Exclusive: Paying for frontier AI models buys 4-month head start at 5x the cost

Mozilla’s State of Open Source AI report finds the performance gap between closed “frontier” models and top open-weights models has narrowed to roughly 4.4 months. The study shows leading open models now approach closed-model capability at a fraction of the price, prompting many firms to use open models for routine tasks and reserve paid models for a small band of high-intensity work.

By AI NewsroomPublished about 1 hour agoUpdated about 1 hour ago0 views

Exclusive: Paying for frontier AI models buys 4-month head start at 5x the cost

Why It Matters

This shift changes procurement and deployment choices for organizations: cheaper open models can handle most routine workloads now, while closed models retain value only for time-sensitive or expert-level tasks. That trade-off will affect costs, vendor strategy, and where AI revenue flows in the market.

Key Facts

Report: Mozillas State of Open Source AI report published September 15, 2026 (shared with Ars prior to publication).
Performance gap: Mozilla estimates the gap between frontier closed models and top open-weights models is 4.4 months.
Kimi K3 vs Fable 5: Moonshot AIs Kimi K3 scores three points below Anthropics Fable 5 on the Artificial Analysis Intelligence Index while costing about 30% of Fable 5.
VALS AI benchmark: When run on the same neutral harness (Terminal-Bench 2.1), Z.ais GLM 5.2 scored within a point of Anthropics Claude Opus 4.7/4.8 while costing about five times less per completed task.
Head start and cost trade-off: Paying for closed frontier models buys roughly a four-month lead at about five times the per-task cost for tasks in the 8–12-hour range.

Mozillas latest State of Open Source AI report, shared with Ars ahead of its September 15 release, finds that the capability gap between the top closed frontier models from U.S. labs and the best open-weights models has shrunk to about 4.4 months. The report highlights a leading open model, Moonshot AIs Kimi K3, which posts a composite score only marginally below Anthropics Fable 5 while costing roughly 30% as much. That narrowing margin helps explain why many companies are shifting routine workloads to cheaper open models. Researchers use several metrics to compare models. The nonprofit METR measures an AI models time horizon by the length of tasks it can reliably complete at a 50% success rate, and finds the best closed model can handle jobs about 1.7 times longer than the best open model. In practical terms, Mozillas analysis shows open models close the gap quickly: a task an open model can do in seven hours is within reach of the closed frontier at about 12 hours, and open models are expected to match the closed frontier for those jobs within roughly four months. That leaves a narrow band — tasks roughly eight to 12 hours in duration — where closed models still offer unique capability. Organizations still pay premiums for closed models because they arrive ready to use with vendor-provided compliance packaging, support, and accountability, and many firms lack the staff to operate open-weights models at scale. Benchmarks are also sensitive to the software harness used: custom harnesses built by labs can boost performance, so independent evaluations try to neutralize that advantage. When Vals AI ran models on a single neutral harness (Terminal-Bench 2.1), Z.ais GLM 5.2 performed within a point of Anthropics Claude Opus versions and did so at about one-fifth the per-task cost, implying that buying closed access currently buys a discrete time advantage at a significant price premium. Market signals back up the technical trends. Open-weights models are dominant by usage on marketplaces such as OpenRouter — eight of the top 10 models by token volume in August 2026 supplied open weights — yet revenue remains concentrated in closed models: a Linux Foundation paper covering May–September 2025 estimated open models earned about 4% of total model revenue versus 96% for closed models. Companies such as DoorDash are already adopting hybrid approaches, using open models for routine work and reserving closed frontier models for the higher-complexity tasks that still demand the extra head start closed models provide.