Recently used Cursor intensively, making a phase summary.

【Quota】
Officially provided $10K credits, currently used about $8,387 (around 83%), expiring around 6/1.
Over 400k AI line edits on the dashboard, running mostly throughout May.
An unexpected discovery:
Credits are automatically deducted for invoices, so Ultra renewed for another month—this was beyond my expectation, but it effectively gives me an extra month of included quota, which is quite worthwhile.
【My current actual usage】
After not renewing Claude this round, I hardly use Claude daily anymore, and I almost never click Opus.
Now I mainly choose between Codex 5.5 and Composer 2.5:
· Composer 2.5 — Running long loops, agent conversations, batch processing in Cursor
· Codex 5.5 — Local execution, repo modifications, MCP, writing back summaries
Occasionally use Grok 4.3, mainly for intelligence retrieval and digesting external information tasks.
Workflow is also being adjusted: both Cursor and Codex serve as entry points, no longer just using Cursor as a background daemon.
【Pitfalls encountered】
Previously didn’t understand billing, manually selecting frontier models in Cursor, which quickly exhausted the API pool of $400–500 per month.
Now mainly switch to burning through Composer 2.5’s included pool, almost no API pool usage.
Another pitfall is the cost of context in huge threads—short confirmations can also be very expensive.
So I try not to let all tasks crowd into the same thread now.
【Overall impression】
AI agent-type apps are indeed becoming more user-friendly.
The agent conversation entry in Cursor and the app in Codex have been very smooth recently.
For me, it’s no longer about “which model is the strongest,” but more about how to better divide tasks between Codex 5.5 and Composer 2.5.
I will continue to use them. Will organize a more complete case study again at the end of the month.