OpenAI GPT‑Image‑1.5 tops image arenas at 1264 Elo – 20% cheaper tokens

OpenAI quietly made its biggest image move since DALL·E by shipping GPT‑Image‑1.5 across ChatGPT (Free, Go, Plus, Edu, Pro) and the API, then wiring it into a new Images tab. The model is up to 4× faster than GPT‑Image‑1 and moves to token billing with roughly 20% cheaper image input and output tokens, resolutions up to 1536×1024, and discounted cached inputs so high‑volume apps don’t get punished for retries.

On benchmarks, the jump is real: GPT‑Image‑1.5 debuts on LMArena’s text‑to‑image board at 1264 Elo, about 29 points ahead of Gemini 3 Pro Image / Nano Banana Pro, and chatgpt-image-latest lands #1 on the edit leaderboard. Artificial Analysis shows similar gains, with +147 Elo over GPT‑Image‑1 on generation and +245 on editing. Builders report much sharper, identity‑preserving edits—"change only what you ask for" is finally more than marketing—and vastly better in‑image text for UI mocks, posters, and fake newspapers.

The new ChatGPT Images surface adds reusable style chips ("3D glam doll", "Plushie", "Holiday portrait"), prompt starters, and mobile branching so image work feels more like a design tool than a chat log. Day‑0 integrations on fal, Replicate, and Figma Weave mean you can A/B it against Nano Banana Pro today—many still prefer Google for dense infographics and party scenes. One caveat: early jailbreaks and a mobile partial‑diffusion bug suggest you should layer your own safety filters in front of whatever OpenAI ships by default.