Demand
OpenRouter Weekly Token Volume
Complete Monday–Sunday UTC weeks; all named rows plus the daily Other bucket.
Total public token volume represented in OpenRouter’s rankings, summed across every returned model and the daily Other bucket, then aggregated into complete Monday–Sunday UTC weeks. Values are trillions of prompt-plus-completion tokens and measure OpenRouter rankings traffic, not global AI usage.
OpenRouter Weekly Token Volume
For the complete week beginning 2026-08-31, OpenRouter rankings represented 115.5T tokens/week.
Methodology
The source is OpenRouter’s official authenticated GET /api/v1/datasets/rankings-daily Data API. The updater requests daily UTC observations from 2025-01-01 through 2026-08-31, the latest completed UTC date as of 2026-09-01. Requests are split into inclusive 30-day windows to keep response sizes bounded, sent sequentially at no more than 30 requests per minute, and retried on rate limits. Each complete raw JSON response is archived unchanged in parsed content, and a companion manifest preserves every response’s meta.as_of, meta.start_date, meta.end_date, meta.version, filename, and SHA-256 digest.
For each UTC date, total_tokens is parsed from its returned decimal string with arbitrary-precision integer arithmetic. The daily platform total represented by the rankings dataset is the sum across every returned row: up to 50 named models plus the single aggregated other row. OpenRouter defines total_tokens as prompt_tokens plus completion_tokens. The Other row is always included, so the calculation does not mistake the named top 50 for the complete represented total.
Each daily observation is assigned to the Monday 00:00 UTC week containing it. A weekly value is emitted only when all seven distinct dates from Monday through Sunday are present. The five-day opening fragment from January 1–5, 2025 is dropped, as is the current week until its Sunday has completed. Any week containing a missing source date is also omitted rather than interpolated, prorated, carried forward, or shown as an artificial decline.
The exact weekly integer sum is divided by 10^12 for display in trillions of tokens per week and rounded to six decimal places only in the chart package. Exact integer daily and weekly totals are retained in dailyAudit and weeklyAudit, while the raw API responses remain the reproducibility source. Duplicate date-and-model rows, negative or malformed totals, missing Other rows, more than 50 named rows, and non-seven-day plotted weeks cause the updater to fail.
The latest callout compares the week ending 2026-08-30 with the complete week four observations earlier, beginning 2026-07-27; the current ratio is 1.99×. The longer-run multiple compares the latest week with the first complete observation beginning 2025-01-06; the current ratio is 226.63×. These are ratios of exact weekly integer totals, not rounded chart values.
OpenRouter says public rankings exclude private models, private endpoints, and zero-data-retention activity. Counts use upstream providers’ tokenizers, which do not segment identical text uniformly. The series therefore measures public OpenRouter rankings token volume rather than audited total platform throughput, requests, users, revenue, compute, or global AI consumption. Changes can reflect OpenRouter customer and application mix, routing, catalog, visibility rules, and tokenizer mix as well as underlying demand.
Why this matters for the AI industry: token volume is a high-frequency usage signal that sits downstream of model releases, application adoption, inference prices, and available serving capacity. Its level and growth help contextualize demand for inference infrastructure and model access, while sharp changes can flag platform-specific shifts in workloads. Because OpenRouter is only one routing marketplace and token counts are not standardized compute units, this chart is best used alongside revenue, request, user, pricing, and infrastructure measures—not as a standalone estimate of the whole industry.
Frequently asked questions
What does one point on the chart measure?
One point is the exact sum of seven daily OpenRouter rankings token totals for a complete Monday-through-Sunday UTC week, divided by one trillion for display. Each daily total includes prompt plus completion tokens across all returned named models and the aggregated Other row.
Why is the OpenRouter Other row included?
Other contains all models outside that day’s named top 50. Excluding it would systematically understate represented OpenRouter usage and make the size of the omission vary as the long tail changes. Including every named row plus Other closes the published daily denominator.
Why are only complete seven-day weeks shown?
Weekly totals are comparable only when they cover the same duration. The chart drops January 1–5, 2025, the still-forming current week, and any week with a missing source day. It never annualizes, prorates, interpolates, or substitutes a shorter week, avoiding false cliffs at either edge or around source gaps.
Why does the chart use Monday-starting UTC weeks?
OpenRouter’s underlying observations are UTC calendar days. Assigning Monday through Sunday in UTC creates a deterministic seven-day convention that matches the referenced weekly chart and avoids mixing local time zones or trailing windows.
Why is weekly token volume relevant to the AI industry?
It is a timely proxy for the scale and growth of model inference activity visible on a major multi-model routing platform. Rising volume can indicate more application use, larger workloads, lower effective prices, or changing model behavior, all of which matter for serving capacity, accelerator demand, provider economics, and the competitive model ecosystem. It cannot identify which mechanism caused a move on its own.
Does this chart measure all OpenRouter traffic or global AI usage?
No. It measures token volume represented in OpenRouter’s public rankings dataset. OpenRouter excludes private models, private endpoints, and zero-data-retention activity from public rankings, while direct-provider traffic and other platforms are outside the dataset. The chart must not be described as audited OpenRouter throughput or global market size.
Are tokens a standardized unit across models and providers?
Not perfectly. OpenRouter reports token counts using each upstream provider’s tokenizer, and different tokenizers can split the same text differently. Tokens remain the API’s most consistent published usage unit, but comparisons over time can be affected by shifts in model and tokenizer mix.
What is the four-weeks-earlier comparison?
The latest complete week is divided by the complete weekly observation four positions earlier. For the week ending 2026-08-30, that benchmark is the week beginning 2026-07-27, producing 1.99×. The comparison uses exact integer totals before display rounding.
Can the chart identify usage by an individual model?
This total series does not need to. Summing all rows, including Other, preserves the complete represented total. A model-level history would be less complete because a model falling below the daily top 50 becomes inseparable inside Other for that date.
How can the chart be reproduced?
Use the archived official API responses for the 2025-01-01–2026-08-31 source window, verify their saved meta blocks and SHA-256 digests, parse total_tokens as integers, sum every row by date, assign each date to its Monday UTC week, retain only weeks with all seven dates, and divide exact weekly totals by 10^12. The updater path and command are recorded with this chart package.