If you've used Cursor's Auto mode since it launched, you probably remember it being pitched as effectively free to use on a paid plan — pick Auto, let Cursor route your request to whichever model made sense, and don't think about cost. That framing has been quietly coming apart through 2026. A Cursor forum thread titled "Cursor Pro – Auto Mode 'Unlimited' usage no longer feels unlimited," opened July 9, 2026, pulled 36 replies and nearly 9,000 views — and by August, Cursor's own forum moderators confirmed why: Auto usage now counts against the same usage pool as everything else you run.
This post walks through what actually changed, how Cursor's usage-pool pricing works in 2026, and what your options are if you want a number you can predict in advance instead of a pool you have to watch.
What changed with Auto mode
| Period | Auto mode behavior |
|---|---|
| Original framing | Marketed as unlimited on paid plans — pick Auto, usage isn't metered against your plan |
| Through mid-2026 | Growing user reports that Auto usage was being throttled or capped despite the "unlimited" framing |
| August 2026 | Confirmed: Auto usage draws from the same Cursor Models usage pool as manually-selected models |
| August 2026 (Teams/Enterprise only) | Auto reworked into three "Cursor Router" variants — Cost, Balance, and Intelligence — each billing at the routed model's list price |
On Individual plans (Hobby, Pro, Pro+, Ultra), the Cursor Router rollout hasn't landed yet — you still get the older Auto, which Cursor forum staff described in August as "the classic Auto, which in the new terms is called Auto Cost." It uses a flat per-token rate regardless of which model it routes a given request to, rather than metering against each provider's list price individually. What happens when your pool runs out mid-session, and whether Auto can select pricier "Fast" model variants, weren't clearly documented as of late August 2026 — Cursor's own staff said they'd check internally after users asked.
The practical upshot: "Auto" no longer means "not counted." It means "Cursor picks the model for you," and that pick still draws down the same pool a manually-selected Claude or GPT request would.
How Cursor's usage pools actually work
Cursor prices on usage pools rather than a per-request or per-seat allowance. Each paid plan includes two pools:
- Cursor Models — Cursor's own models (the Composer family, Grok) plus Auto mode usage, priced on the generous end.
- Other Models — third-party frontier models (Claude, GPT, Gemini), which draw down the pool at that provider's own API rate.
| Plan | Price | What you get |
|---|---|---|
| Hobby | Free | Limited Agent requests, access to Composer |
| Pro | $20/mo | Usage pool roughly equal to the $20 price |
| Pro+ | $60/mo | Roughly 3x Pro's pool |
| Ultra | $200/mo | Roughly 20x Pro's pool, aimed at all-day agent use |
| Teams Standard | $40/user/mo | Pro-level pool per seat, plus SSO, analytics, centralized billing |
| Teams Premium | $120/user/mo ($96/mo billed annually) | Roughly 5x Standard's pool per seat |
| Enterprise | Custom | Pooled usage across the org, invoice/PO billing |
Once a pool is exhausted, Cursor can bill on-demand usage in arrears if you've enabled it — at the underlying model's API rate, with no Cursor markup on Individual plans. You can set a spend limit to cap that on-demand spending for the billing cycle, though enforcement isn't instant: usage can briefly exceed the limit before the system catches up, and raising the limit mid-cycle can make previously-credited overage billable after the fact.
Cursor itself doesn't publish exact dollar sizes for the Pro+ and Ultra pools beyond the "roughly Nx Pro" framing — if you want the precise current number, it's in your own account's usage dashboard, not the public pricing page.
Why "unlimited" rarely survives contact with real usage
This is the same pattern that hit Claude Code subscribers when Anthropic's weekly usage boost expired in September 2026: a shared pool sized for typical usage gets stretched by agentic workloads that re-send context and loop repeatedly, and at some point the provider either throttles it quietly or restructures the pricing to make the limit explicit. Cursor's Auto mode and Anthropic's Claude Code limits are different products with the same underlying economics — a "pool" is a business decision about how much average usage a flat price can absorb, not a promise that heavy usage is free.
If your Cursor sessions are long, agentic, and daily, you were always going to find the edge of whatever pool Cursor sized Auto for. That's exactly the profile of user who showed up in that July 2026 forum thread.
Usage pools vs. a flat-rate gateway
| Cursor usage pools | Direct provider API keys | APIClaw (flat-rate) | |
|---|---|---|---|
| Billing model | Dollar-denominated pool per plan, on-demand overage after | Per-token, passthrough rate | Flat monthly price, daily request allowance |
| Who sets the pool size | Cursor, can restructure (as with Auto in Aug 2026) | No pool — bill scales with usage | You, by the tier you pick |
| Cost predictability | Pool size is fixed, but which requests draw it down can change | Variable bill, no usage ceiling | Fixed price, fixed daily ceiling |
| Model access | Cursor's models + third-party, billed differently per pool | Whichever provider you hold a key for | Claude, GPT, Gemini, and more via one key |
| Works outside Cursor | No — Cursor-specific | Yes, any OpenAI/Anthropic-compatible tool | Yes — Cursor, Claude Code, and others |
None of these is strictly better — they trade predictability for different things. Cursor's pools give you a fixed subscription price with a usage ceiling Cursor can reshape, as it just did with Auto. Direct API keys remove the ceiling but make your bill the variable instead — the cheapest Claude API breakdown walks through what that costs for a typical agentic workload. A flat-rate gateway fixes both sides of that trade: price and daily allowance are both yours to pick, and they don't get restructured out from under you mid-cycle.
If you want a number you can actually predict
Cursor's model picker and Auto mode both run on top of whatever API key you point Cursor at — Cursor itself is the editor and agent orchestration layer, not the only possible backend. Pointing it at a flat-rate gateway instead of billing everything against Cursor's own pools is a settings change, not a rewrite:
# In Cursor: Settings → Models → "Add Custom API Key" → OpenAI-compatible
# Base URL:
https://apiclaw.biz/v1
# API Key:
sk-your-apiclaw-keyFull walkthrough, including which model names to use and how team billing works: Cursor custom API setup. The same "unlimited tokens per request, fixed daily allowance" model is explained for Claude Code specifically in the Claude Code API setup guide, and the general flat-rate-vs-per-token tradeoff is covered in the flat-rate AI API and unlimited tokens API comparisons.
Worth noting: switching Cursor's model backend doesn't require giving up Cursor's tab completion, Composer, or agent UI — those stay as they are. What changes is which account gets billed for the model calls those features make.
When Cursor's own plans are still the right call
If you're a light-to-moderate Cursor user — tab completion most of the day, Agent mode for a handful of focused tasks — Pro's $20/mo pool is probably enough, and you get Cursor's own models (Composer, Grok) included at generous rates you won't replicate through a pass-through key. Cursor's native billing also means one invoice and one dashboard for usage across every model, which a multi-provider setup doesn't give you for free. The pool restructuring mainly bites people whose Agent/Auto usage had grown large enough to live near the edge of their plan's pool — if you're nowhere near that edge, the August 2026 change is mostly academic.
The bottom line
"Unlimited" in Auto mode never meant uncounted — it meant Cursor hadn't yet built the plumbing to meter it precisely, and once they did, in August 2026, Auto usage started drawing down the same pool as everything else. That's not unique to Cursor; it's the same shared-pool economics behind Anthropic's own Claude Code limit change a month later. If you want your coding-tool costs to stop depending on how a vendor chooses to size and reshape a shared pool, the alternative isn't finding the next "unlimited" claim — it's picking a billing model where the ceiling is a number you set, not one a changelog resets for you.