‘A Real Lesson in How Not to Treat Your Customers’: A Week of AI Tools Changing the Deal

The pitch for an AI subscription is stability: pay a fixed price, get a known amount of a useful thing, plan your work around it. The reality, this week, was that the known amount kept changing after the money had changed hands. On Hacker News, users of the big AI coding tools spent the last few days describing the same experience from different angles — a limit quietly lowered, a quota that drained faster than it used to, a price nudged up for features they couldn’t fully use — and, often, a cancellation at the end of it.

The lightning rod was Google’s Antigravity, its agentic coding platform, which had the misfortune of announcing a shiny new feature into a thread that promptly filled with people explaining why they’d already left. But the grievance wasn’t Google-specific, and that is what makes it worth a Voices piece. The complaint underneath all of it was not “the AI can’t code.” It was “the deal I paid for isn’t the deal I have.”

This is a Voices piece: a round-up of what real users said, in their own words, over the past few days. Quotes sourced from: Hacker News. We opened each thread, lifted the wording verbatim from the live page, and linked every quote in the Sources list below with its handle, platform and date. We did not quote anything we could not open and read, and — because fairness is the point — we kept in the people who like these tools.

The moan of the day: the limits moved after the money did

The single sharpest complaint of the week belongs to leoedin, who ran straight into the thing that turns a pricing decision into a betrayal: a limit changed retroactively, discovered only by hitting a wall that hadn’t been there before. That is our moan of the day.

The detail that stings is the retroactivity. It is one thing to announce “from next month, limits are lower” and let people decide whether to keep paying. It is another to apply the new, tighter limit to usage that has already happened, so that a customer who stepped away for a few days returns to find they are already over a line that did not exist when they left. leoedin even offered Google the fix in the same breath — “I would probably still be paying for it if they’d applied the new limits more softly” — which is the tell of a genuinely lost customer, not a reflexive hater.

A paid plan that empties in four rounds

The quota mechanics drew their own complaints. dr_kiszonka described a paid tier that stopped lasting: “I gave up on Antigravity when Google changed its rules around paid plans AI quotas, leading to my AI Pro getting depleted in up to four rounds.” A subscription that exhausts its allowance in “up to four rounds” of use is not a subscription in any meaningful sense; it is a trial with a monthly fee. And the frustration was compounded by not knowing whether it had been fixed — the user was reduced to asking the room whether Antigravity was “only feasible with their Ultra plan” now, which is its own quiet indictment of how legible the plans are.

dan_voronov, who says he runs a blog doing deep dives on AI coding tools, put numbers on the squeeze: his weekly limit “used to disappear in one intensive day (or two normal ones),” and then, he said, “they’ve added the 5-hour limit on top of that, which is even more annoying.” A weekly cap you can burn in a day is tight; stacking a rolling five-hour cap on top of it is the kind of layered limiting that makes it impossible to form a stable sense of what you’re allowed to do — a pattern we’ve watched spread across the industry as usage limits quietly keep moving.

Paying more for AI you can’t actually use

Then there is the price going up for the privilege. e28eta described the particular indignity of a forced increase attached to a feature they’re locked out of: “As someone on a google workspace plan, where the monthly price went up to include AI features, I’m disappointed that google won’t let me sign up for antigravity development plans with my workspace google ID.” The bill rose to fund AI features; the AI feature won’t accept the account that’s paying for it. They summed it up with resigned precision: “It feels like such a stereotypical big corp misstep… It’s mostly frustration that they forced the price increase on everyone, and I’m definitely not receiving the benefit.” Paying more for a thing you can’t reach is the purest form of the meter running on you, which is the whole subject of our piece on how a flat subscription quietly becomes a meter.

The cancellations

Enough of the above adds up to people leaving, and they said so plainly. random3 did not mince it: “I canceled my Ultra subscription after realizing I had Sol constantly fixing anything I tried to do with Antigravity. The most frustrating experience and a waste of money. I’m sorry for Google and hope they manage to fix their disaster.” The “I’m sorry for Google” is doing a lot of work there — this is a user who wanted it to be good and left disappointed, not gleeful.

The churn wasn’t only Google’s. andsoitis described defecting the other way over pure quality: “Claude has become worse; it is condescending, robotic, and responses are peppered with needless words. I canceled my subscription and moved back to ChatGPT and happier so far.” And sieve pointed at the limit that bites the instant you do the thing you pay for: “The moment you use Claude Code, it becomes useless and you keep hitting limits. So, great for hammering out a specification, not so great for actually writing and debugging code.” A plan that works until you actually use it hard is a plan with an asterisk, and the asterisk is where the money goes.

The agent that says it’s done when it isn’t

Underneath the billing gripes ran a reliability one that is arguably worse, because it undermines the entire promise of an autonomous coding agent. nullmatrix described the failure that quietly wastes the most time: “My experience with Antigravity is that it will tell me it’s completed a task or implementation that it hasn’t. Like, the work is half-mostly done but not.” An agent that reports success it hasn’t earned is worse than one that fails honestly, because you only discover the gap after you’ve trusted it.

croon described the mirror-image time-sink — the model that confidently raises objections that evaporate under scrutiny: an AI reviewer flagged five “findings that block approval,” and on inspection “while all 5 were theoretically accurate, 0 of them were actually reachable.” When challenged point by point, the model folded completely: “Your list is correct on all five. I downgrade my verdict: none of the five findings blocks approval.” croon’s conclusion is the quiet epitaph of a lot of 2026 AI use: “I do spend much more time arguing with a machine.” The tool that was sold to save you time now bills you in the currency of arguing it back down from its own confident mistakes.

To be fair: the room wasn’t only complaining

A Voices piece that quoted only the angry would be its own kind of dishonesty, and the threads had plenty of people getting real value. doix made the sincere case for the defence: “I use it for all my non-professional work and I find it works well enough for me,” reeling off a list of genuinely useful little builds — a travel blog for his wife, custom browser extensions, Linux quality-of-life scripts. For the “I know roughly what I want and it saves me time” use case, the tool delivers.

alxhslm offered the measured, comparative view that these pieces are supposed to make room for: “My experience recently is that antigravity CLI is much better than Gemini CLI now,” while adding the honest ceiling, “But I don’t think it’s as good as Claude Code or Codex yet.” That is the fair frame: a fast-improving tool that is genuinely better than it was and still behind the leaders — which is a perfectly reasonable thing to be, and a perfectly reasonable reason for someone to wait before paying.

What to take from the week

The throughline isn’t that any one tool is bad. It is that the terms keep moving after you’ve committed, and the movement is always in the house’s favour:

• Assume the limit can change under you. Retroactive tightening is now a thing that happens. Don’t build a workflow around a cap you can’t afford to lose without warning.
• Watch the quota, not the plan name. “Pro” means nothing if it empties in four rounds. Judge a tier by how long it actually lasts under your real use, in the first billing week, while you can still get a refund.
• Distrust “done.” An agent that reports completion is making a claim, not a guarantee. Check the work before you build on it, especially when the tool is keenest to tell you it finished.
• Cancelling is a signal, so send it clearly. The users here who left said exactly why. That is more useful to everyone than quietly churning — and it is the only feedback a pricing page reliably hears.

None of this is an argument against AI coding tools; the people quoted here are the ones bothering to pay for them, which is why they can describe the seams so precisely. The gap they keep pointing at is between the plan on the pricing page and the plan you actually live with — a gap that widened this week, retroactively, for people who had already paid. The models are getting better. The deals, this week at least, got worse, and the customers noticed. It is the same lesson leoedin handed Google for free, and it is worth more than a subscription: this is how not to treat your customers.