Scale · founder · 7 min read

Cursor Rebills Auto on Monday, and the Only Notice Was an Email

On August 24, Cursor's Auto mode stops being a flat rate and starts drawing at the routed model's price. Included usage rises too. Here's the net.

Updated August 24, 2026 ·

pricing cursor agents budget model routing

On Monday, August 24, the way Cursor bills its Auto mode changes. Auto stops costing one flat rate per request and starts drawing at the rate of whichever model your request gets routed to. Cursor’s own email says that for most requests, this means a higher rate than today.

At the same time, the included usage pool for Cursor Models — Auto included — goes up.

Both things are true. Anyone telling you it’s only a price hike or only a generous limit increase is selling you half the story.

Update, August 24: this may not apply to you at all

The change lands today, and Cursor still hasn’t published the new included limits or a per-model rate table. The pricing docs remain unchanged. But a Cursor staff member has now clarified something on the community forum that materially narrows who this affects, and it’s worth reading before you panic about your bill.

On Pro Individual, Auto is still classic Auto — what the new terminology calls Auto Cost — and it’s still a flat per-token rate regardless of which model the router picks underneath. The three-mode router (Cost, Balance, Intelligence), where Balance and Intelligence bill at the actually-selected model’s rate, is Teams and Enterprise only and has not shipped to individual plans.

So if you’re a solo founder on Pro, the honest read as of today is: probably nothing changed for you yet, and the per-model billing arrives whenever Router does. If you’re on a Teams plan, it’s live and it’s the thing described below.

One useful number from the same thread, because it explains why Auto is cheaper than it looks: in long agent sessions most of your tokens are cache reads, not fresh input or output. Auto Cost’s cache-read rate is roughly half of Grok 4.5’s — about $0.25 per million versus $0.50 — which is why Auto often costs less than manually picking the model Auto would have picked anyway.

Also confirmed, and mildly annoying: there is no user-adjustable reasoning-effort setting in Auto. The router picks the model and the effort level per request. If you want predictable reasoning effort, you have to pick a model by hand.

What to do: check Settings → Billing for the new included number rather than trusting any recap post, including this one. We’ll correct this page again when Cursor publishes real figures.

What actually changes

Three points, from the email Cursor sent subscribers:

1. Auto goes per-model. Every Auto request is metered at the price of the model it was routed to, instead of one flat Auto price. Cursor’s stated reason is that the flat rate was too cheap to keep serving the efficient everyday models.
2. Included usage goes up for Cursor Models, including Auto. It applies to the billing cycle you’re already in — no action needed.
3. Routing does not change. Auto still picks the model best suited to the task, and the set of models it can pick from is the same. Claude, GPT, and Grok are all still in the pool.

Your $20/month doesn’t change. What changes is how fast that $20 gets used up.

Why this is worth your attention even if you don’t use Cursor

Cursor didn’t post a blog entry or a changelog note. The pricing docs still show Auto as a flat rate. The only notice was an email, plus a viral X thread the same evening — which is why half the takes circulating this week are wrong.

That’s the pattern worth internalizing. Seat prices are stable and public. Metering rules are unstable and semi-private. Every AI tool you pay a flat monthly fee to is running the same play: the sticker stays at $20 or $25 so the marketing page doesn’t have to change, and the actual economics move in the allowance underneath it.

If you budget by sticker price, you will be surprised. If you budget by watching your usage page, you won’t.

The part nobody is connecting

Five days before the billing email, on August 19, Cursor shipped a cloud agent release that makes agents dramatically more expensive to run — and it’s a good release.

The headline features:

• Subscriptions. A cloud agent can now attach itself to an event source — a pull request, a Slack thread — and wake up when something happens. Agents automatically subscribe to PRs they open and drive them to done, fixing CI and answering bot comments along the way.
• /goal. Give the agent a long-lived objective and it works toward it until complete. Cursor’s own example: /goal fix all flaky tests and make CI green.
• Subagents on their own machines. Each subagent gets an isolated copy of the project on its own VM, so you can run a swarm of them without collisions.
• Steering. You can send a mid-run message without cutting the agent off.

Read those two announcements next to each other. On Wednesday, Cursor gave agents the ability to keep working indefinitely toward a goal, wake themselves up on events, and spawn copies of themselves. On Monday, the meter under all of that switches from a predictable flat rate to the rate of whatever model each step happens to need.

An agent that holds a goal is an agent that decides how many requests to make. A /goal that turns out to be harder than you thought is a loop that keeps buying frontier-model requests until it either succeeds or you notice.

Nobody has published a horror story yet. Give it a month.

What to actually do

If you use Cursor on a paid plan:

Screenshot your Usage page today, before Monday. You want a normal week of Auto activity on record — the model column, and the split between included and on-demand. Without a before, you can’t tell whether the bigger pool covers the higher draw.

After Monday, run one routine edit and one hard reasoning task through Auto and compare cost-per-request to your baseline. If routine work is still cheap and hard work now looks like you picked the frontier model yourself, that’s the change working as designed, not a bug.

Cursor hasn’t published the new included numbers or the per-model Auto table. Check Settings → Billing after the 24th rather than trusting anyone’s recap post, this one included.

If you’re about to try /goal or subscriptions:

Wait a week, then start on something small and bounded. “Fix this one flaky test” before “make CI green.” Watch what a completed goal actually costs before you point one at a real backlog. And be deliberate about which PRs and Slack threads agents subscribe to — an agent watching a busy channel is an agent waking up all day.

If you don’t use Cursor:

Do the equivalent audit on whatever you do use. Open your billing page, find the usage meter, and write down what a normal week costs you. Lovable, Replit, Bolt, and Base44 all run credit systems that can be repriced with an email. The specific mechanism differs; the exposure doesn’t.

The bigger pattern

This is the third repricing this year that can’t be reduced to a single number. Replit split compute from agent credits. Grok 4.6 put a rate cliff at 200,000 tokens that doubles the price of the whole request retroactively. Now Cursor has made its default mode’s rate a function of a routing decision you don’t make and can’t see in advance.

The common thread: the vendor is moving cost out of the sticker and into the meter, and the meter is getting harder to predict at exactly the moment agents are getting better at consuming it. Those two trends are not a coincidence. Longer agent runs are the product, and longer agent runs are also the revenue.

None of this makes Cursor a bad buy. Per-model pricing is arguably more honest than a flat rate that quietly subsidized expensive requests with cheap ones. And a bigger included pool is a real thing you’re getting.

But the era where a $20 subscription meant a $20 month is over across this whole category. Budget for the meter, not the sticker, and check your usage page like you check your bank balance.

Sources

• Cursor changelog: Cloud Agents and Cursor Harness Improvements (Aug 19)
• explainx.ai: Cursor Auto pricing — per-model rates and higher limits on August 24
• FinTech Weekly: Cursor faces backlash over plan pricing shift
• Cursor pricing docs