Demand
7-Day Average of OpenRouter Daily Top-Three Token Share — Models Can Change Daily
The line is a 7-calendar-day trailing average of exact daily OpenRouter token shares for each day’s three largest models, with the three-model membership reranked and able to change every day. A point appears only when all seven daily source observations exist; this is not a fixed cohort or global market share.
7-Day Average of OpenRouter Daily Top-Three Token Share — Models Can Change Daily
For the seven days ending 2026-09-05, the average exact daily share captured by each day’s independently selected top three OpenRouter models was 35.8%.
Methodology
The source is OpenRouter’s authenticated /api/v1/datasets/rankings-daily Data API. The updater requests inclusive windows of no more than 30 days, sequentially with a delay and retry handling for rate limits. It begins on 2025-01-01, the first date available in this official dataset, and ends on the latest completed UTC day. Every complete raw response, metadata block, and SHA-256 digest is retained in a dated archive.
For each available UTC day, the API returns the 50 named models with the highest token totals and one aggregated other row representing all remaining models. The updater requires exactly 50 unique named models and exactly one other row for every returned date, parses token totals as arbitrary-precision integers, and rejects duplicate dates and malformed totals.
For each day d, the updater independently ranks the 50 named rows by total_tokens. The exact daily share is 100 × [T1(d) + T2(d) + T3(d)] / [T1(d) + … + T50(d) + A(d)], where A(d) is the aggregated other total. Integer token numerators and denominators are retained as exact fractions before smoothing. Point notes show all seven raw daily shares rounded to six decimals plus the ending day’s model identifiers and exact numerator context.
The displayed value for day d is the exact rational arithmetic mean of the seven daily shares from d−6 through d, rounded to six decimals only after averaging. It is emitted only when all seven calendar dates have validated source observations. Models are reranked independently on every underlying day, so the three-model membership can change within a seven-day window and across adjacent plotted points; the line never follows a fixed cohort.
OpenRouter returns no rows for 2025-06-15 and 2025-07-15. Each missing daily observation invalidates all seven trailing windows that include it, producing 14 explicit smoothed-series gaps: 2025-06-15 through 2025-06-21 and 2025-07-15 through 2025-07-21. The updater does not interpolate, carry values forward, or average six-day or other variable-size windows. Any new source gap, or either known gap unexpectedly gaining data, stops the update for review.
This is concentration within traffic routed through OpenRouter, which is one marketplace and API routing service. It does not represent all model usage, all API traffic, unique users, requests, revenue, capability, or market share across the broader AI ecosystem. OpenRouter’s coverage, customers, model catalog, routing behavior, and token accounting can change over time.
Frequently asked questions
What does each plotted percentage measure?
It is a seven-calendar-day trailing average of seven exact daily shares. Each daily share is the portion of represented OpenRouter tokens generated by that day’s three largest named models. The leaders are selected again every day and can change within the averaging window, so the line does not track a fixed cohort.
How is each exact daily share calculated before smoothing?
For each UTC day, the updater sums the token totals of the three largest named models and divides by the sum of all 50 named-model totals plus OpenRouter’s aggregated other total, then multiplies by 100. Integer numerator and denominator totals are summed before percentage rounding.
Are the same three models averaged over time?
No. The 50 named models are reranked independently each UTC day, and that day’s three leaders form only that day’s numerator. A seven-day point averages seven daily shares whose top-three membership may differ from day to day.
Why use a seven-day trailing average?
Daily shares can be noisy. A seven-day average makes the trend easier to read while retaining daily updates. It is a strict calendar window: the updater emits a point only when all seven daily source observations exist and never substitutes a shorter variable-size window.
What happens when a source day is missing?
The missing day remains unfilled, and every seven-day trailing window containing it is omitted. The two verified source gaps on June 15 and July 15, 2025 therefore create 14 missing smoothed endpoints, June 15–21 and July 15–21. No interpolation, carry-forward, or partial window is used.
What is included in the denominator?
Each underlying day’s denominator is the sum of all 50 named model totals returned for that day plus OpenRouter’s aggregated other row, which represents models outside the top 50. No returned traffic category is omitted.
Why does the chart begin in January 2025?
The official daily ranking dataset begins on January 1, 2025. Because the display requires seven complete calendar days, the first plotted trailing average is January 7. The chart does not estimate earlier values or use a shorter start-of-series window.
Does this represent the global AI model market?
No. It covers only tokens represented in OpenRouter’s official rankings and is shaped by OpenRouter’s customers, catalog, applications, and routing behavior. It does not measure all providers, direct API traffic, consumer products, enterprise deployments, or global market share.