OpenRouter Rankings: Real Usage Data, Model Popularity, and How to Choose Wisely
OpenRouter's rankings measure which AI models developers actually route traffic to, ranked by tokens processed rather than by benchmark scores — and the current leaderboard looks nothing like the marketing.
........
Rankings count prompt plus completion tokens processed through the OpenRouter API, aggregated into daily UTC buckets.
As of usage data through September 1, 2026, DeepSeek V4 Flash 0731 leads with 12.1 trillion tokens, ahead of GLM 5.3 Flash at 10T and GPT-5.6 Luna at 9.52T.
Chinese-origin models have gone from under 2% of OpenRouter traffic in late 2024 to a substantial share of the current top ten.
OpenRouter states plainly that these rankings measure adoption, not quality, and do not reflect the whole market.
Rankings data is licensed CC BY 4.0 and accessible as JSON through the Data API.
··········
WHAT THE RANKINGS ACTUALLY COUNT.
Each model is ranked by the number of tokens it processed through the OpenRouter API, counting both prompt and completion tokens.
Usage is aggregated into daily buckets in UTC and totalled per model variant, which means a free variant of a model is ranked separately from its paid variant.
Requests that a user or an app marks private are excluded before aggregation, so the totals cover public traffic only.
The date shown above the rankings is the newest usage bucket the leaderboard was computed from — not the moment the page was rendered — so it moves only when new usage data is aggregated.
··········
THE TIME WINDOWS.
The leaderboard offers four views, each covering a different trailing window.
Today, This Week, and This Month cover trailing windows of one day, seven days, and thirty days respectively, each ending with the most recent complete daily bucket.
Trending works differently: it compares the trailing seven days against the seven days before that, and ranks models by percentage change in tokens.
To stop a tiny base from producing a huge percentage swing, Trending only includes models with at least one million tokens in the current window.
New models with no prior week are listed first, up to five of them.
··········
WHO IS ACTUALLY ON TOP RIGHT NOW.
The current top of the leaderboard, from usage data through September 1, 2026, is dominated by cheap, fast models rather than frontier reasoning models.
DeepSeek V4 Flash 0731 leads with 12.1 trillion tokens processed in the week, up 4% on the prior week.
GLM 5.3 Flash from z-ai enters at number two with 10 trillion tokens, flagged as new.
GPT-5.6 Luna, OpenAI's cheapest tier, sits third at 9.52 trillion tokens with a 129% week-over-week jump — the steepest climb among established models in the top five.
Xiaomi's MiMo-V2.5 holds fourth at 7.2 trillion tokens, and Tencent's Hy3 fifth at 5.89 trillion.
··········
OpenRouter top ten by tokens processed (week ending September 1, 2026)
Rank | Model | Author | Tokens | Change |
|---|---|---|---|---|
1 | DeepSeek V4 Flash 0731 | DeepSeek | 12.1T | +4% |
2 | GLM 5.3 Flash | z-ai | 10T | new |
3 | GPT-5.6 Luna | OpenAI | 9.52T | +129% |
4 | MiMo-V2.5 | Xiaomi | 7.2T | +27% |
5 | Hy3 | Tencent | 5.89T | +18% |
6 | Hy4 preview | Tencent | 5.72T | new |
7 | DeepSeek V4 Flash 0423 | DeepSeek | 5.16T | +7% |
8 | Nemotron 3 Ultra (free) | NVIDIA | 4.63T | +14% |
9 | Ox Alpha | stealth | 4T | +83% |
10 | MiniMax M3 (free) | MiniMax | 3.77T | >999% |
··········
WHAT THE LEADERBOARD IS NOT.
OpenRouter is unusually explicit about the limits of its own data, and the caveats matter more than the ranking itself.
The rankings measure adoption, not quality. They do not rank models by accuracy, reasoning ability, or benchmark performance — those live in a separate Benchmarks section.
They describe traffic routed through OpenRouter specifically, not the whole market, and not usage on a model provider's own API. A model that most of its users reach directly through the vendor's API will be under-represented here.
Token volume is not a count of requests, users, or spend. Models differ in verbosity and tokenization, so a chattier model racks up more tokens for the same amount of work.
The practical consequence: a model can top this leaderboard because it is cheap and fast and wired into a few very high-volume apps, without being the best choice for any particular task.
··········
WHY CHEAP MODELS DOMINATE.
Look at the composition of the top ten and a pattern is obvious: almost every entry is a Flash-class, Luna-class, or free-tier model.
Two of the top ten are explicitly free variants — NVIDIA's Nemotron 3 Ultra and MiniMax M3 — and free traffic naturally accumulates tokens fast.
The rest are the cheapest tier from their respective families, which is exactly what you would expect from a metric that counts tokens rather than value delivered.
High-volume production workloads — classification, summarization, background agents, content pipelines — burn enormous token counts and are precisely the workloads that get routed to the cheapest model that clears the quality bar.
Meanwhile a team using a frontier model for a handful of high-stakes decisions a day contributes almost nothing to token volume, no matter how much value that usage produces.
··········
THE GEOGRAPHIC SHIFT.
The most striking structural change in OpenRouter's data over the past two years is where the traffic comes from.
Chinese-origin models — from DeepSeek, Xiaomi, Tencent, MiniMax, Alibaba, and Moonshot — occupy six of the current top ten positions.
According to analysis of OpenRouter's April 2026 data, Chinese-origin models had by then crossed 45% of all traffic on the platform, up from under 2% in late 2024.
The driver is not mysterious: aggressive pricing, open-weights releases, and models that land close enough to frontier quality for the workloads that generate the most tokens.
··········
HOW TO ACTUALLY USE THIS DATA.
The useful read of the leaderboard is as a shortlist, not a verdict.
A high ranking tells you a model is cheap enough, fast enough, and reliable enough that a meaningful number of developers are willing to put production traffic through it. That is real information, and it is not something a benchmark can tell you.
What it cannot tell you is whether that model is right for your specific task, at your specific quality bar.
The sensible pairing is to treat the rankings as the candidate list, then decide with the benchmark board that matches your workload — coding, reasoning, agentic — and with your own evaluation on your own data.
The question worth asking per task is what the cost of being wrong is, and then the cheapest model that clears that bar.
··········
ACCESSING THE DATA DIRECTLY.
OpenRouter publishes the underlying rankings data rather than keeping it locked to the web page.
A Data API endpoint returns the top 50 public models per day by total token usage, plus a single aggregated "other" row per day summing everything outside that top 50.
Token totals in the API are prompt tokens plus completion tokens, matching the public chart, and each row is a distinct date and model pair.
Optional filters let you slice by period (day, week, month), by output or input modality, by tool-calling activity, and by request context length.
The rankings data is licensed under CC BY 4.0, so it can be reused and republished with attribution.
··········
·····
FOLLOW US FOR MORE.
·····
·····
DATA STUDIOS
·····
