Chinese AI Models Now Trail US Frontier Models by Just 4 Months, Costing 70% Less

Open AI Models Rapidly Close Performance Gap

The performance gap between open and closed models has shrunk dramatically, with Chinese AI models now trailing Silicon Valley's frontier AI models by just 4.4 months according to 1

1

published September 15. 1

1

's 1

1

achieved a composite score on the Artificial Analysis Intelligence Index just three points behind 1

1

's 1

1

closed frontier model while costing only 30% as much. This represents a fundamental shift in the AI landscape, with 1

1

becoming viable alternatives for most organizational workloads.

Cost Efficiency Drives Organizational Adoption Strategies

The cost advantage of 1

1

is reshaping how companies deploy AI. Raffi Krikorian, 1

1

's chief technology officer, told Ars Technica that "closed earns its premium in a few places: expert professional work, high-intensity retrieval, and 1

1

." Companies like DoorDash exemplify this strategy, using Kimi for routine work while reserving Fable for difficult tasks requiring human expert time. The 2

2

surveyed roughly 1,400 developers and analyzed 2

2

traffic data, revealing eight of the top ten models by August token volume were open weights, seven built by 2

2

companies.

Benchmarking Data Reveals Narrowing Technical Advancements

1

1

from multiple sources confirms the convergence. Research nonprofit METR measures AI models by task time horizon—how long human experts need to complete jobs that models can finish with 50% reliability. The best 1

1

currently handle tasks 1.7 times longer than top open models. If open models reliably complete seven-hour jobs, closed ones manage 12-hour tasks. Within four months, open models catch up to that 12-hour threshold while closed models advance to 20-hour capabilities. 2

2

's 2

2

scored within one point of 2

2

4.7 and 4.8 on 2

2

's Terminal-Bench 2.1 while costing one-fifth per completed task when run through the same neutral harness.

Hardware Constraints and Revenue Reality Check Market Claims

Despite usage gains, 2

2

limit practical deployment of leading open models. Kimi K3's native checkpoint runs approximately 1.56TB across 96 shards, requiring at least eight GB300 GPUs with multiple nodes for production traffic—infrastructure beyond most organizations' reach. The report's hardware analysis shows the best open model fitting one server scored 52.6 points while single-GPU models reached 40, representing drops of 10 and 23 points respectively from top performers. This reality explains why closed providers captured 96% of 2

2

on OpenRouter from May-September 2025 despite lower token volume, as reported by the Linux Foundation. Organizations continue paying premiums for closed models that work out-of-the-box with compliance packaging, support, and accountability—capabilities many lack staff to replicate with open-weights models.

Future Implications and Strategic Considerations for Organizations

The accelerating pace of open model development signals a strategic inflection point. Open capability doubles every 3.9 months versus 5.5 months for closed models according to Mozilla's fitted estimate on METR data. Krikorian advised that paying for closed frontier models makes sense when deadlines land before open alternatives catch up, but routine work continuing next quarter should use open models at one-fifth the cost. Tasks requiring eight to 12 hours represent the current sweet spot where closed models justify their premium. However, caveats remain—the four-month gap measurement uses API-to-API hosted endpoints at list price, and the report's data stopped September 1. Since then, Artificial Analysis moved to index v4.3 with different evaluation sets, showing Claude Fable 5.1 at 53 versus Kimi K3 at 44 on highest effort settings. Additionally, a September 8 NSA/CISA/FBI joint advisory alleged Moonshot extracted Claude Fable 5 data to train K3 through distillation, though Mozilla notes this claim remains "asserted, and unshown." Organizations should watch whether regulatory scrutiny or technical barriers emerge as Chinese models continue narrowing the performance gap with Silicon Valley's frontier offerings.