There are several reasons:
1. Neither of those statements are true. Carbon is substantially more expensive per task than Opus. The reason it doesn't look like it here is because Opus is set to Max. Max effort is sorta like an experimental mode that forces the model to overthrow everything and burn way, way more tokens than is necessary. You should literally never use it in real world work; you should use High or xHigh, which are not only vastly cheaper, but also actually give better outputs than Max.
2. Carbon does not perform better than Opus. For one, their score would actually be tied here if Opus were benched at xHigh rather than Max. More importantly, Google models always perform far better on benchmarks than they do in real world work.
3. "Taste" isn't represented in these benchmarks, nor are communication skills, both of which Opus is massively superior than Carbon in.
@theo@KennesLode It's funny how even in an app for using agents, people forget that many things can be done with agents now instead of requiring an app feature.
@brainiac6996@TimJayas It would except I very extensively use many other models and do not at all have the same issues with them. Opus in particular feels leagues above Carbon.