Open-weight AI models are rapidly catching up with U.S. big tech’s closed, cutting-edge models, changing how companies use AI. They use low-cost open models for routine, repetitive tasks and deploy higher-cost closed models only for difficult work.

According to IT outlet Ars Technica on Sept. 15 local time, Mozilla said in its latest report that the performance gap between open models and closed, cutting-edge models has narrowed to about 4.4 months.

The key is efficiency relative to cost, rather than performance alone. Moonshot AI’s open model Kimi K3 trailed Anthropic’s closed, cutting-edge model Fable 5 by 3 points on a composite AI performance metric, but cost only about 30 percent as much.

Raffi Krikorian (라피 크리코리안), Mozilla’s chief technology officer, pointed to areas where closed models still have a clearer edge, including expert-level work, high-intensity search and long-context processing. He said companies should choose models by weighing costs by task, rather than using expensive closed models for every job.

Mozilla explained the gap between models by measuring how long it would take a human expert to do the work. It said the current best closed models can reliably handle tasks about 1.7 times longer than the best open models. If an open model can handle a 7-hour task, that means a closed model can handle tasks of about 12 hours.

But that gap is also shrinking quickly. Mozilla forecast that in about 4 months, open models could handle tasks at the level currently performed by closed models.

From a company’s perspective, the range of tasks that truly require closed models is narrower than many think. Both types can handle work that would take a human expert 8 hours or less, leaving wide room to use cheaper open models. For high-difficulty tasks requiring about 8 to 12 hours, closed models’ performance edge can justify the cost.

Companies are already adopting this approach. Food delivery platform DoorDash splits usage by using Kimi for repetitive work and Fable for more complex tasks. It pays for closed models only for work that requires fast results until open models catch up in performance.

Other evaluations reached similar results. In ValisAI’s Terminal-Bench 2.1, Chinese Z.ai’s open-weight model GLM 5.2 came within 1 point of Anthropic’s Claude Opus 4.7 and 4.8. The per-task cost was about one-fifth that of closed models.

Usage is also shifting. Eight of the top 10 models by token usage on OpenRouter in August 2026 were open-weight models. But revenue is still centred on closed models. A Linux Foundation paper said that from May to September 2025, open models accounted for 4 percent of revenue and closed models 96 percent.

Mozilla also pointed to the problem of regional concentration in the open-model ecosystem. With many major open models currently tied to China, it said research institutions in the United States and Europe also need to join the race to develop open models. It presented as an alternative an approach involving public compute, neutral foundations, companies and philanthropic funding.

In the end, companies’ AI model selection is moving away from a race to pick a single best-performing model. A growing trend is to divide roles, using cost-effective open models for daily work and closed models for demanding tasks where performance matters. As the performance gap for open models continues to narrow, the performance premium companies must pay for closed models could gradually shrink.