Three new models went from zero to real usage share on OpenRouter's image leaderboard in a single week
Machine.net's daily OpenRouter tracker flagged three simultaneous zero-to-nonzero jumps in its image-generation usage leaderboard for the week of September 21, all three large enough that the tracker's own significance threshold marked them 'article-worthy' rather than noise. OpenAI's openai/gpt-6-luna-20260922 opened the week at 0.0% of tracked image-generation usage and closed it at 6.3%, the single largest share gain of any model, edging past Google's gemini-2.5-flash-lite as a top-five entrant in its debut week. A second, unattributed 'stealth/space-bunny-alpha' listing gained 5.3 points in the same window, and Qwen's qwen3.7-flash-20260727 added 4.0 points — together, the three new entrants captured 15.6% of tracked image-generation usage that previously didn't exist for any of them.
The corresponding losses are concentrated, not diffuse. Google's gemini-2.5-flash-lite fell from 17.0% to 13.5% of usage, a 3.5-point drop that alone accounts for more than a fifth of what the three new entrants gained, and OpenAI's own prior-generation gpt-5.6-luna-20260709 gave up 1.7 points in what reads as internal cannibalization ahead of its successor's launch. DeepSeek's deepseek-v4.1-flash held roughly flat, losing only 0.3 points despite the reshuffling around it — evidence that this week's movement was concentrated at the newer end of the model lineup rather than a broad reordering of the whole leaderboard.
The same tracking run also caught a smaller, more targeted signal worth noting alongside it: in OpenRouter's separate text-model usage leaderboard, Anthropic's week-old anthropic/claude-opus-5.5-20260921 climbed seven rank positions to enter the top 15 overall, with token volume up 20.3% day over day as of September 29. That's a real, measurable usage data point following up on the pricing and efficiency claims machine.net covered when Opus 5.5 launched — early adoption is showing up in the same aggregator data this image-model story is drawn from, not just in the vendor's own marketing copy.
None of this is enough data to call durable share shift yet — a single week of a stealth-labeled model and two vendors' new releases is exactly the kind of spike the tracker's own methodology treats as a lead rather than a conclusion, and OpenRouter's aggregate usage is a proxy for developer traffic through one routing layer, not the industry as a whole. But three unrelated labs debuting into meaningful real usage in the same seven-day window, immediately after release, is a faster cold-start than image-generation models have typically managed on this leaderboard historically — worth machine.net checking again in a few weeks to see whether any of the three entrants holds its share or fades back toward zero.