This is huge... At least from what I've seen, many companies (even if they don't know it) are largely sloppy R&D shops. Most people are on projects that are highly speculative and unlikely to generate revenue. But it often isn't framed that way. Everyone believes their project is the obvious way forward.
This varies wildly person to person but for me, it's not only about the means. It's a combination of means and end. If I want the computer to do something, and the only way to get it to do so is read a book and tinker, the reading and tinkering are very fun. If I could just ask the computer in plain English (which I can now), I would. Of course it's way less of a fun journey then. But just reading and tinkering with no extrinsic element is not a fun journey.
The game changes when it's not just "Opus vs Sonnet", but "Opus vs GLM". The amount saved is way more than even $0.04. And it's not only money but speed. Some providers can serve GLM crazy fast - I'll even go outside of my subscription to pay extra money for the quick results.
This sounds like not the best workflow. If your prompt is that long, you may not be spending your time very efficiently.
Splitting problems into smaller problems is huge. You want a concrete idea that you still own. Only tell an agent to do something you know it can nail - this will likely be one subsystem. The subsystems and how they talk is on you.
I'll agree that GPT 5.6 may well be the best given the above contstraint, but for run-of-the-mill dev tasks (real ones, not benchmark ones), GLM 5.2 still blows every other model out of the water.
Cost per task as a metric is a bit ridiculous because there are so many types of tasks. GPT-5.6 can do some tasks GLM could only dream of, but GLM can do some tasks 100x cheaper and better than GPT-5.6.
Oh My Pi has it. I'm a big OMP shill right now. Seems not very popular, but it has the stability of Pi with the features Opencode (and more I think; OMP has web browsing and a more advanced edit system too). OMP often outperforms Claude Code and Codex for me.
The artificialanalysis cost per task chart has DeepSeek as the clear winner and Fable as the clear loser. But I would still pick Fable for some tasks, so that also can't be all there is to it.
But I agree that price per token figure is not great. It seems even the tokens per character can vary between models, so it's basically useless.
> But artificial intelligence, far more than any tool we’ve ever created, intends us not just to sit forward and behave, but to cease to think critically, to cease to imagine, and, most temptingly, to cease to feel struggle and pain.
If I don't think critically, AI causes me to feel struggle and pain.
I consider myself an LLM maximalist, at least in the context of software engineering. Even after delegating everything you can to LLMs, there's still work left for you. And it's far more interesting. Even the delegation to LLMs is interesting - it's not that easy to make it work well.
In my domain (web app CRUD), we build the same thing over and over and over. And it's not beautiful or interesting like furniture. It feels like with LLMs, all that repetitive stuff is gone and I can focus myself on the tricky problems.
Right I saw them saying something along the lines of "they're good at subagents". But this seems true even with third party harnesses. So I'm wondering what Codex is hiding.
I've also seen Google indexing pages with random values in the path that don't get linked to statically (server asks for the URL then redirects to it immediately). I'm pretty sure they index straight out of the Chrome address bar.
I guess this implies that non-Codex harnesses get a little bit worse? In wondering what's so special about their subagents system that they feel the need to hide these messages...