On the OpenRouter marketplace, where developers send their AI traffic to whichever model actually suits the task at hand, DeepSeek's text model is now handling 25.3% of all requests. OpenAI's flagship model sits at 18.6%. Anthropic, the company behind Claude, holds just 2.9% (Digital Today, 26 September 2026). After two years of headlines about which lab has built the smartest model, that is a striking number: the model getting used the most isn't the one making the most noise.

The people who put real money behind these companies have started saying so out loud. Scott Wilson, chief investment officer at Washington University's endowment, told reporters that "OpenAI and Anthropic are highly valued relative to the obligations they have taken on." Vinod Khosla, one of the most prominent investors in AI, argues the real competitive advantage isn't the model at all - it's the infrastructure underneath it: chips, data centres, and the software layer that makes a model useful day to day. Both are versions of the same question a growing number of businesses are quietly asking themselves: does the newest, priciest model actually do more for us than the one we already have?

For most UK SMEs, the honest answer is usually no - not because cheaper models are always just as capable, but because most day-to-day AI use doesn't need frontier-level intelligence. Drafting a client email, summarising a call, tidying up a spreadsheet: these need a tool that is reliable, that someone on the team has actually taken the time to learn properly, and that fits how the business already works. Switching to whichever model topped this week's benchmark chart is a way of staying busy without getting anywhere - the AI equivalent of moving your savings every time a marginally better interest rate appears in an advert.

A second piece of news this week quietly reinforces the same point. The major AI labs are all handling one of their new legal duties - labelling AI-generated content so people can tell it apart from the real thing - in noticeably different ways: Anthropic uses invisible watermarking with signed provenance data, Google DeepMind has watermarked since 2023, OpenAI relies on metadata for images having dropped an earlier text-detection tool over accuracy concerns, and Microsoft offers a visible watermark as an option (Stephenson Harwood, Neural Network, September 2026). If the world's best-resourced AI companies are still finding their own way through a single compliance requirement, a smaller business does not need to wait for perfect regulatory clarity before making a sensible choice of its own.

What actually matters, legally as well as practically, is being able to show your working. The UK's Jurisdiction Taskforce has confirmed that existing contract and tort law already applies to AI-related harm, which means a business that picked an AI tool carelessly - without testing it or checking it fit the task - could face an ordinary negligence claim, no new AI-specific legislation required. The protection isn't using the most advanced model on the market. It's being able to say: we tried it against a real piece of work, we checked it did the job, and we chose it on purpose.

None of this is an argument for standing still, or for assuming last year's tool will always be the right one. It's an argument for choosing on purpose rather than on headlines. The businesses that look sensible in two years' time won't be the ones that adopted every new model release the week it launched. They'll be the ones that picked a handful of tools that genuinely fit how they work, learned them properly, and can explain why they chose them. That is a far calmer way to keep up than trying to stay ahead of a leaderboard that reshuffles every quarter.

Go wisely.