Fastest fix: Set your default model to Auto (Composer input bar, model dropdown). On a Pro plan Auto runs effectively unlimited and does not drain your $20 included credit pool, so the “You’ve hit your usage limit” wall stops appearing. Reserve a hand-picked frontier model (Claude Opus 4.7, GPT-5.5) only for the prompts where Auto genuinely fails.
First, an important correction. Cursor retired the old “fast requests / slow requests” system in June 2025. If you read an older guide promising “500 fast requests then unlimited slow,” that no longer exists. As of June 2026, every paid plan ships a monthly credit pool equal to the plan price (Pro = $20 of model usage), credits drain at each model’s real API rate, and when the pool empties you either stop or switch to pay-as-you-go on-demand billing. There is no free “slow” fallback queue to fall into anymore.
So the modern version of the original complaint is: “My $20 Pro credits looked like plenty, then mid-month Cursor said I hit my limit and started charging me / refusing requests.” That is what this guide fixes.
How the current billing actually works (June 2026)
| Layer | What it is | Drains credits? |
|---|---|---|
| Auto | Cursor routes your prompt to a cost-efficient model it picks | No — effectively unlimited on Pro, billed separately at low flat rates |
| Composer 2.5 | Cursor’s in-house agent model | Very cheap; large included allowance |
| Frontier models (Claude Sonnet 4.6, Claude Opus 4.7, GPT-5.5, Gemini 3.1 Pro) | Hand-picked premium models | Yes — at each model’s real API price |
| Tab / autocomplete | Inline completions | No — unlimited |
| On-demand | Usage past your included pool | Yes — pay-as-you-go in arrears |
The $20 Pro credit pool is denominated in dollars of API usage, not a request count. A quick one-line question on Sonnet 4.6 costs cents; a multi-step agent run on Opus 4.7 over a large file can cost a dollar or more in one prompt. That is why the pool feels like it vanishes — you are spending real API dollars, not “requests.”
Common causes
1. You default to a frontier model instead of Auto
Manually picking Claude Opus 4.7 or GPT-5.5 as your everyday default bills every prompt against the $20 pool at full API rate (as of June 2026, Opus 4.7 is $5/$25 per 1M input/output tokens; GPT-5.5 is $5/$30). A handful of big agent runs can clear $20 in a week.
How to judge: open the Usage page (below) and group by model — most users find 60-80% of spend went to one premium model.
2. One agent “turn” fires many model calls
In agent mode the model chains read_file → grep → edit_file → run_terminal_cmd, and each step is a billed model call. “Fix this bug” can be 3-8 calls, not one — and each call’s input grows as it accumulates context, so token cost compounds.
How to judge: compare the request count on the Usage dashboard to the number of prompts you actually typed today. A 3x gap means agent steps are the spend.
3. Max Mode multiplies token cost
Composer’s Max Mode sends the model a much larger context window per call, so each call bills far more tokens. One large refactor in Max Mode can spend several dollars on a single prompt.
How to judge: look at the model name in the Composer input bar — a Max badge means it is on. Max Mode is opt-in per model.
4. You hit your credit cap, not a rate limit (they are different)
Two separate things can stall you, and people confuse them:
- Credit cap: your
$20pool is empty. You see “You’ve hit your usage limit” with options to enable on-demand spending or upgrade. Nothing queues; manually selected models just stop until you pay or the month resets. - Rate limit: during peak load even Pro requests can queue for tens of seconds rather than fail. This is not the same as the old “slow pool” — it is temporary throttling and does not mean you are out of credits.
How to judge: Settings → Usage (or cursor.com/dashboard). If the credit bar is full but replies are slow, it is rate limiting; if the bar is empty, it is the cap.
5. On-demand spending was left on, so “running out” silently became “getting charged”
If you enabled on-demand (usage-based) pricing, you never hit a hard wall — Cursor keeps serving frontier models and bills the overage in arrears. That is the source of surprise charges.
How to judge: Settings → Billing → look for an on-demand / usage-based spend toggle and your monthly spend limit.
6. Usage figures lag a few minutes across devices
After switching machines or networks the Usage numbers can trail real spend by a few minutes, making it feel like “I haven’t used anything.”
How to judge: refresh cursor.com/dashboard or wait 5 minutes; the web dashboard updates faster than the in-IDE panel.
Before you start
- Identify which entry point is spending: Chat, Composer/agent, or Cmd+K. Same billing rules, very different frequency.
- Use the web dashboard at
cursor.com/dashboardfor Usage, not the in-IDE panel — numbers update faster and break down by model. - Note your Cursor version and current default model (Composer input bar dropdown). Different models bill at very different rates.
Info to collect
- Cursor version, current plan (Hobby free / Pro
$20/ Pro+$60/ Ultra$200/ Teams Standard$40seat), and whether on-demand spending is enabled. - Screenshots of
cursor.com/dashboard→ Usage, grouped by model and by day. - Roughly how many prompts you sent today, which models, whether Max Mode was on.
- Whether you are using your own API key (BYOK); if so, confirm each model shows your key, not Cursor’s pool.
Shortest fix path
Ordered by impact on remaining credit.
Step 1: Read your actual spend
Open cursor.com/dashboard → Usage, group by model, and look at the last 30 days. You are looking for the one or two models eating most of the $20. This single screen usually explains the whole month.
Step 2: Make Auto your default
Set the default to Auto (Composer input bar → model dropdown → Auto). On Pro, Auto runs effectively unlimited and bills at low flat internal rates instead of draining your $20 pool. For most everyday coding this alone ends the mid-month wall.
Per-prompt: Composer input bar -> model dropdown -> Auto
Default: same dropdown -> set Auto, escalate to a frontier model only when Auto failsStep 3: Use Composer 2.5 for cheap agent work
When you do want a named model for agent tasks, prefer Cursor’s in-house Composer 2.5 — it has a large included allowance and is far cheaper per task than Opus 4.7 or GPT-5.5. Save the frontier models for genuinely hard reasoning.
Step 4: Turn off Max Mode unless you need it
Leave Max Mode off by default. Enable it only when you must feed the model a very large file or whole-repo context, then turn it back off — it multiplies the tokens billed per call.
Step 5: Lean on Tab + Cmd+K
Cursor Tab autocomplete and Cmd+K inline edits do not draw from your credit pool. For routine completions and “tweak this one line” edits, use Tab/Cmd+K and save Composer/agent runs for real tasks.
Step 6: Decide on on-demand spending deliberately
cursor.com/dashboard → Billing → on-demand (usage-based) pricing. Turning it off gives you a hard stop at $20 (predictable, no surprise bills). Turning it on with a spend limit lets you blow past the pool when you need to, at API rates, without a plan upgrade — usually cheaper than jumping to a higher tier for one busy month.
Step 7: Consider BYOK for heavy frontier use
Settings → Models → add your own Anthropic / OpenAI / Google API key. Calls then bill to your provider account instead of your Cursor credit pool. Trade-off: some newer Cursor agent features need Cursor’s own routing, so BYOK can lag slightly on capabilities.
How to verify the fix
- Switch to Auto, work for a few days, then recheck
cursor.com/dashboardand confirm the$20pool is draining slowly enough to last the month. - Sign into the same account on another device and confirm Usage matches, ruling out a frontend cache delta.
- If you turned off on-demand spending, confirm Billing now shows a hard cap and no pending overage.
If it still fails
- Reduce the repro to one prompt with one model and watch how many dollars a single agent turn actually spends on the dashboard.
- Check whether a recent Cursor update changed your default model or re-enabled Max Mode after the upgrade.
- Search forum.cursor.com for your model’s current pricing; bring your Usage screenshots.
- Grab
View → Output → Cursorlogs and post to the Bug Reports forum; the billing team monitors that channel.
FAQ
What happened to “500 fast requests”?
Cursor retired the fast/slow request system in June 2025 and apologized for the rollout. Plans now include a dollar credit pool ($20 on Pro) that drains at real API rates. If you still see “slow request” wording, you are on a stale client or reading an old guide — update Cursor.
Is Auto mode really unlimited on Pro?
In practice yes for typical use: Auto does not draw from the $20 credit pool and bills at low flat internal rates, so most developers who stay on Auto never hit the cap. Heavy peak-hour use can still get rate-limited (queued), but that is throttling, not running out of credit.
Why did Cursor charge me extra when I “ran out”?
You almost certainly had on-demand (usage-based) spending enabled. Once the $20 pool empties, on-demand keeps serving frontier models and bills the overage. Turn it off in Billing for a hard stop, or set a monthly spend limit.
Why does one “Fix this bug” prompt cost so much? Agent mode chains multiple model calls (read, search, edit, run), and Max Mode enlarges each call’s context. The bill is tokens consumed across all those steps, not one flat request.
Is upgrading from Pro to Pro+ worth it for credits?
Pro $20 includes $20 of usage; Pro+ $60 includes $60; Ultra $200 includes $400. If you only spike occasionally, enabling on-demand with a spend limit on Pro is usually cheaper than the next tier. If you reliably overshoot every month, the higher tier’s bonus usage wins.
Prevention
- Check
cursor.com/dashboardweekly so you can pace spend through the month. - Default to Auto / Composer 2.5; treat Opus 4.7 and GPT-5.5 as escalation models for hard problems only.
- Treat Max Mode as an explicit, temporary action — never default-on.
- Use
.cursorrulesto keep replies focused, indirectly cutting multi-turn agent calls. - For big refactors, agree the plan in Chat first, then let the agent execute once — this avoids expensive agent flail.
Related reading
- Cursor models list out of date
- Cursor rules not loaded
- Cursor Composer confused in large projects
- Cursor Tab not suggesting
- Cursor Agent Mode Tool Call Stuck Pending
Tags: #Cursor #Troubleshooting