Demand
OpenRouter Model Author Token Share
Monthly shares of public token usage routed through OpenRouter by model author, from 2025-01-01 through 2026-08-31. Every named top-50 model row is assigned by its reviewed publisher prefix; the source’s unknown-author tail and unselected authors are Other. This is OpenRouter traffic, not global market share.
OpenRouter Model Author Token Share
- Anthropic
- DeepSeek
- OpenAI
- xAI
- MiniMax
- Xiaomi
- Alibaba / Qwen
- Z.ai
- Tencent
- Other
In 2026-08-01, models by Other represented 25.2% of public token usage routed through OpenRouter.
Methodology
The updater authenticates to the official GET /api/v1/datasets/rankings-daily endpoint and requests period=month with no modality, context, category, or language filters. It splits 2025-01-01 through the latest completed UTC day (2026-08-31) into calendar-year requests, each no longer than the documented 366-day maximum, runs requests sequentially more than two seconds apart, and retries rate limits without logging the API key.
For each month, the denominator is the exact integer sum of total_tokens over every returned row: up to 50 named public model rows plus OpenRouter’s aggregated other row. total_tokens is prompt_tokens plus completion_tokens. Every named row is grouped by the publisher prefix before the slash in model_permaslug using an explicit reviewed map to a stable model-author ID and public label.
The fixed lineup was chosen with hindsight on the inaugural 20-month snapshot ending 2026-08-26. Authors were ranked by arithmetic mean monthly token share, giving each month equal weight so later high-volume months do not dominate. The top ten authors—Google, Anthropic, DeepSeek, OpenAI, xAI, MiniMax, Xiaomi, Alibaba / Qwen, Z.ai, Tencent—were frozen. Future runs audit ranks but do not change the legend automatically.
Author identity is stable across successor model versions because all model_permaslug values sharing a mapped publisher prefix are combined. The reviewed map also supports prefix aliases resolving to one canonical author ID; in this snapshot each observed prefix maps once, including qwen → Alibaba / Qwen, x-ai → xAI, meta-llama → Meta, and mistralai → Mistral AI. A new prefix causes the updater to stop for editorial review rather than guess.
OpenRouter’s aggregated other row has no recoverable model-author identity and always remains inside chart Other. Other also includes every named top-50 row mapped to an author outside the selected ten. If a selected author has additional usage hidden in the source tail, it is not estimated or reallocated; named author shares are exact for exposed rows and the full distribution remains exhaustive.
Percentages are calculated from integer token counts and rounded to nine decimal places. The ten named model-author values are independently rounded; Other receives the residual rounding unit so every packaged month sums to exactly 100.000000000%. Exact author numerators, the source tail, unselected named tokens, and denominators are retained in periodAudit.
The latest 2026-08 bucket covers 2026-08-01 through 2026-08-31 and is complete. OpenRouter excludes private models, private endpoints, and zero-data-retention traffic. Provider tokenizers differ, so token counts are not perfectly comparable across model authors. The chart measures public traffic routed through OpenRouter and must not be described as global market share.
Frequently asked questions
Does this chart show global model-author market share?
No. It shows model-author shares within public token usage routed through OpenRouter. Private models, private endpoints, zero-data-retention traffic, direct provider usage, and activity on other platforms are outside its scope.
How are models aggregated into model authors?
Each named top-50 row uses the publisher prefix before the slash in OpenRouter’s model_permaslug. An explicit reviewed map resolves that prefix to a stable author ID and public label, so dated versions, free variants, betas, and successor models from the same mapped author are combined.
What does Other include?
Other includes OpenRouter’s aggregated other row, whose model-author identity is unknown, plus every named top-50 row from an author outside the selected ten. The unknown tail is never allocated to authors, even when it may contain usage from a selected author.
Why use these ten model authors?
They were the top ten mapped authors by arithmetic mean monthly token share across the complete 2025-01-01–2026-08-26 selection window, with each month weighted equally. This is a hindsight-selected, frozen legend: it improves historical continuity but must not be read as a contemporaneous or global top-ten ranking.
Are tokens comparable across model authors?
Only approximately. OpenRouter reports counts based on upstream tokenizers, and tokenizers can split identical text differently. The shares are exact relative to OpenRouter’s returned counts, but a token is not a perfectly standardized unit across model authors.
Is the latest month complete?
Yes. The final bucket covers the complete UTC calendar month beginning 2026-08-01 and ending 2026-08-31.
How is an unknown or new publisher prefix handled?
The updater refuses to guess. Every observed named-row prefix must exist in the reviewed map; a new prefix stops the update for editorial identification. OpenRouter’s literal aggregated other row is different: it has no prefix and remains Other by design.