ChatGPT Images 2.5 just came out, and the part I wanted to test was not whether it could make a prettier single image. I wanted to know whether it could survive a real edit chain.

I built this little rabbit sequence as a continuity test. Frame 01 set the character, watering can, pot, camera, lighting, and clay style. From there, every new frame was edited from the immediately previous accepted frame. The scene had to stay physically consistent while the rabbit raised the can, started watering, stopped, reacted, and the plant grew. I was surprised by the result for how little time went into it.

The clip is 13 accepted stills played in order over five seconds. There is no video model creating the movement. What you are seeing is the image sequence itself.

That distinction matters. A good one-off image is useful, but production work usually asks something harder: change this one thing, keep everything else, then do it again. OpenAI specifically calls out better precision editing and multi-turn consistency in Images 2.5, and that is the improvement I wanted to test here.

It still took judgement. I rejected frames when the rabbit froze, the can drifted, watering continued when it should have stopped, or the scene changed too much. The model did not remove the need to direct the sequence. It made that direction more usable.

If you are testing Images 2.5, I would test edit continuity before judging it from a one-shot prompt. Ask whether you can keep working on the same image without it slowly becoming a different image.