Earlier in the year, there was no real argument about which coding agent to use. It was clear cut Claude Code which had a huge lead over everything else, but how things quickly change in the tech (AI) space.
I've spent the last few weeks actually using Codex and some of the newest open-weight models, and I can't say that anymore. The gap has closed in a couple of spots these competitors are actually ahead of Claude Code.
The harness was the real lead, and it's the thing that got copied
People always talked about Claude Code's lead as if it were the model. It wasn't, or not only. The models were excellent, but the actual moat was the harness around it, the way the agent planned, used tools, read a codebase, ran a command, looked at what broke and kept going without me holding its hand. That loop felt a bit like magic and it was a generation ahead of everything else, it really was.
The problem with a harness lead is that a harness is just an idea, and good ideas get copied. Codex now runs a loop that feels basically the same in day-to-day use. What actually surprised me is the open-weight side. You can now point an open agent harness at one of the new models and get something that, for a lot of ordinary tasks, I genuinely could not have told you wasn't Claude Code.
The models caught up fast and OpenAI feels like they're giving you more usage
The other half is the raw capability, and I've written before about how quickly the open models were closing the gap. I framed it then as a thing that was coming. It came.
The latest Kimi is the one that made it concrete for me. On the kind of everyday coding I actually do, feature work, refactors, the boring plumbing, it holds its own, and it's open weight. Codex, meanwhile, has stopped being the secondary afterthought you reach for when you reach your Claude limits and started being a real first choice for some tasks. I still think Anthropic's best model Fable has an edge at the toughest tasks but it's very minor. The everyday middle, which is where most of the work lives, is a tie now, and a tie is a loss when you used to be the only game in town.
Then there's what it costs, which is where OpenAI is ahead. I pay $200 a month to Anthropic for 20x Max and I pay $100 a month to OpenAI for the 5x Codex plan, and the $100 plan feels like it goes further than the $200 plan. Well, I guess anecdotal evidence and feeling isn't something that's really evidence but I still can't shake that feeling. The two companies count usage differently, neither gives you a real meter, and the numbers you do get are approximate by design. But given similar tasks for a week on both plans, I had around 10% left over on my Anthropic $200 plan while my OpenAI $100 plan had 30% left over.
The Fable safety guards are what actually pushes me toward the door
Using Claude Code lately means never being sure which model I'm talking to. I've written about the safety classifiers quietly swapping me from Fable down to Opus mid-session, and about Fable going in and out of the plans. Put those together with the normal churn of new models landing and old ones moving around, and the day-to-day experience is that I don't have a stable tool. I have a rotating cast. One session behaves one way, the next behaves differently, and I burn attention just working out who I'm dealing with before I can trust the output.
A competitor's pitch, the open-weight one most of all, is now just this: it's the same model every time, and it's the one you picked. That sounds small. After a few weeks of it, it isn't. Consistency turns out to be a feature, and it's the one Anthropic keeps interrupting.
What would actually win it back
Anthropic has done the magic-trick before with ingenuity, where a release just resets the field, and the underlying research is still top tier. But the lead came from two things, the best model and the best harness, and right now the harness is commoditized and the model is matched for almost every task. Being tied on capability while being the most annoying to actually use is not a position you hold for long.
What would actually win me back
None of this is a bigger benchmark number although it helps. It's mostly about making the tool feel stable enough that I stop looking around.
Kill the weekly limit altogether but keep the five-hour one. This is the change I'd make first. The rolling five-hour window is a fair rate limit, because it stops abuse, it resets on a schedule I can predict, and if I hit it I lose an afternoon and get on with my day. The weekly limit is a different animal though. Hit that on a Wednesday and your weekly limit resets on a Saturday so you are done until the window rolls, which usually leads me switching to Codex until the limit resets. You cannot plan around it either, because you cannot see your burn rate until you have already burnt it, and the usage numbers you do get are approximate by design. One limit is a speed bump. The other one can end your week.
Stop moving the model under me. A classifier deciding my session needs a different model should only be done on questions that warrant it, not a question about to check my security logs to see if there is anything that stands out.
Pick a lineup and leave it alone for a while. Fable in, Fable out, Fable back at half limits, Opus 5 as the new default. Stability is a feature, and right now Codex is quietly winning on it by simply not doing this. They offer their best model without giving you a new tier of limits.
And get properly ahead on the hard tasks. The everyday tie only matters because there is nothing above it that clearly justifies staying. Give me a class of work where nothing else is close and the tie stops mattering.
I moved my whole workflow here when Claude Code was obviously the best thing available. It stopped being obvious.
