The capability gap between US frontier AI models and the best Chinese open-weight models has narrowed to 4.4 months, according to Mozilla's State of Open Source AI report published September 15. Moonshot AI's Kimi K3 now trails Anthropic's closed Fable 5 model by just three points on the Artificial Analysis Intelligence Index — at 30% of the cost. The finding reframes the buy-versus-run calculus for every enterprise AI budget written this quarter.
Mozilla's argument is that most organizations should default to open models for routine work and reserve closed frontier models for a narrow band of tasks where the four-month head start actually pays for itself. DoorDash, cited in the report, already runs Kimi for routine workloads and reserves Fable for harder jobs that would take human experts longer to complete. That split is becoming the template.
Closed models still command a premium in a specific set of workloads: expert professional work, high-intensity retrieval, and long-context tasks. They also ship with compliance packaging, support, and accountability that many organizations lack the internal staff to reproduce around a downloaded open-weight model. Mozilla CTO Raffi Krikorian framed the decision as workload-specific rather than a wholesale platform choice.
Key facts
- 01Mozilla measures the capability gap between US frontier closed models and top Chinese open-weight models at just 4.4 months.
- 02Moonshot AI's Kimi K3 sits 3 points behind Anthropic's Fable 5 on the Artificial Analysis Intelligence Index at 30% of the cost.
- 03On Terminal-Bench 2.1 with a neutral harness, Z.ai's GLM 5.2 scored within one point of Claude Opus 4.7 and 4.8 at 5x lower per-task cost.
- 048 of the top 10 models by token volume on OpenRouter in August 2026 were open-weight.
- 05Open models earned 4% of AI model revenue vs 96% for closed models from May through September 2025, per a Linux Foundation paper.
The performance gap can be measured through METR's time-horizon methodology, which tracks how long a task an AI model can reliably complete with a 50% success rate. On that scale, the best closed model can handle a job 1.7 times as long as the best open model. Both curves are doubling on a cadence that has been accelerating.
The current window where closed models are uniquely useful is tasks running eight to 12 hours of expert human time. Nothing on the market yet handles jobs longer than 12 hours reliably. Anything under eight hours can be offloaded to an open model at a fraction of the cost, and Krikorian expects that ceiling to keep rising in step.
Harness design complicates head-to-head comparisons because closed-model providers ship custom software layers tuned to their own models. Benchmarking firm Vals AI addressed this by running every model through a neutral harness on Terminal-Bench 2.1. GLM 5.2 from Chinese lab Z.ai (Zhipu AI) landed within one point of Claude Opus 4.7 and 4.8 while costing about five times less per completed task.
That result crystallizes the trade: paying for a closed frontier model today buys roughly a four-month capability lead at five times the per-task price, and only for the eight-to-12-hour task band. For anything a business will still be doing next quarter, the open model will catch up before the invoice ages, and the model won't be the bottleneck anyway.
Usage data confirms the shift. On OpenRouter, the AI gateway that routes developer traffic across hundreds of models, eight of the top 10 models by token volume in August 2026 provide open weights. Revenue tells a different story: a Linux Foundation paper by Frank Nagle and Daniel Yue found open models earned just 4% of model revenue versus 96% for closed models between May and September 2025, though that data predates the past year's open-model surge.
“The Chinese labs are running the same playbook the Americans ran with Android—give it away, but own the ecosystem around it.”— Raffi Krikorian, Mozilla CTO
The concentration question is where the report gets pointed. The best open-weight models today overwhelmingly come from Chinese labs, and the best closed frontier models overwhelmingly come from US companies. Krikorian argued that the plural ecosystem around open-weight AI is now largely funded by Chinese capital — a plurality and a concentration at the same time — and called on US and European labs to compete in the open lane rather than ceding it.
Mozilla's proposed counterweight borrows from the Linux playbook: public compute programs funding fully open reference models (Switzerland's national compute producing Apertus is cited as a template), neutral foundations holding open protocol ground, commercial beneficiaries of commodity models paying in, and philanthropy funding evaluation and audit infrastructure. Krikorian pushed further, arguing that open weights alone aren't enough — training data, pipelines, and evaluation transparency matter for trust.
The one caveat worth naming is that open-weight releases still typically withhold training data, data pipelines, and training code, so 'open' here is narrower than the free-software model that produced Linux. Buyers relying on Kimi K3 or GLM 5.2 are trusting benchmarks and vendor-supplied documentation rather than full reproducibility.
For the AI market, the practical implication is that model choice is becoming a workload-routing decision rather than a vendor-loyalty decision. If GLM 5.2 delivers Claude-adjacent performance at one-fifth the cost on a neutral harness, closed-model pricing power collapses to the specific tasks where the four-month lead is worth 5x markup. That is a much smaller addressable market than the current $500M-plus annualized revenue lines at frontier US labs are pricing in — and it explains why the American labs are simultaneously racing on capability and lobbying on policy. The commodity layer is arriving on schedule; the question is who ends up owning the ecosystem built on top of it.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
