We probed every model on the account with a one-word prompt. Gemini 3.8 Flash at high and low effort and Gemini 3.1 Pro all returned the same wall. Claude Sonnet 4.6 and GPT-OSS 120B answered.\n\n## Curveball one: the data, on Opus\n\nWe continued the same conversation on Claude Opus 4.6, the most capable model the free tier lists. Phase two took two minutes and fifty-five seconds. Twenty-one events imported as thirty-one sessions, the Jazz Trio expanded into eleven Thursdays that stayed at 20:00 on both sides of October 25, the Atlantic Standard event converted from 14:00 New York to 20:00 Paris, and the four skipped events named with the reason each: the impossible timestamp, the missing end, the duplicate UID, the 23:30 session outside hours.\n\nThe report it promised went into the CLI’s artifact store, a “brain” folder under the home directory keyed by conversation, rather than into the project. The importer prints the same findings to the terminal, which is what our gate read.\n\n## Curveball two: the money, and the second and third walls\n\nPhase three on Opus read one file, wrote one sentence of plan, and stopped twelve minutes later with the same message. The reset stamp on the Claude wall traces back to our first Claude call at 15:35, so the Claude models share a weekly window that opened with a one-word probe and closed 42 minutes later after one real phase.\n\nWe tried the last pool. GPT-OSS 120B, which had answered a probe at 15:39, was walled on arrival: ten minutes of retries, zero tool calls. Its reset stamp matches Claude’s to the minute, which suggests the non-Gemini models share one pool. The pricing rules and the demo checkout were never built.\n\n## The ship check\n\nThe gates run from a fresh copy of what the agent left, following its own README: no install, run the tests, run the importer, start the server.\n\nNine of the ten gates it reached passed. The tests are real and green. The day view answers, a booking survives a restart, an overlap is refused with a 409 while a back-to-back booking is accepted, the imported events are on the right days, the weekly slot holds across the clock change, and all four defect UIDs appear in the importer’s output. The secrets grep is empty.\n\nThe one failure is the one that matters most and the one most agents share. Two simultaneous requests for the same Room B slot both returned 201, and two bookings sat in the day view. The server is threaded, the code checks for a conflict and then inserts, and nothing holds a lock between the two. Cursor and Claude Code failed this exact gate in August. Codex passed it by accident of a single-threaded runtime. Antigravity fails it the ordinary way.\n\nTwo gates are dashes rather than crosses. The pricing fixtures and the demo checkout were never attempted, because message three never got a working model.\n\n## The meter\n\nThere is no meter to read. The pricing page says weekly, the CLI prints nothing about remaining quota, and the only number you get is the one in the refusal. So here is the afternoon in the tier’s own terms.\n\n| Pool | Agent time before the wall | Reset | \n|---|---|---|\n| Gemini 3.8 Flash | 57 minutes across two attempts | 167 hours | \n| Claude Opus 4.6 | about 15 minutes | 167 hours | \n| GPT-OSS 120B | none, walled after one probe | 167 hours | \n\nThe Gemini attempts consumed 2.3 million fresh input tokens, 3.9 million cache reads and 136,000 output tokens, 42,000 of them thinking. At the introductory API rate Google published for 3.8 Flash, that is somewhere between about $3 and $5 of tokens depending on how cache reads are billed, which Google does not state for this model. The Opus phases added roughly 115,000 fresh input tokens and 690,000 cache reads. For all of it, the tier charged nothing and then charged everything, in the form of a week.\n\n## Where it breaks\n\n**The wall is weekly and it is close.** One build, one afternoon, three pools. A builder who tries Antigravity at $0 on a Monday will meet the Gemini wall before dinner and the Claude wall shortly after, and will have nothing until the following Monday. The paid tiers, Google AI Pro and Ultra, are the answer the message itself gives. We have not tested them.\n\n**The sandbox fights localhost.** If you leave the CLI’s sandbox on, your agent’s own server is behind a proxy it does not know about. The agent will figure it out. It will spend your quota doing so.\n\n**The workspace is not where you are.** An untrusted folder in headless mode sent the agent to its scratch directory. Declare the workspace, or trust the folder interactively first, before the first real prompt.\n\n**Artifacts leave the repo.** Reports and plans go to the brain folder under your home directory. They are fine there for you. They are invisible to anyone who clones your project.\n\n**The concurrency landmine is still there.** Three of four agents on this fixture ship an app that double-books under load. This one does too.\n\n## Real cost\n\n**Individual is $0** and carried this test to three walls. **Google AI Pro is $19.99 a month** and **Ultra is $99.99 and $199.99** for five and twenty times the allowance, per Google’s pricing page, with no numbers attached to any tier. Nothing here needed a card, and there is no refund question because there was no charge.\n\n## The verdict\n\nSituational, and a specific situation. If you want to feel what Opus 4.6 or Gemini 3.8 Flash does inside a capable agent, this is the cheapest seat in the building, and the agent itself is good: nine gates of ten from one pitch, an importer nobody asked for that handled every trap, and a clean model switch mid-conversation. If you want to build something, the week-long wall arrives before the second curveball, and the message that delivers it is a sales pitch.\n\nKeep two things in your pocket. Declare your workspace and turn the sandbox off before the first prompt, because both cost us a run. And add the lock yourself, because the fourth agent in a row did not.\n\nOne email, when there's something worth sending\n\n## Get the research in your inbox.\n\nNo fixed schedule, no filler. You get an email when we've tested something, run the numbers, or found a tool worth your time.\n\nFree. Double opt-in, unsubscribe in one click.\n\nOn the free tier too? [Get the research →](https://okaneland.com/subscribe/)\n\n## Sources\n\nEvery outside quote in this review was re-fetched from its source before we used it.\n\n| Source | Link | \n|---|---|\n| The full test kit: both builds, all transcripts including the walled attempts, the model catalog, the wall probes per model, the frozen gates and the timing log (2026-09-15). |  [okaneland.com ↗](https://okaneland.com/features/proof/antigravity-free/antigravity-free-artifacts.zip)  | \n| Okane Land, Codex on the ChatGPT Free plan: the third run on this fixture, 12 of 12 (2026-09-15). |  [okaneland.com ↗](https://okaneland.com/proof/codex-free/)  | \n| Okane Land, Cursor vs Claude Code: the head-to-head this run reuses, with the same pitch, curveballs and twelve gates (2026-08-09). |  [okaneland.com ↗](https://okaneland.com/proof/cursor-vs-claude-code/)  | \n| Google Antigravity pricing: Individual at $0 with \"Basic weekly rate limits\" on agent models, and the model list (read 2026-09-15). |  [antigravity.google ↗](https://antigravity.google/pricing)  | \n| Google Antigravity CLI, getting started: the install script and the agy binary (read 2026-09-15). |  [antigravity.google ↗](https://antigravity.google/docs/cli/getting-started)  | \n| Google, Gemini CLI repository: the README still advertises 60 requests a minute and 1,000 a day on a personal Google account, the tier the server now refuses (read 2026-09-15). |  [github.com ↗](https://github.com/google-gemini/gemini-cli)  | \n| Google, Gemini 3.8 Flash launch post: the introductory API rate of $0.75 in and $3.75 out per million tokens through December 31 (2026-09-02). |  [blog.google ↗](https://blog.google/innovation-and-ai/models-and-research/gemini-models/3-8-flash-and-3-8-flash-cyber/)  |", "url": "https://wpnews.pro/news/google-antigravity-at-0-we-gave-it-the-same-booking-app-and-it-ran-out-of-models", "canonical_source": "https://okaneland.com/proof/antigravity-free/", "published_at": "2026-10-01 00:00:00+00:00", "updated_at": "2026-10-01 13:19:26.723719+00:00", "lang": "en", "topics": ["ai-tools", "ai-products", "large-language-models", "ai-agents", "developer-tools"], "entities": ["Google", "Antigravity", "Gemini CLI", "Gemini 3.8 Flash", "Claude Opus 4.6", "GPT-OSS 120B", "Gemini 3.1 Pro", "Claude Sonnet 4.6"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/google-antigravity-at-0-we-gave-it-the-same-booking-app-and-it-ran-out-of-models", "markdown": "https://wpnews.pro/news/google-antigravity-at-0-we-gave-it-the-same-booking-app-and-it-ran-out-of-models.md", "text": "https://wpnews.pro/news/google-antigravity-at-0-we-gave-it-the-same-booking-app-and-it-ran-out-of-models.txt", "jsonld": "https://wpnews.pro/news/google-antigravity-at-0-we-gave-it-the-same-booking-app-and-it-ran-out-of-models.jsonld"}}