Open-Weight vs Closed AI Models in 2026: Token Volume, Spending Share and When Small Specialised Models Win
Open-weight models now handle most tokens on gateways like Vercel's and OpenRouter, yet closed models still take about 86% of estimated spend. What the September 2026 data shows, why Chinese open models lead, and how a 9B longevity model topped frontier systems on a domain benchmark.
Open-weight AI models are now doing most of the work on some major inference platforms, but closed models still collect most of the money. Vercel’s AI Gateway Production Index for August 2026 found open-weight models accounted for 56% of tokens but only 14% of estimated spend, while closed models took 44% of tokens and 86% of spend. On OpenRouter, open-model usage has grown from about 1 trillion tokens a week in September 2025 to roughly 80 trillion. Meanwhile, a Cell study showed a 9-billion-parameter specialised model outscoring every frontier model it tested on an aging-biology benchmark. This article explains the data, the reasons behind the split, and how to decide when an open or specialised model is the better choice.
The token versus spend gap
The New Stack summarised Vercel’s September 17 AI Gateway Production Index:
| Models (Vercel AI Gateway, August 2026) | Share of tokens | Share of estimated spend |
|---|---|---|
| Open-weight models | 56% | 14% |
| Closed-weight models | 44% | 86% |
| Anthropic, all models | n/a | 64% |
| GPT-6 Astra, first 12 days after launch | n/a | 7.7% |
Important caveats from the source: this covers gateway traffic only, and spending is estimated from labs’ published list prices, so actual bills may differ. Vercel also says its current open-weight classification is broader than in earlier reports; it previously reported open-weight share at 7% in December 2025. Other findings: the average price per token on the gateway fell 23.2% in August, the third straight monthly decline, and among teams running more than ten million tokens in both comparison months, the median team paid 7.6% less per token.
The same article reports that open-weight models made up roughly 60% of OpenRouter’s US-originating token consumption in August, mostly Chinese models such as DeepSeek V4 Pro, Kimi K3 and GLM 5.2 served from US data centres. OpenRouter made US in-region routing generally available for business and enterprise customers, processing requests inside the country or rejecting them if that cannot be guaranteed. A Deloitte survey cited in the piece found 77% of companies consider an AI solution’s country of origin in vendor selection.
Why open models lead on volume
Nathan Lambert of Interconnects published an expanded version of a congressional briefing on September 21, 2026. His main points:
- Growth. OpenRouter usage of open models grew from about 1 trillion tokens in a week of September 2025 to about 80 trillion tokens per week.
- Chinese leadership. Chinese open-weight models overtook American ones around 18 months earlier. On Hugging Face downloads, China’s lead has grown to about 1.6 billion, with 3.2 billion total downloads, twice America’s.
- Benchmarks. As of September 14, 2026, the top Chinese open models on the Artificial Analysis Intelligence Index were GLM-5.3 (45), Kimi K3 (44) and GLM-5.3-Flash (42), compared with Thinking Machines’ Inkling and Inkling Small (26 each) and Nvidia’s Nemotron 3 Ultra (23) for the US.
- Gap to the frontier. He estimates Chinese open-weight models are about 2 to 5 months behind the closed American frontier, and American open-weight models about 6 to 9 months behind.
- Where they are closest. Chinese labs are closest on tasks with clear demand such as agentic coding, and further behind on open-ended scientific tasks.
He also notes that faster release cycles and a narrower focus can flatter Chinese models slightly on public benchmarks, while stressing they are genuinely strong. Lambert estimates that fully preventing distillation from closed models would widen the gap by only one to two months.
Lambert distinguishes open-weight models (weights released under a licence) from fully open-source models that also release training code and data, such as Ai2’s Olmo. For more on that debate, see our article on the open-weight AI models debate.
Why closed models still take most of the spend
The price gap explains much of the split: closed frontier models charge far more per token, so even a minority of tokens generates most of the spend. The New Stack’s article quotes Apelogic CEO Boris Renski arguing that much of the premium buys enterprise plumbing such as identity integration, connectors and observability. It also cites CNCF executive director Jonathan Bryce describing paying ten times more for a four-month capability lead as “a very expensive form of lock-in.”
A practical reading: teams send high-volume, routine work to cheaper open models and reserve closed models for the hardest tasks or where they rely on vendor tooling. The New Stack also reported a small private-codebase benchmark (10 tasks) where the best agent setup succeeded 38.8% of the time, a reminder that even top models struggle with real-world work and that the “harness” around a model matters.
When small specialised models win
On September 17, Insilico Medicine announced a Cell cover study, conducted with Liquid AI, the Buck Institute, and Harvard Medical School and Brigham and Women’s Hospital, introducing three open resources:
- LongevityBench, a benchmark testing reasoning across clinical data, genetics, epigenetics, transcriptomics and proteomics, designed to reduce success through simple recall.
- Longevity-LLMs, five compact open-source models from 0.6 billion to 9 billion parameters, fine-tuned on aging-specific clinical and multi-omics data and built on Liquid AI’s LFM2 and Alibaba’s Qwen3 and Qwen3.5 architectures.
- Longevity Claw, an open-source agentic platform that runs multi-step research workflows.
According to Insilico, no single frontier model was best across all five data types, and performance varied significantly with question phrasing. The best specialised model, L-Qwen3.5-9B, achieved the highest overall score among all 26 systems in the study, outperforming every frontier model tested, including Gemini 3.1-Pro. Even the smallest model, at about 0.6 billion parameters, outperformed most frontier systems. Longevity Claw nominated 328 genes as potential aging targets; one, KDM1A, had been independently validated in a separate study. These are results reported by the company and its collaborators in a peer-reviewed paper.
The lesson generalises beyond biology: curated domain data and targeted fine-tuning can matter more than model size for specialised tasks, and compact open models can be run on local infrastructure, which helps with data security and cost. For a related example of turning research into usable agents, see our article on Paper2Agent and living papers.
How to choose between open and closed models
| Situation | Likely better fit |
|---|---|
| High-volume, well-defined tasks (classification, extraction, summarisation) | Open-weight or small fast models |
| Hardest reasoning or long multi-step agent tasks | Frontier closed models, verified on your tasks |
| Specialised domain with good training data | Fine-tuned compact open model |
| Strict data residency or on-premises requirements | Self-hosted open models or in-region routing |
| Heavy reliance on vendor tooling and integrations | Closed model platform |
Practical steps:
- Measure your own mix. Log tokens and cost per task type to see where spend concentrates.
- Route by task. Send routine calls to cheaper models and escalate only when needed. Fast models like Inception’s Mercury 2.5 are designed for this kind of supporting work.
- Benchmark on your data. Public leaderboards may not reflect your domain, as LongevityBench showed.
- Separate model origin from data location. Decide policy on where data is processed independently of who built the model.
- Review licences. Open-weight licences vary in commercial terms.
For a frontier model perspective from the same month, see our coverage of GPT-6 Astra.
Our analysis
The September 2026 data shows a maturing market rather than a winner-takes-all one. Open-weight models, led by Chinese labs, now carry most token volume on major gateways because they are cheap and good enough for much routine work. Closed models keep most of the spending because they still lead on the hardest tasks and come with enterprise tooling. The Insilico results add a third option: small, specialised open models that can beat frontier systems in a narrow domain. For most teams, the right strategy is a mix, guided by measured cost and quality on their own tasks.