Everyone keeps asking who’s going to win the AI race like it’s a heavyweight title fight with one belt. OpenAI or Anthropic. Pick a side, place your bet, watch one of them get knocked out. And I get why people frame it that way. It’s clean. It’s dramatic. It makes for good Reddit threads. But I’ve been watching this thing closely for a while now, and I’m pretty convinced the whole „winner takes it all“ framing is already wrong. Not „will be wrong eventually.“ Wrong right now. We just haven’t admitted it yet because we’re all still high on the latest model dopamine hit.
Let me back up.
the upgrade treadmill is real and it’s brutal
There’s this thing that happens every time a new frontier model drops. You try it, and within about ten minutes you physically cannot go back. Mythos came out. Fable 5 came out. GPT-5.6 landed. And the reaction on X and Reddit was always the same energy: „I cannot use Opus 4.8 anymore, it feels broken now.“ „GPT-5.5 feels like talking to a toddler after this.“ People aren’t being dramatic for clicks. They genuinely feel it. The jump is real enough that going backward feels like downgrading from a smartphone to a flip phone.
And this is the part that makes the „winner takes all“ crowd feel right. If users always sprint to the newest, most capable model the second it ships, then whoever has the best model this quarter owns the market this quarter. Simple. The frontier is the whole game.
Except it isn’t. And the reason it isn’t is hiding in plain sight.
not everyone is doing the same job
Here’s the thing nobody wants to internalize. The reason people abandon Opus 4.8 the second Fable 5 ships is that they’re doing hard work. Coding. Complex reasoning. The stuff where a 5% capability gap actually shows up in your output and ruins your afternoon. In those tasks, the difference between the best model and the second-best model is the difference between shipping the feature and spending two hours untangling slop the model confidently generated.
But coding is not the whole market. It’s not even most of the market. It just happens to be the loudest part of the market because the loudest people online are engineers.
Think about the actual boring enterprise use case. You’ve got a knowledge base. A few thousand internal docs, some policies, a product catalog, whatever. And you want a chatbot that lets people ask „what’s our refund policy for enterprise customers in the EU“ and get the right answer. That’s it. That’s the job. You wire it up with a RAG pipeline, or you fine-tune something small, and you let it summarize facts back to the user.
Now tell me with a straight face that this job needs Fable 5. It does not. Throwing a frontier model at a glorified document lookup is like renting a Formula 1 car to drive to the grocery store. Yeah, it’ll get there. It’ll also set your money on fire doing it. Frontier models are unbelievably good at summarizing retrieved facts, but a model from two years ago is also good at summarizing retrieved facts, because that was a solved problem two years ago. You’re paying frontier prices for a commodity capability.
This is where the open weight models walk in.
glm-5.2 and the „good enough“ army
So GLM-5.2 just showed up, and the benchmarks are making bold claims. The marketing wants you to believe it’s right there with Opus 4.8. And look, I’ll be honest with you the way I’d be honest with a coworker. In a real coding environment, it is not Opus 4.8. It’s not close. Benchmarks are a vibe, not a guarantee, and anyone who’s actually shipped code with these things knows the gap between „scores well on the eval“ and „doesn’t quietly hallucinate an API that doesn’t exist“ is enormous.
But here’s what I’d put GLM-5.2 at. It feels roughly like early Sonnet 4 or Opus 4. And you know what? That’s a real model. That was a model people were building production systems on not that long ago. It’s open weight. You can run it. You control it. And for a massive swath of tasks that are not „write me a distributed system,“ it is more than good enough.
That’s the trap in the benchmark framing. We keep comparing GLM-5.2 to Opus 4.8 and going „ha, not even close.“ Wrong comparison. The right question is: is GLM-5.2 good enough for the chatbot, the classifier, the summarizer, the internal tool that doesn’t need a genius? And the answer is yes, obviously, easily.
google didn’t lose the model race, they lost the room
Quick detour, because I can’t talk about this without addressing the elephant that left the room. Nobody’s even arguing about Google in the coding and tech spaces anymore. Go read the Reddit threads. The energy around Gemini 3.5 is basically a roast. People openly saying it’s not even as good as Opus 4.5, and at this point the new benchmark people joke about is whether your model is getting banned by the US government before it’s whether it can write a clean function. That’s where the conversation has gone.
And it stings a little, because Gemini 2.5 was genuinely beloved. As a conversational model it was great. People loved talking to it. But „loved over a year ago“ is a death sentence in this market. Nobody who’s tasted the new capable models will go back to a model from last year for hard work. The treadmill doesn’t care about your feelings or your fond memories.
But, and this is the part that should make you pause, the Gemini app is doing fine. The actual chat interface that normal humans use is performing well. ChatGPT is still number one by a mile. And the Claude app, where you’d go to chat with Anthropic’s models? Dead last. Something like eleven percent market share. Eleven.
So Anthropic is getting demolished in the consumer chat app while simultaneously crushing it in coding and enterprise. How does that make any sense under „winner takes all“? It doesn’t. It only makes sense if you accept that these are different markets with different winners.
anthropic figured this out before the rest of us
This is why Anthropic went all in on coding and enterprise. They looked at the board and went, we’re not going to win the casual „help me write a birthday card“ chat war against ChatGPT’s brand and Google’s distribution. So they planted their flag where the frontier capability actually matters and where the money is willing to follow it: developers and enterprise engineering budgets. And it’s working. They’re winning the segment where being the best model is the entire value proposition, because in coding, second best is genuinely painful.
That’s not „winning AI.“ That’s winning a lane. A very lucrative lane. But a lane.
my actual prediction
Here’s where I land, and I’ll say it plainly so you can quote me when I’m wrong.
Coding stays a frontier game. Forever, basically. As long as the capability gap shows up in real output, developers will always reach for the most capable model they can get relative to price. We are capability-greedy and we have good reason to be, because the model is doing load-bearing work. I will pay for the best coding model the way I’ll pay for good tools. It’s the cheapest expensive thing I buy.
But the catch is that word „relative to price.“ Right now all of this is subsidized to hell. The Mythos and Fable tier pricing makes sense because there’s a land grab happening and everyone’s burning investor money to win mindshare. That ends. It always ends. And when the subsidies dry up and the real cost of running these monsters lands on the bill, a lot of people who currently default to the frontier model are going to look at their invoice, choke, and start asking which tasks actually need it.
That’s the moment the smaller and open weight models stop being the budget option and become the smart option. You’ll run GLM-5.2 or whatever’s current as your daily driver for the 80% of work that doesn’t need a genius, and you’ll reach for the frontier only when the task earns it. Not everyone can afford Fable money for every keystroke. Most people can’t. Most companies won’t, once finance gets involved.
So no, it’s not winner takes it all. It’s not even winner takes the most, not for long. What we’re actually heading toward is a stabilized market where different players own different jobs. Frontier coding goes to whoever’s best this quarter. Cheap high-volume inference goes to the open weight crowd. Consumer chat goes to whoever has the brand and the distribution, which right now is OpenAI. Enterprise knowledge bots go to whatever’s cheap and good enough, which is a two-year-old model running quietly in a RAG pipeline that nobody tweets about.
The race we’re all watching has one finish line. The actual market has a dozen. And the companies that figure out which lane they’re in are going to do a lot better than the ones still trying to win a belt that doesn’t exist.
Stop asking who wins AI. Start asking who wins your specific job. The answer is almost never the model you think.