The author signed up for Antigravity under one set of usage terms and within days found rate limits dropped, model routing silently downgrading premium calls to cheaper models, and quota-exceeded errors on workflows that worked 72 hours earlier. They frame this as a deliberate bait-and-switch because Google never publicly acknowledged the change, leaving the launch marketing as the only 'contract' developers had to rely on.
The editorial argues Antigravity is the third act of a now-familiar playbook repeated by Cursor, Windsurf, Replit, and even Claude Code: generous beta to drive viral adoption, then a quiet throttle, then official repricing. The structural issue is that AI coding tool 'terms' are marketing pages rather than SLAs, which makes silent downgrades a defensible vendor strategy across the entire category.
Commenters who inspected network traffic report that request payloads are being downgraded server-side — users select Gemini 2.5 Pro in the IDE but receive responses consistent with Flash. They argue this is worse than a rate limit because it silently degrades output quality without any user-visible signal, breaking the basic trust that the model selector means what it says.
A blog post titled *Google's Antigravity Bait and Switch* hit the front page of Hacker News this week with 642 points and a comment section that reads like a support group for AI-tool early adopters. The author, writing at 0xsid.com, signed up for Google's new Antigravity — the Gemini-powered agentic IDE Google launched in late 2025 as its answer to Cursor and Windsurf — under one set of usage terms and found, within days, that the terms had quietly shifted underneath them.
The specifics are familiar to anyone who's been through a Cursor pricing change or a GitHub Copilot quota tightening: launch-day messaging implied generous (or unlimited) access to high-capability models, the product onboarded developers fast, and then the meter started running tighter. Rate limits dropped, model routing silently swapped premium calls for cheaper ones, and the "free" tier began returning quota-exceeded errors on workflows that worked fine 72 hours earlier. Google has not publicly acknowledged a policy change, which is exactly the problem — when the contract is a marketing page rather than a SLA, "we never said that" becomes a defensible position.
The HN thread surfaces dozens of corroborating reports. Multiple developers describe burning through what they assumed was a daily quota within an hour of normal agentic coding. Others note that the model selector in the IDE no longer reliably honors their choice — ask for Gemini 2.5 Pro, get something that smells like Flash. A handful of commenters dug into the network traffic and found request payloads being downgraded server-side.
This is not a Google-specific story. It's the third act of a pattern that's now repeated across every major AI coding tool launched since GitHub Copilot's pricing reset in 2024: generous beta, viral adoption, quiet throttle, official repricing. Cursor did it. Windsurf did it before the Google acquisition. Replit did it. Even Claude Code's usage limits have walked backward from their early-2025 generosity. Antigravity is just the latest, and Google's scale makes the bait wider and the switch sharper.
The economics force the pattern. Frontier-model inference costs real money — a single agentic coding session can chew through $5-15 of provider-cost tokens if the agent loops on a hard bug. No free tier survives contact with a few thousand power users running overnight agent swarms. The honest move would be to launch with metered pricing from day one and let the market decide. The dishonest move — and the dominant one — is to launch with "unlimited" framing, harvest signups, lock in the workflow integration, and then introduce the meter once switching costs are high.
What makes the Antigravity case sting more than the others is the marketing posture. Google launched Antigravity with explicit language positioning it as the *developer-friendly* alternative to the closed ecosystems of Cursor and Windsurf. The pitch was: Google has the model, Google has the compute, Google has the patience to absorb the cost while you build muscle memory. The HN consensus is that Google had exactly enough patience to clear the press cycle and not a day more.
There's also a deeper structural problem here that goes beyond pricing. Agentic IDEs are not interchangeable text editors. They train your workflow. The mental model of "talk to the agent, let it iterate, review the diff" is sticky in a way that syntax highlighting never was. Once a developer has spent two weeks shaping their prompts and project context to a specific tool, the cost to migrate isn't just the IDE — it's the rebuilt rapport with a different agent. Vendors know this. The throttle comes after the rapport, not before.
If you or your team is evaluating Antigravity right now, three concrete moves:
Assume the current quota is the floor. Whatever rate limit, model access, or context window you see today is the maximum you can plan around. If your workflow needs more than that to be productive, the tool doesn't work for you — regardless of what the landing page says. Don't build internal tooling, agent configs, or team training around capabilities that aren't contractually committed.
Demand pricing in writing before any team rollout. For anything beyond solo experimentation, get a paid tier with a published SLA and a price-change notice period in the contract. Free tiers from frontier-model vendors in 2026 are loss leaders with a 90-day half-life. Enterprise contracts with 60-day price-change notice clauses are the only thing that survives the next round of CFO pressure on inference margins.
Keep your prompts and project context portable. Store your agent rules, system prompts, and project conventions in plain markdown files in your repo — not in the IDE's proprietary settings UI. Cursor's `.cursorrules`, Claude Code's `CLAUDE.md`, and Antigravity's equivalent should all be able to read the same source-of-truth file. If a vendor's pricing changes overnight, you want a one-day migration, not a two-week rebuild.
For solo developers and weekend hackers, the calculus is different: ride the free tiers, switch tools when they tighten, and treat this as the cost of frontier-model inference being genuinely expensive. The mistake is assuming any single AI coding tool is a long-term commitment in 2026 — the half-life of "the best AI IDE" is measured in months, and the half-life of any specific pricing tier is measured in weeks.
The next twelve months will sort the AI coding tool market into two tiers: vendors who price honestly from launch (and survive on margin) and vendors who keep running the bait-and-switch playbook (and burn community trust until the next round of VC funding runs out). Google has the balance sheet to do either. The Antigravity launch suggests they've chosen the latter — at least for now. Watch for the official pricing announcement that follows every one of these episodes, usually 4-8 weeks after the first wave of HN complaints. Until then, code on it if you must, but don't unplug your Cursor license yet.
For Mac users, I wrote (using Antigravity) a self-contained, zero-dependency Python script to restore everything. It safely shuts down background processes, merges your VS Code settings, updates extension pathways, and merges the global SQLite databases using raw base64 protobuf concatenation to res
"We should be able to trust that our tools will remain the tools we actually signed up to use."Yep... well that's what free software and open-source is for. You can't trust corporations so you MUST have the actual code. Harsh lesson but at least if something is learned and the mi
I never really used the Antigravity IDE, but had it installed. The update also made me do a double take and wonder what the hell was going on.It seems like Google is hitting the reset button on the product they call "Antigravity", existing users be damned. Fine, if you've never instal
Google made its lack of interest in Antigravity IDE obvious from very early. Updates were few and far between and app-breaking bugs stuck around, despite tons of reports.Google's lack of focus is astounding. They sprinkle random products here and there and seem to then tepidly pick the product
Top 10 dev stories every morning at 8am UTC. AI-curated. Retro terminal HTML email.
One of the first things I look for when trying out software is the "Updates" toggle. If it doesn't have a way to disable them, most of the time I'll go look for another option.