I spent some time studying Cursor Ultra’s billing system: two quota pools, Composer Standard / Fast, Auto, MAX Mode, Remote Control, Cloud Agent — all cross-checked against official docs and the dashboard. This is the conclusion version for anyone confused by their bill.

【TL;DR】
1. Auto looks convenient but often costs more than directly locking Composer 2.5 Standard.
2. Fast and Standard use the same model with the same intelligence; Fast is about 6x more expensive — don’t leave it as default.
3. Ultra has two pools: the Composer pool is cheap with generous limits; the frontier model pool is ~$400/month and only gets consumed when you manually select Opus / GPT / Fable.
4. Remote Control ≠ Cloud Agent, but both run the agent loop in the cloud → MAX lights up. Difference: RC’s tools still run on your local Mac; Cloud Agent’s tools run on a cloud VM.
5. Composer 2.5 has a fixed ~200k context; even with MAX enabled you can’t extend it to 1M. To get 1M you need to switch to a frontier model + MAX (expensive).

【Price Comparison｜Per 1M tokens｜Official pricing】
Composer 2.5 Standard: Input $0.50 · Output $2.50
Auto: Input $1.25 · Output $6.00
Composer 2.5 Fast: Input $3.00 · Output $15.00
Quick math: Auto input ≈ 2.5× Standard; Fast output = 6× Standard.

【How the two pools are deducted】
Pool 1 (Composer + Auto): daily agent work mainly consumes this pool — cheapest.
Pool 2 (frontier models at API pricing): Ultra includes ~$400/month. When Pool 1 runs out, overflow goes one-way into Pool 2.
Reconciliation: → Usage → group by model.

【200k vs 1M context】
Composer 2.5: fixed ~200k context, long conversations get compressed — this is a model limitation, not something MAX can fix.
Frontier model + MAX: only then can you reach 1M window, suitable for huge repos, but it eats Pool 2 and burns through your bill fast.
Not an intelligence upgrade — just a bigger window + higher cost.

【Remote Control and MAX — your testing is correct】
The official mobile docs clearly state: once Remote Control is enabled, the agent loop moves to the cloud, but tools continue running on your local machine.
The Cloud Agent docs state: when the agent loop is in the cloud, it’s always in Max Mode, with no toggle to turn it off.
So when you select Remote Control, MAX lights up automatically — that’s normal, not a UI bug.
But the model you choose remains: Composer 2.5 still at 200k — don’t mistakenly assume you’re already on 1M.
To actually get 1M, you must manually switch to Opus / GPT or other frontier models, knowing that it consumes Pool 2.

【How mobile users should choose】
To control your Mac from your phone → set run location to “Local · Remote Control”, don’t select “Cloud” Cloud Agent.
RC: loop on cloud + tools local, MAX lights up, Composer still 200k.
Cloud Agent: loop + tools both in cloud VM, MAX forced, typically consumes more Pool 2.

【How to avoid accidental Cloud Agent】
Don’t select Cloud as run location · Don’t “Move to Cloud” in conversations · Check automations · A cloud icon on recent projects just means history — switch back to local ~/Projects.

【A real billing sample】
Ultra user, first 6 days of July: Included ~$413, On-Demand still $0.
July 4→5 steep increase, main driver: frontier model Fable 5; composer-2.5-fast is a slow bleed.
Remote Control solves whether your phone can control it, not the window size; long frontier agents are the biggest hole.

Architecture diagram below👇
Ref: · ·