Our daily tracker reads OpenRouter’s image-generation rankings. In the week of 5 October 2026, measured against the week of 28 September, one model moved from 7.3% of all image-generation tokens to 55.9%.
z-ai/glm-5.3-flash-20260826 gained 48.6 points of share in a week and now accounts for a clear majority of image generation on the platform, at 2,072.7M tokens.
The whole table moved
| model | share | prev | Δ |
|---|---|---|---|
z-ai/glm-5.3-flash-20260826 | 55.9% | 7.3% | +48.6 |
Others | 18.0% | 41.5% | −23.4 |
deepseek/deepseek-v4.1-flash-20260910 | 8.4% | 7.2% | +1.3 |
google/gemini-2.5-flash-lite | 5.5% | 11.3% | −5.9 |
openai/gpt-6-luna-20260922 | 3.4% | 5.3% | −1.9 |
xiaomi/mimo-v2.6-flash-20260921 | 2.3% | 4.5% | −2.1 |
google/gemini-3.8-flash-20260902 | 1.8% | 5.0% | −3.2 |
anthropic/claude-haiku-5.5-20261007 | 1.7% | 0.0% | +1.7 |
anthropic/claude-opus-5.5-20260921 | 1.7% | 0.0% | +1.7 |
meta/muse-spark-1.3-contributor-20260902 | 1.2% | 0.0% | +1.2 |
Read the second row before the first. Others — the aggregate of everything outside the top ten — fell from 41.5% to 18.0%. That is where most of GLM-5.3-Flash’s gain came from, not from the named competitors.
The named models did lose ground: Gemini 2.5 Flash Lite dropped 5.9 points, Gemini 3.8 Flash 3.2, MiMo 2.1, Luna 1.9. But together those losses total under 14 points. The long tail lost 23.4.
Why that distinction matters
A model taking share from gemini-2.5-flash-lite is a product competition. A model absorbing the long tail is something else: a large number of people who were each using a different small model stopped, at roughly the same time, and moved to the same one.
That is the signature of a price or capability threshold being crossed rather than of a gradual preference shift. The long tail of image models on an aggregator exists because dozens of models are each marginally better or cheaper for some specific use — a particular aspect ratio, a style, a language, a cost ceiling. When that tail collapses by more than half in a week, the most likely explanation is that one model became good enough and cheap enough to make the specialisation not worth the routing complexity.
The three entrants from zero are worth noting alongside it: Claude Haiku 5.5 at 1.7%, Claude Opus 5.5 at 1.7%, and Meta’s Muse Spark 1.3 contributor build at 1.2% — all three at 0.0% the week before. Haiku 5.5 was released on 7 October, so part of what the table shows is simply new availability.
Three honest caveats
The window excludes the trailing partial week. OpenRouter’s image rankings only advance when a complete week closes, so this comparison — 5 Oct against 28 Sep — is the most recent complete pair. It is not this week’s data.
Share is measured in tokens, not images. A model that produces more tokens per image, or that is being used for image editing rather than generation, moves the number differently from one producing single images. GLM-5.3-Flash is also the number two model in OpenRouter’s overall text table at 7.3% of all tokens, so it is a general-purpose model doing image work, not an image specialist.
OpenRouter is an aggregator, not the market. It reflects the behaviour of developers routing through one API, which skews toward people building products and away from people using a first-party app. Nothing here says anything about what is happening inside Midjourney, Adobe or ComfyUI.
What it is worth doing about it
If you route image generation through an aggregator, this is a reason to look at what your fallback chain actually selects. A long tail collapsing means several of the models in a typical routing config are now carrying far less traffic than they were, which over time means less attention from their providers.
And if you are deliberately using a small specialised model because it is better at your specific case, that is still a good reason — but it is now a minority position, and worth re-testing against the model two thirds of the platform has moved to.