Claude or Codex: Which AI model fits your team?

Claude Opus 5 writes better code than GPT Astra. No, I don’t mean Fable.

I test every new release from Anthropic and OpenAI in production codebases. I don’t run benchmarks (who has time for that?), but I do look at how the models perform when you give them real problems.

I’m always surprised by the gap between what the online community says about OpenAI releases like GPT Astra and what I see in my own tests. For me, using Codex mostly means frustration and rework. How can that be?

Over the weekend, I had the privilege of talking to a talented engineer who describes himself as a “Codex power user.” He was visibly shocked when I told him how disappointing my tests with GPT Astra had been.

Then came what I wish we had much more of, one follow-up question after another: “Which harness do you use?”, “Do you use Superpowers only with Claude, or with Codex too?”, “What integrations with external tools have you set up?”, “Do both harnesses get the same instructions, guidelines, and information?”

And suddenly, you’re having a constructive conversation. The takeaway: He uses the “Superpowers” skills, and they help him a lot. I had tested them and deliberately decided against them. Superpowers helps agents with stricter rules and more explicit instructions. That makes Codex much better. Instead, I rely on agents to infer what they need to know from the existing codebase. Claude does that very well; Codex apparently less well. But Claude does take much longer.

Two good engineers, two completely different experiences. Both seem to be valid. What can we learn from this?

1. The first sentence in this post, which I’m sure annoyed a few people, is true. But only for me and the way I work.
2. Other engineers’ opinions tell you nothing about how well you or your team will get on with an LLM, a workflow, or a harness. Absolutely nothing. You can no more trust the senior engineer who says an open-weight model from China is just as good as Claude Fable 5.1 than you can trust me when I keep demonizing Codex.
3. You can’t and shouldn’t impose a rigid workflow on an engineering team. “You all have to use Skill X” doesn’t work. Instead, you should enable each engineer to find the best workflow for them: Give them room to experiment. Provide a platform where engineers can exchange ideas and learn from each other. Talk to each person about the changes they’re experiencing.

Different engineers work in different ways and need different conditions to do their best work. That was true before, and it’s still true today. How well an AI model works in your team depends on the team members. You can’t tell that from a benchmark.