OpenAI's GPT Image 2.5 Flare leads the Artificial Analysis image arena at 1188 Elo. It bills $210.72 per 1,000 images.
Meta's Muse Image sits seventh at 1111 Elo — 77 points back, about 6% — and bills $10.00 per 1,000. That is a 21× price difference for a gap most buyers would struggle to see in a blind test. Which is precisely what the arena is: a blind test.
The ladder
Three things the table shows
The premium is not linear. Going from Muse Image to GPT Image 2.5 Flare costs 21× more and buys 6.9% more Elo. Going from Muse Image to MAI-Image-2.6 costs 3.9× more and buys 3.2%. Where that stops being worth it depends entirely on whether a human is looking at every output.
The fastest model is not the most expensive. Grok Imagine Image 2.0 returns in 5.1 seconds — the quickest in the top ten — at $60 per 1,000. GPT Image 2.5 Sunburst takes 106.6 seconds, twenty times longer, for 28 more Elo. If you are generating inside a user-facing flow, that difference matters more than the ranking does.
Microsoft is the value pick at the top end. MAI-Image-2.6 is 41 Elo behind the leader at 18% of the price. Nothing else in the top six comes close on that ratio.
What changed this month
Until recently this board was effectively an OpenAI product line with company. The last month added Grok Imagine Image 2.0, MAI-Image-2.6-Flash and Muse Image, and Google's Nano Banana 2 arrived earlier in the year. Four vendors now hold places between 1100 and 1160 — a band that did not exist as a competitive market in early 2026.
Elo, prices and generation times are from the Artificial Analysis image arena and its model pages, read 21 September 2026. Elo is blind human preference, not a capability benchmark: it says which image people picked, not which model is more capable. Every figure carries its source and date on the model pages.
