Six of the Top Ten Image Models Are One Family

OpenRouter's image leaderboard for the week ending October 26 reads like a product catalogue rather than a competition. Six of the top ten slots belong to one model family. The top slot has 1.57 million requests. The tenth has 253,000.

The list, in order: Nano Banana 2 (Gemini 3.1 Flash Image) at 1.57M, up 13% week over week; GPT Image 2.5 Sunburst at 1.25M, up 6%; Meta's Muse Image at 824K, up 23%; GPT Image 2.5 Flare at 793K, up 26%; Nano Banana 2 Lite at 726K, up 18%; GPT Image 2 at 633K, down 4%; Nano Banana at 450K, flat; Nano Banana Pro at 417K, up 5%; Nano Banana 2.1 at 383K, up 21%; and Nano Banana 2 in its preview form at 253K, down 14%.

By author, the week starting October 19 splits like this: Google 43.3% of image requests, OpenAI 33.2%, Meta 7.8%, Black Forest Labs 4.8%, ByteDance Seed 3.9%, Recraft 1.8%, SpaceXAI 1.8%, Tencent 1.1%, Qwen 0.9%.

Two companies take three quarters of the requests. That is the story, and it is a less interesting story than the one hiding in the tier structure.

Five tiers of the same model

Google's entries are not five competing models. They are one model lineage sliced by price and capability: Nano Banana 2 at the top, Nano Banana 2 Lite below it, Nano Banana Pro above that in quality, Nano Banana 2.1 as a mid tier, and the plain Nano Banana from the previous generation still pulling 450,000 requests a week.

OpenAI does the same thing with two entries: Sunburst and Flare are both GPT Image 2.5, positioned at different speeds and prices, and together they hold second and fourth place.

This is not a trick. Tiering is how you serve a market where one customer needs a 4K hero image and another needs 900 thumbnails before lunch. But it does mean the leaderboard is measuring something other than model quality. A vendor with four tiers occupies four slots with one model. A vendor with one great model occupies one slot.

The clearest evidence that these are distribution decisions rather than quality decisions is the tenth entry: Nano Banana 2 in its preview form, still receiving 253,000 requests a week, nine weeks after the production version shipped. Nobody chooses a preview model on purpose. Someone's pipeline was built against that model string and nobody has touched it since. Request volume measures inertia as much as preference.

The one number that is not about Google or OpenAI

Meta's Muse Image at 824K is the only entry in the top four that is not from the two leaders, and it grew 23% in a week, faster than either of them. Meta's pitch for Muse has been that it plans a composition before rendering, which is the same architectural bet Canva made with Craft this week. Two companies arriving at planning-then-rendering in the same month is worth noticing, though it may say more about where the easy gains are than about a shared insight.

The other number that stands out is Black Forest Labs at 4.8%, up 83% week over week. That is the fastest growth rate on the board by a wide margin, and it is still under 5%. A fast riser at low volume is usually a sign of one large customer onboarding, not a market shift. Worth watching, not worth concluding from.

The tail of the table is where the interesting absences are. Recraft at 1.8% and up 29%, Tencent at 1.1%, Qwen at 0.9%, SpaceXAI at 1.8%. These are not failures. They are models with a specific job: Recraft has always been the choice for vector and brand-consistent assets, Tencent's image models matter inside Chinese-language products, and Qwen's open weights get run on the customer's own hardware, which means the requests never appear here at all. That last point is the biggest blind spot in any router leaderboard. A model distributed as open weights can be extremely widely used and invisible in this data, because the people running it are running it themselves. The request counts measure hosted adoption, which is a proxy for developer convenience, not for total deployment.

What request counts do not tell you

This is a request count, not revenue and not tokens. A request on a Lite tier may cost a fraction of a request on a Pro tier, so a vendor with a cheap default can top the chart while earning less per call than the vendor in third place. OpenRouter also reports a separate spend-share view for exactly this reason, and the two do not always agree.

There is a second gap, which is that OpenRouter is a router. It sees the traffic that developers route through it, which skews toward people building products, agents, and batch jobs. It does not see the traffic that goes direct to a vendor's own app, and it does not see the enormous consumer volume inside tools like Canva or Photoshop. Google leading at 43.3% here is a statement about developer defaults. It is not a statement about which model makes the most pictures on earth.

What a builder should take from it

If you are choosing an image model this week, the leaderboard is a reasonable proxy for three things: which models have the most examples and community answers, which ones are unlikely to be deprecated without notice, and which ones your future hires have probably already used.

It is a poor proxy for the thing you actually care about, which is whether the model handles your specific images. A 1.57 million request lead tells you nothing about how it renders your product's packaging.

The more useful pattern in the data is the tiering itself. If two vendors are now shipping four and five price points of the same underlying model, then price is negotiable in a way it was not a year ago, and the correct move is to test the cheap tier first. Most production work does not need the Pro model. It needs the Lite model and a decent upscale.

And the preview entry at number ten is a small reminder for anyone maintaining a pipeline: the model you hardcoded in March is still running, still billing, and probably still the preview build. It costs an afternoon to check.

The cost of a tier nobody thinks about

There is a quieter cost hidden in this structure. Every extra tier is an extra migration somebody will eventually have to do, and the tiers age at different speeds. A vendor shipping five price points of one lineage retires the bottom ones first, because those are the least defended commercially. The preview build sitting at 253,000 requests is not a problem today. It becomes one the day it stops accepting calls.

The teams that handle this well do one boring thing: they keep a short list of which model string each product uses, who owns it, and what the fallback is. That is not interesting work. It is the difference between a Tuesday afternoon and a Tuesday outage.