Update: ChatGPT Images 2.5 adds @Sketch and rolls out to Codex
Sketch a reference and it generates from your drawing; comment on any region to revise it. Latency's down 50%, and Codex agents now drive real lab gear.
Copy markdown
Draw it, don't describe it
Type @Sketch in ChatGPT to hand Images 2.5 a rough drawing as a visual reference instead of a wall of prompt text. Built-in templates cover flyers and product shots, so laying out a real design beats coaxing one out of words. Live now on desktop, mobile and web across every tier.
Comment on the image to fix it
In-image commenting lets you drop a note on one region and get just that part re-rendered, leaving the rest of the composition and your brand untouched. With generation latency cut up to 50% versus Images 2.0, the edit-and-check loop is finally fast enough to iterate live.
It's inside Codex now
Images 2.5 shipped to ChatGPT, ChatGPT Work and Codex on every plan, so image generation lives in the same agent that writes your code. The API runs on GPT-Image-2.5 Flare (standard, ~50% lower latency) and Sunburst (precise edits), billed around $30 per 1M image-output tokens.
Elsewhere: Codex starts running the lab
OpenAI showed GPT-5.6 Sol plus Codex autonomously calibrating superconducting qubits, wiring an agent into lab software to run measurements overnight while a researcher checks in remotely. It needed help on weak, noisy signals, but the pattern is portable: point an agent at your own instrument or control software and let it drive the loop.