What are the hottest AI models on OpenRouter this week? GPT, Claude, DeepSeek, and Gemini usage and growth trends

Ask which of GPT, Claude, DeepSeek, and Gemini ran hottest this week, and the leaderboard will not hand you one name. Fold each family into a single line and a model swap looks like a collapse, while a tiny base looks like a takeover.

What follows lines up what OpenRouter’s week view actually counts, then splits the four families into the snapshots inside the top 20, the token totals, and why the percentage and the absolute change keep telling different stories.

What this board counts

OpenRouter’s model board sums prompt tokens and completion tokens routed through OpenRouter, by model. “This week” is the trailing seven days ending on the latest complete UTC day. The change column compares that window with the seven days before it. This fetch is 09:12 China Standard Time on 30 September 2026, which is 2026-09-30T01:12:31Z. The window closes on 29 September.

Keep three limits next to every table below:

• This is not global share. First-party traffic in ChatGPT, claude.ai, and the Gemini app is absent.
• A token is not the same unit across vendors. OpenRouter says counts come from each provider’s own tokenizer. The “T” figures below match the page. They are not one ruler.
• Family totals include only the top 20. A model that fell out is not in the sum. One Google row does not mean Google has only one model in production.

This week’s top 10: the leader is outside the four

The single-model leader is stealth’s Space Bunny Alpha. The highest of the four families is DeepSeek V4.1 Flash. No Claude model and no Gemini model made the top 10.

The top 10 already shows the split that follows. DeepSeek has three snapshots on the board, and the two older Flash rows are down. OpenAI has two Lunas: one slightly down, one printed as >999%.

Summed up: largest and fastest-growing are different questions

The table adds only the rows from these four families that sit inside this week’s top 20. Prior-week volume is backed out from this week’s tokens and the change ratio, so the absolute delta is visible. For rows printed as >999%, the prior week is tiny; the backed-out figure is kept to 0.01T.

DeepSeek is still the largest block of the four. The largest absolute gain this week is GPT, about 3.68T. Claude has the steepest percentage and still about a tenth of DeepSeek’s top-20 total. Gemini is the only one of the four that shrank, and it is a single model.

DeepSeek: the family is up because a new snapshot replaced older ones

The page rounds V4.1 Flash to 22T and +24%. On the raw counts it added about 4.2T in the week. The two older V4 Flash snapshots gave back about 1.6T combined. The family’s net +2.57T is what remains after that handoff.

If a route is pinned to 0731 or 0423, “DeepSeek usage” on that pin can fall while the brand is still rising on this board. That is a migration, not an exit.

GPT: 5.6 Luna is still the base, the gain is GPT-6 Luna

GPT-5.6 Luna is still OpenAI’s largest row here, only slightly down. The more expensive GPT-5.6 Sol fell further and sits at 16. The family’s +35% is almost entirely GPT-6 Luna: about 0.07T the week before, 4.38T this week. The page prints >999% because the denominator is small, not because the slope is something you can extend.

“GPT surged” and “the Luna you already route is down” can both be true. A default-model change follows the snapshot you pinned, not the family total.

Claude: doubling did not move its place on this board

Neither row is in the top 10. Opus 5.5 was about 0.03T the week before and 1.69T this week, which is nearly the whole +105%. Sonnet 5 lost about 0.06T, close to flat. Combined, 3.12T is about a tenth of DeepSeek’s top-20 total.

If the workload is already Sonnet-class, this table does not say “move everything to Opus.” It also does not describe usage on claude.ai.

Gemini: only 3.8 Flash made the top 20, and it slipped

Google’s only top-20 row this week is Gemini 3.8 Flash, rank 14, 2.05T, down 7.5%, about 0.17T less than the week before. No second Gemini model appears in this top 20.

That is a quiet week on the router. It is not a weekly report for the Gemini app. How much Google traffic sits outside the top 20 is a question this table leaves open on purpose.

Three easy misreads

1. Treating >999% as a trend slope. GPT-6 Luna and Opus 5.5 both sit on a prior week under 0.1T. Next week the percentage can fall hard even if the volume is still large, as soon as it stops multiplying.
2. Reading an older snapshot’s decline as a brand failure. DeepSeek’s two older Flash rows are down. V4.1 Flash added about 4.2T. A route pinned to a snapshot ID is watching a migration, not a vanished share.
3. Subtracting this board from a table published months ago. The June pricing and usage note on this site covers a different week and a different set of snapshots. Sitewide weekly volume and the leading model do not line up with today’s board. Do not subtract 3.43T from 22T and call the gap “growth.”

How to use the table when you route

• Match the snapshot before you match the vendor. DeepSeek’s net gain is what is left after V4.1 Flash covers the older Flash rows. If the default is still 0731 or 0423, check whether the newer snapshot in the same family is taking the traffic.
• GPT-5.6 Luna is still OpenAI’s base on this board. GPT-6 Luna is this week’s increment. It is a comparison candidate, not a reason to move every call because of one percentage. Sol shrank this week.
• Sonnet 5 is nearly flat. Opus 5.5’s gain comes off a tiny base. A Sonnet-class workload does not get a migration order from this table.
• Gemini 3.8 Flash slipped, and it is still the only Google model in the top 20. Whether it stays on the route depends on your own task hit rate, not on this week’s −7.5%.

Why price pulls volume apart is in the OpenRouter pricing note. How a token is counted is in the token and unit-price comparison. An earlier week on the routing layer is in the previous usage piece. Do not subtract that piece from this one. The window and the snapshots differ.

Questions

Which model led the week? The page’s first row is Space Bunny Alpha, 23.4T, marked new. Among the four families, DeepSeek V4.1 Flash is largest, about 22T on the page, +24%.

Did GPT rise or fall? Split it. GPT-5.6 Luna is −4%, GPT-5.6 Sol is about −14%, and GPT-6 Luna went from about 0.07T to 4.38T. The three together are 14.33T, about 3.68T more than the week before.

Did Claude’s doubling catch DeepSeek? No. Claude’s top-20 total is 3.12T. DeepSeek’s is 32.55T. The doubling is almost all Opus 5.5 off a small base. Sonnet 5 is close to flat.

What about Gemini? Only Gemini 3.8 Flash is in the top 20, at 2.05T, −7.5%. There is no second Google model.

Does this board stand in for ChatGPT, the Claude site, or the Gemini app? No. It counts tokens routed through OpenRouter.