The most-used coding model came from a phone company

For most of the last two years the assumption was simple. If you wanted the best coding model, you paid one of the big labs. That assumption is quietly breaking.

Xiaomi''s MiMo-V2-Pro showed up on OpenRouter under the codename "Hunter Alpha," with no branding attached. It climbed to the top of the daily usage charts before anyone knew who built it. Developers weren''t picking it because of a keynote or a benchmark tweet. They were routing real work to it, and it held up.

By early April, Xiaomi''s models accounted for around 21% of all traffic on OpenRouter. That is roughly three times OpenAI''s share, and it was growing about 42% week over week. A company most people associate with phones is now moving more coding token volume than the lab that kicked off the current wave.

Why builders switched

The pitch is boring in the best way. MiMo-V2-Pro runs about $1 per million input tokens and $3 per million output, with a 1-million-token context window and strong agentic coding performance. For teams running agents that read large codebases, plan changes, and open pull requests, context size and price per task matter more than a two-point lead on a leaderboard.

When your agent burns tokens all day, a 5x price difference isn''t a rounding error. It''s the difference between shipping the feature and killing it in a cost review. Developers noticed. They moved.

The lesson isn''t "use the cheap model"

It would be easy to read this as "switch to MiMo and save money." That''s the wrong takeaway. The real signal is that model choice has stopped being a branding decision and become a measurement decision.

The teams winning here aren''t loyal to a lab. They benchmark on their own tasks, watch cost per solved task, and route accordingly. "Hunter Alpha" topped the charts anonymously precisely because nobody could lean on reputation. People judged it by output.

So the question for your team is not "which model is best." It''s "do we actually know what each model costs us per completed task, on our code, with our prompts." Most teams can''t answer that. They picked a default a year ago and never revisited it.

What to do this week

Build a small eval set from your own tickets. Ten to twenty real tasks, not toy prompts. Run your current model and two challengers through it and score them on completion, not vibes.
Track cost per solved task, not cost per token. A cheaper model that fails half the time is more expensive. A pricier model that one-shots the work can win.
Keep your integration model-agnostic. If swapping a provider means a week of refactoring, you''ve built a cage. Route through an abstraction so you can move when the numbers move.

The market is going to keep doing this. A new model tops the charts, the price floor drops again, and the "obvious" default shifts under you every few months. The teams that treat model selection as a live, measured decision keep catching those gains. The teams that hard-code a favourite keep paying for last year''s choice.

We''re here to help founders and teams design and build digital products that are built to scale with you, not slow you down. If you''re looking to build something, get in contact with us today!