Picture a small business owner who chose an AI tool last year, learned it, wired it into a few daily tasks, and quietly assumed that decision was done. It probably is not. The ground under AI model choice for small business has moved fast this year, and most of the movement is in the owner’s favor. It is also a little unsettling, because the safe default everyone reached for is no longer the obvious one.
Here is what shifted. On OpenRouter, a neutral marketplace that routes real production traffic across dozens of models, the share of tokens going to US models fell from around 70 percent in June 2025 to roughly 30 percent a year later, according to reporting from Office Chai drawing on OpenRouter’s own figures. Because that data reflects what people actually run in production, not what a vendor claims in a keynote, the drop is hard to wave away.
Why AI model choice for small business suddenly matters
Open-weight models, many from Chinese labs like DeepSeek, Tencent, Xiaomi and Minimax, went from about 11 percent of tokens in April 2026 to 29 percent in June, nearly tripling in two months, per OpenRouter’s own insights blog. A Yahoo Finance report put Chinese models at up to 46 percent of US enterprise token usage. The cost gap is stark. DeepSeek’s V4 Flash reportedly runs near $0.14 per million input tokens where a leading closed model sits closer to $5.00. Open models now handle close to a third of all tokens while accounting for under 4 percent of the spend.
Set the geopolitics aside for a moment. On the ground the meaning is simpler. No single company owns the thing a business depends on. The tool an owner picked is one option among many, and moving between them has never been easier. That is the opportunity. It is also where the low hum of overwhelm comes from, because more choice is more decisions, and most owners did not sign up to track a model leaderboard.
The freedom cuts both ways
Cheaper and more interchangeable is not the same as safer. MarketScale noted that US compliance teams now need a real plan for open-weight models moving through their systems, since where a model runs and where data travels are fair questions for any business handling customer information. Forbes framed the price cutting bluntly, warning it could become a race to the bottom that pressures the labs’ own economics. Both things are true at once. Prices are genuinely falling, and the ground is genuinely less settled.
This connects to two shifts covered here before: crashing AI costs and the quiet rise of vertical AI built for one job. Put together, they point at the same conclusion. Being tied to one provider is riskier than it looks.
The more useful question
The more useful question is not which model is best this quarter. That answer will change again by the next one. The better question is whether a business’s setup lets it change its mind cheaply. A tool that locks in one model, one price and one vendor is a weaker bet in a market moving this fast than a setup that treats the model as a part that can be swapped. It is also why the choice between a cloud or desktop AI agent is worth thinking through before committing.
One small step
Nobody needs to rip anything out this week. A calmer first move is to take one repeatable task already running on an AI tool, and try the same prompt on a competing model, ideally through a marketplace that bills per use so there is no contract to sign. Compare the output, the speed and the cost. Even that small test teaches something a leaderboard cannot: how much the choice actually matters for the specific work a business does.
The monopoly feeling that made AI seem like one big company’s product is fading. What replaces it is messier and, for a small business willing to stay a little flexible, mostly better. The open question is how much flexibility an owner is willing to build in now, before the next model everyone is talking about arrives.
