Most of the AI you use is rented. You pay each time you use it, the company sets the price, and if it retires the model one day, that model is gone.
A smaller group works more like buying. These "open weights" models can be downloaded and run on your own machines, and the copy is yours to keep whatever the maker does next.
So how good is the best AI you can own? Pretty good. Cheap, it isn't.
The best one you can own is #16
Epoch AI's capability index rolls many standard tests into a single score. Sort its 256 models from the top and look for the first one you can download. You pass 15 names before you find it.
It's Kimi K3.1 Pro, from Moonshot, with a score of 161. None of the 15 models above it is listed as open. The leader, Claude Fable 5.2, scores 173.
Seven closed models are clearly ahead of Kimi. The places just above Kimi are a photo finish, so "#16" is where the sort puts it, not a settled verdict. Fairest version: Kimi leads the second pack, a clear step behind the front.
It isn't behind at everything. On GPQA Diamond, a hard science test in Epoch's data, Kimi and Fable 5.2 both score about 94%. On FrontierMath, a maths test, Fable 5.2 gets 95% and Kimi 78%. Different league.
The price twist
You'd expect the model you can own to be the cheap option. At the top, it isn't.
The prices here are OpenRouter's listed developer prices: what companies pay to plug a model into their own apps, not an app subscription. They're in US dollars per million tokens, the units AI text is billed in, and I use one "blended" figure that mixes what you send in with what comes back. All of it is on our compare page.
Rented by the token, Kimi K3.1 Pro costs $4. That's exactly what GPT-6.1 Sol costs, and Sol is one of the seven clearly ahead. GPT-6.1 Terra, also clearly ahead, costs $2: half the price for more capability. If all you want is answers piped into an app, Kimi's flagship is hard to justify.
The money is further down the list. DeepSeek V4.1 Flash 1017 is next among open models, at #22, less than two points behind Kimi. It costs about 15 cents. For what a million tokens of Kimi cost, you get about 26 million from DeepSeek.
Before anyone declares a win for open models: Claude Haiku 5.5, a closed model, sits between them at #18 and costs 20 cents. On score, it and DeepSeek are neck and neck. Open models are cheap, but so are the cheap closed ones.
What owning actually gets you
A model's weights are the huge list of numbers it learned in training. Having them lets you run the model, adjust it for your own job, and keep it.
Two catches. First, the licence. DeepSeek's page says V4.1 Flash 1017 is licensed under MIT, which lets you do almost anything with it. Kimi K3.1 Pro has its own licence with extra conditions for big businesses: past certain revenue or user numbers, they need a separate agreement with Moonshot or must show the Kimi name in their product. The revenue bars start in the tens of millions of dollars, so a student project won't come near them.
Second, size. Moonshot lists Kimi K3.1 Pro at 2.8 trillion parameters, those learned numbers. DeepSeek's page shows V4.1 Flash 1017 at 763 billion. These are not laptop models. My read: for most students and hobbyists, owning still means renting someone else's server, just with more choice about whose.
When I'd pick open
Why this matters even if you never touch the tech: my view is that good open models keep the rented ones honest. When a model you can download, like DeepSeek's, scores close to the paid ones and costs less than any model that beats it, the companies renting AI to you have less room to squeeze.
Open wins when control matters more than the last few points of capability: private data you'd rather not send to someone else's servers, or a product that must keep running for years on terms you can read today. For that I'd start with DeepSeek V4.1 Flash 1017. It's second among open models, MIT-licensed, and the cheapest model in the data that scores at least as high, open or closed.
It's the wrong choice when you just want the strongest answer plugged into an app. Seven closed models are clearly more capable, and one of them costs half what Kimi does.
The best AI you can own is #16. The best deal you can own is #22, and that's the one I'd build on.
