What OpenRouter rankings showed Oct 11–13, 2026

OpenRouter publishes a weekly usage leaderboard. Because the window rolls, two snapshots taken 48 hours apart can disagree about the same week's ranking, and neither is wrong. That makes it a useful case study in how easily a leaderboard gets misread.

Below are two windows, captured on Oct 13 and Oct 15, 2026, covering usage data through Oct 11 and Oct 13 respectively. Both are "All models / This Week", and "This Week" on that page means a trailing seven days, not a calendar week.

The two windows, side by side

Window A — usage data through Oct 11, 2026. Captured Oct 13.

Window B — usage data through Oct 13, 2026. Captured Oct 15.

What actually changed between the two captures

The headline did not change: DeepSeek V4.1 Flash stayed first in both windows. What changed is more interesting than the ranking itself.

Step 5 Preview moved from fourth to second, the largest position change in the table. Its token volume rose from 8.84T to 15.1T while its label still reads "new", which means the model was not in the comparison base a week earlier.

Claude Opus 5.5 remained in the doubling band, +104% then +93%, and its absolute volume rose from 5.02T to 5.64T. DeepSeek V4.1 Flash, by contrast, kept first place while its growth rate cooled from +60% to +37%; its absolute volume actually declined slightly, from 41T to 40.5T. A model can hold rank one and lose momentum at the same time.

At the bottom of the table, Space Bunny Alpha fell out of the top ten entirely. Its previous entry was already marked -89%. GLM 5.3 entered at tenth. This is the kind of churn that a single snapshot hides completely: if you only look at one capture, you see a stable list; if you look at two, you see the list is a rotating door below the top few.

The structural observation recorded with the second window is worth repeating precisely: six of the top ten authors are identifiable as Chinese labs, and the top five were all Chinese labs — the first time that had happened across the observed windows.

Why the ratio moved, and why that matters less than it looks

The top-ten volume ratio between Chinese and US labs was about 4.3x in the Oct 11 window and about 4.06x in the Oct 13 window. That looks like a narrowing gap. It is not necessarily one.

The denominator is a sum of ten ranked entries. When a stealth model drops out and is replaced by another Chinese-lab model, the numerator and denominator both move for reasons that have nothing to do with any lab's actual growth. A ratio computed over a fixed-size leaderboard is a composition effect, not a market share.

This is the single most common mistake people make with these rankings: treating top-ten aggregates as if they were total market volume. They are not. The page offers a separate market-share view by author, which is the correct instrument for that question, and it was not used in either capture.

The spend side is a different dataset

The Oct 13 capture also recorded a spend view, which had changed shape: the older classification categories were gone, replaced by "Top models by task". In the Workflow Execution category, one model family accounted for 20.2% of spend, and four of that category's top ten were Chinese labs.

That number cannot be compared with anything from the previous classification scheme. The category definitions changed, so a before-and-after comparison would measure the taxonomy, not the market. The Oct 15 capture did not refresh spend at all and explicitly refused to reuse the old figure.

That refusal is the most useful thing in the whole exercise. A number that cannot be refreshed should be marked stale, not carried forward because it looks authoritative.

Method notes you should carry with you

Three rules make this data usable rather than decorative.

First, "This Week" is a trailing seven-day window. Consecutive captures therefore overlap heavily. They are not independent samples, and treating them as weekly growth is wrong.

Second, attribution matters. The figures come from openrouter.ai/rankings, All models / This week, with usage data through the dates stated above, published under CC BY 4.0. Any reuse should say so, including the exact "through" date, because the same URL returns different numbers a day later.

Third, a rank is not a market. Rank order is stable at the top and noisy below it, and aggregate ratios are sensitive to who enters and leaves the list. If you need a market claim, use a market-share view; if you need a momentum claim, compare the same model across windows and say which windows.

Three rules for reading this data

1. Two captures, two days apart, can disagree about the same week. Say which capture you used and when it was pulled. Rankings are a snapshot of a moving window, not a fixed fact, and blending two snapshots produces a number that never existed.
2. "Top five all Chinese labs" is a ranking, not a market-share claim. Ranking measures position inside a window; share measures volume against a total. Reporting one as the other is the most common way a good chart becomes a wrong sentence.
3. Cite the capture date and the raw file, not the chart. Spend figures are not comparable across captures — the window shifts, the reporting lags, and the currency moves. Anyone reproducing your claim needs the file, not the picture of it.
4. Treat the ranking as a question, not a conclusion. The useful output of a leaderboard is the follow-up question it makes unavoidable; the useless output is a headline that closes the topic.