news

Ideogram 4.5 ships a no-drift editing claim that only Ideogram has scored

The model went live on Ideogram's app and API around September 30 with Runway, Pika, Leonardo and Picsart as partners, and its central consistency claim rests on a comparison Ideogram ran and published itself.

By Daniel Ochoa · 5 October 2026

Ideogram 4.5 is sold on a single promise: edit the same image over and over and the parts you did not touch stay exactly where they were. That is a measurable property, and nobody outside Ideogram has measured it. The company's supporting evidence is a comparison it ran and published itself, applying the same sequence of edits to its model and to GPT Image 2.5 Sunburst, Nano Banana Pro and Nano Banana 2, then pointing at where the rivals degrade.

A precision editor, not a new generation engine

The release is narrower than a version number bump suggests. Ideogram 4.5 targets the workflow problem where every pass through a model nudges pixels, shifts colors and layers on artifacts, so the tenth revision of a product shot or a poster no longer matches the first. Named use cases are product photography, interior design, photo restoration and editing text inside an existing image, which is the capability Ideogram built its reputation on. Reported launch dates differ by a day, with the model logged as released September 30 and most coverage appearing October 1.

Four price tiers and a planned open-weight drop

The Decoder reports four quality tiers running from 0.8 to 22 cents per image, all at native 2K resolution, available on Ideogram's platform and through its API. Launch partners named in that reporting are Picsart, Runway, Pika and Leonardo AI. Recraft says the model is headed to Recraft Studio but has not shipped there yet, and Ideogram says an open-weight release is coming without giving a date.

That last part has precedent. Ideogram 4.0, released June 3, 2026, shipped as an open-weight 9.3B-parameter diffusion transformer with a vision-language text encoder, structured JSON prompting, bounding-box layout control and color palette conditioning. If 4.5 follows, the drift claim becomes something anyone with GPUs can test directly rather than something that has to be taken on trust.

Multi-turn consistency is the one thing public leaderboards do not measure

Every major public image ranking scores single-shot output. Voters see one prompt and one result, or at most a before and after for a single edit, and pick a winner. The image arena that refreshed with 11,311 blind votes across 10 models works that way, and so does essentially every comparable board. None of them run an edit, feed the result back in, run another edit, and score how far the untouched regions have moved after five or ten rounds.

That gap is why a vendor-run comparison is currently the only answer available, and why it should be read as a demonstration rather than a result. Ideogram chose the source images, the edit sequence, the number of rounds and the three competitors. No prompt set, no judge, no error bars and no per-round numbers accompany the claim, so there is nothing to reproduce. Early testers quoted at launch describe consistency across edits as strong, which is a useful signal and not a score.

The honest state of it is that Ideogram has identified a real and underserved failure mode, built a model around it, and published marketing for it. Whether 4.5 actually holds an image steadier than GPT Image 2.5 Sunburst or Nano Banana Pro over ten passes is an open question that will stay open until either an open-weight release or an independent repeat-edit benchmark exists.

Sources: The Decoder, Recraft, Ideogram, AMD.

Daniel Ochoa: Covers model launches, shutdowns and pricing changes as they happen. Reads deprecation notices for a living so you do not have to.