.png)
1 month ago
8
TL;DR. In day-to-day coding, lower latency matters more than a small accuracy edge. Codex CLI (GPT-5 High) often feels more "correct", but Claude Code 2's (driven by Claude Sonnet 4.5) faster loop keeps me from switching tasks and preserves my working context.
"Context is king" when talking about LLMs. Give the model relevant files. Thorough documentation. The right background information. We're becoming good at managing their context.
But we forgot about ours.
At the end of August, Claude had some issues. I switched to Codex CLI with GPT-5. It felt like the right thing to do. I ran a few A/B tests with Claude Code and Codex side by side, and Codex was the absolute winner accuracy-wise. Yes, it took a few times longer to finish the task, but it seemed worth the wait.
My reasoning was: while Codex is doing its job, I'd switch to another task and carry on with it. Rather than instructing Claude how to make fixes, I'd fill the time with another task and return when Codex was done. I did this for a few weeks.
A couple of days ago, Claude Code 2.0 came out. I gave it a shot. The difference was immediate - not in accuracy, but in something I hadn't been paying attention to: I wasn't switching tasks while waiting.
I still find Codex with GPT-5 (High) more accurate, but the whole approach of "give it a task and switch to another one" works poorly for me. Context switching has - surprise! - high costs.
Here's the pattern: You ask the tool a question. Maybe you need to make refactoring, or understand why a test is failing. With a slow tool, you know you have a few minutes. So you switch. Start another task in a separate terminal. Scan HN. Maybe reply to an email.
Then the response comes back. Now you have to reload: What was I doing? Which file was I in? What was the bug? What had I already tried?
This reload cost (the mental overhead of reconstructing where you were) is exactly what we know happens with context switching. It's well documented. But somehow we forgot this applies to us while we wait for our tools.
That's where speed matters. This isn't about being impatient. It's about protecting flow state. When you're deep in a problem, you have this fragile graph in working memory: how the components connect, where the data flows, what you just learned from the last error. Every task switch shatters that graph.
With Sonnet 4.5 (running in Claude Code 2.0), responses come back fast enough that I don't switch. I stay in the problem. The mental model stays loaded: the architecture, the bug, the three things I've already tried, the hunch about what might work next. My personal context stays intact.
Context is king. Both kinds.
