When will the AI price wars begin?
When I say AI is too expensive normal people tend to look at me like I'm crazy. ChatGPT has a pretty generous free tier. For those who choose to pay, $20/mo gets you essentially unlimited AI for normal-people stuff. If you're a software engineer, $200/mo gets you unlimited AI on a "max" plan. In all cases I think people are pretty happy with the ROI.
But those prices are the gateway drugs. They're not the real price, not even close. If you work at a company that pays for OpenAI/Anthropic at "API rates" you'll know what I'm talking about. Claude Opus 4.8 costs $5 for 1M input tokens and $25 for 1M output tokens. Fable 5 (RIP) cost twice as much. At those prices without the drug-dealer discounts of the personal plans its easy to rack up big bills. An engineer on my team spent $500 just messing around with Fable for a day.
That is serious money! Teams who previously never paid an annual SaaS bill over $50-100k are looking at seven-figure AI bills. Yes AI replaces some of human labor at some companies, but you can also buy a lot of human labor for a million bucks.
Sam Altman famously promised "intelligence too cheap to meter" but so far he has delivered the opposite. The meter is only going up. When a new GPT comes out the old one doesn't get any cheaper, it hangs around for a little bit at the same price and then disappears. If you thought you'd be able to hold onto your GPT-4 subscription and watch the price drop over time, you're wrong. Yes the new models are better, but that doesn't help when you were happy with the old ones.
So far most corporate leaders have decided just to eat the cost. Do 10% layoffs to pay the AI bill. Tokenmaxx and stay on the train. Impress your friends with your AI spend. What option to do you have? Everyone else is drinking from the fountain of the magical machine god, you can either pay up or get left behind.
OpenAI and Anthropic have, so far, decided to compete only on intelligence and not on price. Today's best models, GPT-5.5 and Opus 4.8, cost almost exactly the same amount (they're both $5 for input , GPT-5.5 is $30 on output). I'm not accusing them of collusion, butthis is a very convenient arrangement for them! A race to the bottom would make it very hard for them to pay those data center bills.
So that's how you get Anthropic going from ~$20B ARR to ~$50B ARR in a year. OpenAI and Anthropic are both targeting $1T+ valuations. When they IPO they will both instantly be in the top 15 most valuable public companies in the world. These numbers are truly hard to comprehend.
I don't think this will last. It's not that the models aren't amazing, they are, it's that people will get sick of paying. You can see the cracks in the dam starting to form. Uber complaining that they spent their 12 month AI budget in 4 months. Tokenminning is replacing tokenmaxxing as the trending topic of the day. Not everyone needs Mythos to write their work emails and write CRUD code!
Recently the Chinese lab z.ai release a new open-weights model, GLM-5.2. This model release will, in hindsight, be seen as a sea change moment. According to most benchmarks it's as smart as GPT-5.4 or Opus 4.7, but it's completely open weights. You can access it in the cloud for $1 per M input tokens and $3 per M output tokens. So it's 5-10x cheaper than the frontier models and only a few months behind. In my personal experience it's really freaking good.
But the point is not just that it's smart and cheap. Unlike OpenAI/Anthropic models the deal will get better over time. Anyone with enough GPUs can run this model as a service, and nobody can "retire" it. If one provider goes down, another will spin up. If one provider finds a way to lower prices, others will have to keep up. If it's $1 in/$3 out today it'll probably be half of that next year. Because it's open-weights it can be run in your own private data center, no need to trust a big model company with your data. Heck if you have enough money you can run it in your house! Fully offline, low-latency, just pay the electricity bill.
In November of 2025, Opus 4.5 was released and closed-weights AI finally got very very smart. A lot of doubters became believers in that moment. Now we have open-weights AI that's even smarter than that was. It's firmly good enough for most things. Once people get a taste of smart AI they can control, I think we'll see some innovation leading to more step-changes in intelligence and efficiency. What requires a $30,000 Mac Pro to run locally now might run on a $3,000 MacBook in a few years. And when OpenAI and Anthropic start to lose customers to open-weights models they might just break. One of them will decide to compete on price, and then the race to the bottom begins. It is obvious that AI is a commodity. Switching models has no cost for most applications. Whoever gives me the most intelligence for the least money will get my business.
As a consumer I can't wait. If you're a shareholder ... it might be a good time to hedge.