I had a build plan. Nothing exotic: a straightforward v1, the kind of scope you’d hand to any competent coding agent. I gave it to Codex first, running on my ChatGPT Plus subscription.
It consumed an entire week’s worth of usage credits and still wasn’t done.
I gave the same build plan to Cursor with Composer 2.5. It executed the whole thing using about 3% of my monthly allowance, for a $20 subscription.
The quality tradeoff is real, and it’s fine
Composer 2.5 is not as capable as Sonnet 5.6 or Opus 5. The build it produced wasn’t as clean as what those models would have written. But it was a solid v1: a functioning prototype I could immediately start working with and building on top of. Eventually I’ll pull in Sonnet 5.6 to go back and improve specific pieces further.
That’s the actual workflow, not “pick the best model and use it for everything.” Get a working thing fast and cheap, then spend expensive tokens only on the parts that need them.
What you don’t need the heavyweight champions for
Most day-to-day webdev work doesn’t require frontier-model reasoning. It requires an agent that can follow instructions across a lot of files without getting expensive about it:
- HTML and CSS styling
- Basic database queries
- Deployment configs
- Predictable, repetitive changes applied consistently across multiple files
All of that goes to Composer 2.5. It’s cheap, it’s fast, and for this category of work the ceiling on quality doesn’t matter much because the task itself doesn’t demand much judgment.
What you do need them for
Anything where being wrong is expensive gets escalated. A complex authentication flow is the clearest example: that’s when Opus or Sonnet 5.6 comes in, on the specific piece that actually needs the reasoning.
The lesson isn’t “Composer 2.5 is better than Codex” or “Cursor beats Claude Code.” It’s that the price difference between a $20 flat plan and metered frontier-model usage is large enough to change how you should be splitting the work, and most webdev tasks don’t sit on the side of that split where it matters.