Summary: As of June 2026, closed models carry roughly 63% of OpenRouter token volume, Western open-source (RoW OSS) about 23% and Chinese open-source about 14%. Chinese open-source share rose sharply after DeepSeek’s January 2025 releases, fell back to around 7% in mid-2025 and has since recovered. “Token share” is each category’s percentage of all tokens routed through OpenRouter in a week. The live charts below update weekly.
% of all tokens routed through OpenRouter, by week — Jan 2024 to present (Jan–Oct 2024 estimated, shaded; estimated before 29 Jun 2026, live model-catalog measurement after)
Live count from OpenRouter — loading…
Per million tokens (live from OpenRouter, paid models only)
Estimated revenue share (model count × avg price) vs share of total model listings — live from OpenRouter
HuggingFace top-500 text-generation models by monthly downloads (local/self-hosted) vs OpenRouter token routing (cloud API) — updated weekly
Models on OpenRouter whose output is an image or audio, not just text. This tracks what is available and what it costs, not how much it is used: OpenRouter does not publish usage for these, and most image, video and music generation happens on other platforms.
| Model | Provider | Output | Listed | Output price / M tokens |
|---|---|---|---|---|
| Loading… | ||||
Midjourney, Kling, Higgsfield, Runway and HeyGen do not publish usage, so this tracks three public proxies for interest: App Store ratings, Google search interest and website popularity rank. They show relative momentum, not how many people generate or how much. Zing stores each weekly snapshot itself, so the history does not depend on the sources keeping it.
| Service | iOS ratings (US) | Change | Rating | Search interest | Site rank |
|---|---|---|---|---|---|
| Loading… | |||||
Historical data is sourced from OpenRouter's published "Token Share by Source Type and Origin" chart (openrouter.ai), covering weekly token share from November 2024 onwards. The Jan–Oct 2024 portion of the chart (shaded) is Zing’s estimate, not published data. Data points reflect the percentage of all tokens routed through OpenRouter attributed to each category.
Current snapshot (model counts and pricing) is fetched live from the OpenRouter public API (/api/v1/models) at page load. Models are categorised by provider organisation:
Definitions. Token share is a category’s percentage of all tokens routed through OpenRouter in a given week. Revenue share is an estimate, not reported revenue: model count × current average input price per million tokens (USD, as listed by OpenRouter), so it is a rough proxy only.
Limitations. OpenRouter users skew towards developers and API-first teams, so its mix may not represent the wider AI market. The Local vs Cloud panel compares HuggingFace monthly downloads (self-hosted) with OpenRouter token volume (cloud API); these are different measures and are indicative, not like-for-like. The Jan–Oct 2024 history is estimated.
Image & audio generation. Models are counted as generative when OpenRouter lists an image or audio output modality for them; its own router entries are excluded. Prices are the listed image-output or audio-output rate per million output tokens. This shows availability and price, not usage. Weekly history is stored by Zing from our own snapshots, beginning with the first one shown above.
Generative media interest. Midjourney, Kling, Higgsfield, Runway and HeyGen publish no usage figures. Zing records weekly App Store rating counts (Apple’s public lookup, US), Google Trends interest and Tranco website rank, and keeps its own copy of each. These indicate relative interest and momentum, not users or generations.
Average pricing excludes free-tier variants (:free suffix) to avoid distorting the paid market comparison.
Spotted something wrong, or have a source we should include? Let us know.