TL;DR: Every AI agent’s free tier is quietly getting killed across the industry. Cursor’s pricing shift drew backlash and refunds, Google’s open-source Gemini CLI gave way to the closed-source Antigravity, and Claude Code launched with no free tier at all. We did the same thing to our own product, and the number that forced it was uglier than we expected: $405 burned in a week, with one user alone hitting $116.64 in a single 421-turn session. Here’s the actual math behind why this is happening everywhere at once, and the fix we ended up shipping: compute that belongs to whoever uses it.
If you’ve used more than one AI coding or agent tool this year, you’ve probably noticed the same thing happening everywhere: free tiers are shrinking, disappearing, or turning into something with a dollar cap instead of a request cap. Cursor moved from a flat request allowance to usage-based billing in June 2025 and ended up apologizing and issuing refunds after users hit charges they never saw coming. Google open-sourced Gemini CLI, then put its weight behind the closed-source Antigravity CLI. Claude Code never had a real free tier to begin with; it rides on Anthropic’s paid plans, which top out at $200 a month.
| Who | The move | Who pays the compute now |
|---|---|---|
| Cursor | Flat request allowance became usage-based billing, then came apologies and refunds | The user, down to token-level metering |
| Open-source Gemini CLI gave way to the closed-source Antigravity CLI | The user, inside a closed ecosystem | |
| Anthropic | Claude Code shipped with no free tier; paid plans top out at $200/month | The user, from turn one |
| CrossMind | Cloud free tier capped to a trial; local runs on your own machine don’t touch the quota | Split: you pay locally, we cover the small trial |
The common read on this is “AI companies are getting greedy.” We used to believe a milder version of that too, until we ran our own numbers and found we were about to do the exact same thing, for the exact same reason.
The week our free tier broke
CrossMind’s free tier let anyone spin up an agent session and run it against our infrastructure, no credit card, no cap on turns. It felt like the obviously right move for adoption. Less friction between a builder and their first real result, the better.
Then we looked at the actual API spend for a single week: $405. Not from a spike or an attack, just normal usage. One session, a single user running an open-ended brainstorming conversation, hit 421 turns and cost us $116.64 on its own. Nobody did anything wrong. The agent kept the conversation going because that’s what it’s designed to do, and every turn was a metered API call we were paying for on someone else’s behalf.
At the time, we had one paying customer. $405 a week is roughly $21,000 a year in free-tier subsidy, and that’s at the usage level of a product almost nobody had heard of yet. The math doesn’t get better with more users. It gets exponentially worse, because more free users means more open-ended sessions, and open-ended sessions are precisely the ones that don’t have a natural stopping point.
Why “just add limits” doesn’t actually fix it
The obvious fix is what Cursor and Google both tried first: cap it. Fewer requests, lower ceilings, credit-based metering instead of flat access. It buys time, but it doesn’t solve the underlying problem, and it creates a new one: the free tier stops feeling free. Cursor’s backlash wasn’t really about the pricing model being unfair on paper. Users discovered, mid-session, that the thing they thought was unlimited had quietly become metered. That’s a trust problem more than a pricing problem, and no amount of tuning the caps fixes it.
The deeper issue is that a free tier funded entirely by the company’s own API spend was never a growth strategy to begin with. It’s an unpriced subsidy with no ceiling that scales with adoption instead of against it. The better your product gets at getting people to use it, the faster the subsidy bleeds you. That’s a bad position for a company still figuring out whether people want the product at all.
What we’re doing instead
Instead of tuning the caps, we changed who owns the compute. The August desktop release put CrossMind on your machine as a real app, paired with your own coding agents and runtimes. The September update finished the job: runs on your own computer don’t count against your cloud quota at all. Local capacity is yours to spend freely, because we never touch the request and never pay for it. The $405-a-week problem doesn’t shrink under this model, it disappears, because it was an ownership problem the whole time.
The hosted free tier still exists, and it stopped pretending to be a subsidy. It’s a capped trial now: one concurrent agent, basic usage limits, enough to learn the product. Anyone doing real volume either runs locally or moves to a paid plan, where bringing your own API key for custom models lives on Plus and up. Your key, your meter, our orchestration.
The paid product stays intact. It just moves to where the actual value already lives. We wrote earlier about testing whether CrossMind should be a callable MCP tool instead of a standalone agent, and the answer that came out of that research holds here too. Hosted browser identity, team approval routing, and cross-agent scheduling don’t map onto a stateless, local tool call, and those are also the things nobody else in the local-agent space is building, because tools like Claude Code and Cursor are fundamentally single-player. That’s the part we keep charging for. The raw compute for a solo brainstorming session was never where the value was. It just happened to be where the cost was hiding.
The number worth remembering
If you’re evaluating any AI agent tool right now, or running one, ask two questions of any free tier: who is actually paying for the compute, and what happens to their incentives once the usage numbers get real. A free tier that’s still solvent at 100 users and still solvent at 100,000 users almost never uses the same architecture. Ours didn’t, and $405 in a single week is what made that obvious instead of theoretical.
Shipping that number honestly beat waiting for a caps-and-credits rollout to quietly do the same thing with worse optics. The whole thing is live today: desktop app, paired computers, local runs outside your cloud quota. The pricing page shows exactly where the paid line sits.