Tracked sources, refreshed daily

OpenRouter: weekly token totals and trailing-30-day per-model totals from the public data behind openrouter.ai/rankings, 91 days of daily per-model tokens from each lab's page (e.g. openrouter.ai/anthropic), and model names, launch dates and prices from its public model API. opencode: daily totals by model lab from opencode.ai/data, including cached input tokens. Vercel AI Gateway: daily share of tokens and spend per lab from Vercel's open leaderboard export, back to October 2025. These sites publish rolling windows, so Token Tracker saves every day's pull and builds its own history.

Company disclosures, curated

Statements from earnings calls, keynotes, filings and announcements, each tagged as the company's own words or a press report. Volumes are converted to tokens per day: per-minute ×1,440, per-month ×12 ÷ 365.25, totals ÷ the days they cover, and words ×1.33 (Epoch AI's convention). Where a statement names only a month, the point sits mid-month.

Is it speeding up?

Growth is compared period against period. Between disclosures, change is converted to an average monthly rate; statements under 60 days apart are skipped rather than annualized from a few weeks. Platform traffic compares the last 30 days with the 30 before (14 days for opencode). A measure is speeding up when its latest period's log growth beats the previous period's by more than 15%, slowing when it trails by more than 15%, and shrinking when volume fell. Share compares now with earlier: 90 days on Vercel, the first complete week on OpenRouter, four weeks on opencode. Labs under 100B OpenRouter tokens a month, or under 1% of opencode's traffic, are shown there but not judged.

Uses, apps, models and money

OpenRouter also publishes what its tokens are used for (tokens and spend by task over 30 days), its top apps, the median cost of a coding-agent session by harness and model, the week's fastest-climbing models, and weekly tool calls and images. opencode publishes weekly user retention by model (the share of a week's users who come back the next week), token share by country, and cost per session. Vercel adds each lab's share of images and videos generated. Capex is cash spent on property and equipment from Alphabet, Microsoft, Amazon, Meta and Oracle filings (via Quartr), by calendar quarter; Alphabet, Meta and Oracle file year-to-date figures, which are split into single quarters. Prices are OpenRouter list prices, blended three input tokens to one output.

Drivers and revenue

"Tokens added" is the change in a lab's 30-day OpenRouter total. Free or discounted tiers are counted among each lab's top models. Revenue per token divides a lab's share of Vercel spend by its share of Vercel tokens. OpenRouter's doubling time is a log-linear fit over the last 26 complete weeks. T = trillion (10¹²), Q = quadrillion (10¹⁵). Nothing here is investment advice.