For most tasks I tried the choice is either expensive Claude that will actually deliver vs Codex or others that are cheaper but won’t.

Having said that Terra is awesome for clearly defined (by Opus/Fable) research task, and Sol is amazing for adversarial review, but I won’t allow them to touch my codebase any more.

If someone has a workflow where they keep tight control over agents, reviews the code by hand and so on, I can imagine them being useful.

I seriously doubt that. There's some myth going on here (lol, myth->Mythos).

For example, this exact narrative you're framing was said back during Opus 4.6/4.7 and fast forward to today we have open weight models that perform at Opus 4.6 level.