AI underdogs

AI does not end with the frontier models of OpenAI, Google and Anthropic. Around them is a wide circle of other developers. Their models stand out for how open they are, what they cost and who is behind them. For a company, these qualities can matter more than the high-end answers the big labs deliver. Most everyday tasks need far less.

Five of those developers have released new models in recent months.

Meta and xAI are based in the US, DeepSeek and Moonshot in China, and Mistral in Europe. Each of them competes in its own way.

Moonshot competes on openness. It released Kimi K3 with open weights, so any company can take it and run it itself.¹ Kimi K3 even reached third place in the Artificial Analysis tests.²

For most of a company’s work, an AI model at that level is more than enough.

Free to download, except the best

Meta offers open weights too, but only for part of its range. Muse Spark Open comes with them, so any company can take it in-house.³ Muse Wonder, Meta’s strongest model, stays closed.³ It is sold only through a paid coding subscription, under which Meta, by default, trains on the code its users write.⁴

That closed model, in its Muse Wonder Full version, is 15th on Artificial Analysis.⁵ No other US company releases both open and closed models at this level.⁶

Even so, the best model comes with Meta’s conditions attached. A company that writes its code with Muse Wonder hands that code to Meta for training unless it changes the default.

Open, cheap and the most used

Where Meta holds its best model back, DeepSeek keeps its model open and competes on cost. It released V4.1-Flash at a lower price and under the MIT licence.⁷

For a company, that licence means the model can be adapted to its own work without asking permission.

On OpenRouter, more requests now go to DeepSeek’s models than to anyone else’s, Google and OpenAI included.⁸ That reflects what developers actually use, not which model answers best.

By that measure, models from outside the frontier already lead.

Cost can also be the selling point without any openness.

xAI keeps Grok 4.7 fully closed and prices it low.

It charges $2 for every million tokens the model reads and $6 for every million it writes. xAI sells it inside coding tools such as Cursor and GitHub Copilot.⁹ Grok is a name many people know, yet Grok 4.7 sits around 32nd on Artificial Analysis.¹⁰ Like DeepSeek, it competes on cost, and also on being available inside those tools. Unlike V4.1-Flash, though, Grok 4.7 cannot be taken in-house.

Beyond openness and cost lies the question of who owns a model.

Mistral builds its pitch around that question. It presents itself as a European company that keeps its models under European control.¹¹

Its newest model, Large 4, came out first through Mistral’s own service, with open weights promised later.¹¹ Once they arrive, a company will be able to host it itself.

Large 4 leads everything built outside the US and China, yet seven Chinese open models do better on Artificial Analysis.¹² So what Mistral sells is the control it promises, not the best answers.

Artificial Analysis measures how good a model’s answers are. It leaves out whether a company can run the model itself, what each reply costs or who has the final say over its use. Beyond the frontier, each developer handles these questions in its own way. That difference is why the wider circle deserves a look before a company settles on one of the big three. A frontier model may still be the right choice, but only once those questions have been answered.