Cheap models now trail the expensive ones by months, not years

A Mozilla report, covered by Ars Technica, says open models have closed most of the gap. For routine office work, paying frontier prices is usually a choice, not a necessity.

What happened

Ars Technica reports, from a Mozilla State of Open Source AI report dated 15 September 2026, that the gap between US “frontier” models (the newest, dearest closed systems) and the best Chinese open-weights models is now about 4.4 months. Open-weights means the model’s files can be run or hosted without locking you to one vendor’s chat window.

Mozilla’s view, as Ars summarises it, is that most organisations should treat open models as the default for most work. Ars cites Moonshot AI’s Kimi K3 scoring three points behind Anthropic’s Fable 5 on the Artificial Analysis Intelligence Index, at 30 per cent of the cost.

Mozilla’s chief technology officer, Raffi Krikorian, told Ars that a closed model “earns its premium in a few places: expert professional work, high-intensity retrieval, and long context,” and that paying for closed models should be “workload-specific rather than organization-specific.”

Our take

If the board is asking why you have not “done AI,” do not answer by buying the most famous model for every task. The useful split is the job, not the brand. Drafting, sorting, summarising, and first-pass admin are where cheaper models already sit close enough, and the lag is measured in months.

Keep the expensive systems for work that is expert, heavily sourced, or very long. That is also where they still fail quietly: they invent, they miss a clause, they sound sure. A person who owns the file still has to check.

Ask vendors to price and route by task. If they cannot show which jobs need the premium model, you are paying for a nameplate. Nothing here says you must host Chinese models yourself; it says you should not assume Silicon Valley list prices are the only competent option for back-office volume.