Open-Weight Models Now Handle 72% of OpenRouter Token Traffic, Up From 27% a Year Ago
In late September 2026, open-weight models handled 72% of the tokens generated by the top 50 models on OpenRouter. A year earlier, the same figure was 27%. That 45-point shift in 12 months is the largest directional move in inference market share the platform has recorded.
The data comes from an OpenRouter analysis published October 31, comparing open-weight and closed-weight model usage across its API.
The Price Gap Is Not Closing
The token economics are blunt. Across the last four weeks on OpenRouter, the token-weighted list price for open-weight models was $0.28 per million input tokens and $1.62 per million output tokens. For closed-weight models: $1.27 input and $6.36 output — approximately 4.5 times more expensive per token on a blended basis.
The gap on the output side is the key lever. Most production workloads are output-heavy, and at a 4x cost differential on the component that dominates billing, the math tilts toward open weights quickly for developers running volume.
Provider competition amplifies this. Open-weight models are served by 20 to 31 providers on the platform simultaneously, each running different hardware configurations, quantizations, and pricing strategies:
| Model | Weights | Providers | Input price range (per M tokens) |
|---|---|---|---|
| GLM-5.3 Flash | Open | 30 | $0.04 – $1.15 |
| DeepSeek V4.1 Flash 1017 | Open | 31 | $0.03 – $0.60 |
| Kimi K3 | Open | 20 | $0.24 – $4.50 |
| Claude Sonnet 5.5 | Closed | 5 | $2.00 – $2.20 |
| GPT-6 Luna | Closed | 3 | $0.05 – $0.20 |
| Gemini 3.8 Flash | Closed | 2 | $0.38 – $1.35 |
For open-weight models, the cheapest endpoint can be 10 to 30x cheaper than the most expensive one serving the same model weights. For closed models, the range compresses to a few percentage points. Price discovery on open weights is a real exercise.
The 11-Point Intelligence Gap Persists
The market share shift does not reflect a benchmark convergence. On the Artificial Analysis Intelligence Index, the best-performing open-weight model currently sits at 46.3. The best closed-weight model — Claude Opus 5.5 — scores 57.6. That 11-point gap has narrowed over the past two years but has not closed, and the top of the closed-weight leaderboard continues to be updated faster than the open-weight frontier.
For many production tasks, that gap does not matter. Routing-and-classification workloads, code completion, structured extraction, and retrieval-augmented generation do not require the top of the intelligence curve. For tasks that do — complex reasoning chains, novel research synthesis, frontier coding — closed models still hold a measurable lead.
OpenRouter’s guidance in the analysis reflects this directly: start with an open-weight model, test against your task, and escalate to a closed model for the steps that require it. The routing API supports this natively, with fallback chains and price ceilings configurable per request.
Price Inversion at the Edges
One complication: the categories are not cleanly separated by price. GPT-6 Luna is listed at $0.10 per million input tokens on OpenRouter, cheaper than any Kimi K3 provider. The cheapest endpoint for Kimi K3 (an open-weight model) costs less than Claude Sonnet 5.5 (closed), but the most expensive Kimi K3 endpoint costs about twice as much. Category membership is not a reliable price proxy at the margins.
The practical implication: routing by category label is less useful than routing by specific provider endpoint with explicit price ceilings. OpenRouter’s provider object supports a max_price ceiling that drops over-budget endpoints automatically, combined with a sort: throughput parameter that selects the fastest remaining option. The same request can route to a closed model first and fall back to an open-weight endpoint if the closed one exceeds the cap.
What the Shift Signals
The 72% token share figure is a measure of where developers send their volume runs, not their most demanding tasks. It reflects cost optimization behavior at scale rather than a verdict on capability. The models at the top of the intelligence leaderboard — Claude Fable 5.1, GPT-6 Astra, Claude Opus 5.5 — remain closed. What has changed is that a larger and larger share of inference workloads do not need those models, and open-weight options have become good enough, cheap enough, and reliably enough served to capture that portion.
The rate of change — from 27% to 72% in 12 months — suggests the portion of tasks that open-weight models can handle competitively is expanding faster than closed models can hold it.