Every sports day hands out two prizes worth caring about. One goes to whoever crosses the line first. The other goes to the runner who knocked the most off their personal best, and that's usually the better story.
AI just had its own sports day. Between 23 October and 9 November, five companies released new models. Seventeen days, five launches, and each company marks its own homework in its launch post. The fair thing is to put all five on one results sheet, scored by someone who didn't enter the race.
Our results sheet is Epoch AI's capability index, which rolls lots of standard tests into one score and ranks 256 models. Here's where each launch landed.
Now the same list with prices, from OpenRouter's listings on 14 November. That's what companies pay to plug the model into their own apps, per million output tokens (the units AI text is billed in). Nothing to do with a chat app subscription.
- Claude Fable 5.2 (Anthropic): 1st, $40
- Gemini 4 Carbon (Google): 2nd, $10
- GPT-6.1 Terra (OpenAI): 6th, $5
- Kimi K3.1 Pro (Moonshot, open): 16th, $10
- Mistral Large 4 (Mistral, open): 48th, no price in our data
"Open" means you can download the model and run it yourself. The other three you can only rent. Everything is on our compare page if you want to check my maths.
Who crossed the line first
Claude Fable 5.2 (9 Nov) took first place. It also tops Epoch's FrontierMath, a set of very hard maths problems, at 95%. Anthropic says Fable and Mythos are "best for the most ambitious, open-ended technical work", and the maths result backs that up. It's also the most expensive of the five by a distance.
Hold the medal ceremony, though. The top of the table is a photo finish, and Fable 5.2's lead over the next six models isn't settled.
Gemini 4 Carbon (28 Oct) is the only launch near the top of two different scoreboards: second on Epoch's tests, and first on LMArena, a public leaderboard where people vote for the better of two anonymous answers.
Kimi K3.1 Pro (23 Oct) is the best open model on the board, 16th overall, eight places above the older Kimi K3, though that gap isn't settled. Moonshot's launch post calls it its most capable open-weight model in its class. Solid, not a leap.
My winner: GPT-6.1 Terra, most improved
This is the launch that moved the needle. GPT-6.1 Terra (26 Oct) finished 6th. Its predecessor, GPT-5.6 Terra, sits 17th. That's eleven places in one release, and the gap is too wide to be a wobble in the measurement.
And the price went down. Output fell from $12 per million tokens to $5, matching OpenAI's own docs page, which calls it "a fast, efficient model for visual reasoning, precise computer use, and clear, natural writing." Of the 20 highest-scoring models on Epoch's index, only one costs less: Claude Haiku 5.5.
Now put it next to the winner. For the $40 Fable 5.2 charges per million output tokens, Terra writes eight million. And Terra is one of those six inside the photo finish at the top.
Is Fable better? On hard maths, yes: Terra scores 91% on FrontierMath to Fable's 95%. Terra also has no LMArena score yet, so there's no public vote on it. Carbon came close to this spot, but Terra's leap from 17th to 6th while getting cheaper is the cleaner story.
The letdown: Mistral Large 4
I wanted this one to be good. A big open release is exactly what I root for. But the numbers are flat.
Mistral Large 4 (29 Oct) finished 48th, last of the five by a long way. The preview version from 6 October sits 57th, and the two are close enough that Epoch's index can't tell them apart. Mistral's post reports a 13-point jump over the preview on one test, Terminal-Bench, but that's a figure Mistral measured itself.
Our price data has no figure for the final version. Mistral's own listed price would make it the cheapest of the five, and Mistral says the weights are open under what it calls a modified MIT licence. Cheap and open counts for something. But it's only the eighth-best open model on the board.
My take on why any of this matters if you'll never write a line of code: plenty of AI apps are built on top of models priced like these. When a top-six model costs less than half what the last version did, my guess is that shows up later as cheaper or more generous tools. A guess, not a promise.
If you're building on a student budget this month, start with Terra and save Fable 5.2 for the problems Terra can't crack.
