• Ideogram claims 4.5 keeps unedited regions intact across multi-turn edits, but publishes no drift metric or benchmark for that claim.
• The reported 0.8-cent price is the floor of a four-tier range that tops out at 22 cents per image; quality at the cheapest tier is not documented.
• The open-weight release is announced with no date, so reproducible self-hosted testing is not yet possible.
• The failure to test for is cumulative drift outside the mask over many sequential edits, measured per accepted edit rather than per call.
• The claim that rival models become unusable after a few edits is Ideogram's own comparison, not an independent result.

Ideogram has launched 4.5, an image model it says can change one part of a picture and leave the rest alone. The Decoder reports native 2K output, four quality tiers from 0.8 to 22 cents per image, integrations with Picsart, Runway, Pika and Leonardo AI, and an open-weight release promised without a date. The headline claim, that edits do not degrade untouched pixels, is the one thing no source quantifies, and it is the thing that decides whether this ships in a production pipeline.

What Ideogram actually claims

Ideogram's model page calls 4.5 "the most precise edit model" and says it reduces pixel shifts, color changes and texture artifacts across multi-turn edits. It states that GPT Image and Nano Banana outputs become unusable within a few edits, while 4.5 stays clean. The page shows editing on an image of 4,016 × 6,016 pixels, about 24.2 megapixels. It publishes no benchmark, no drift metric and no failure rate, so the comparison is a vendor assertion rather than a measurement.

Price and availability are reported, not yet verifiable

The Decoder reports four quality tiers from 0.8 to 22 cents per image at native 2K, access through the Ideogram platform and API, and integrations with Picsart, Runway, Pika and Leonardo AI. The model page we read carries no pricing and does not mention partners. Cost per call is therefore a range of roughly 27x, and the 0.8-cent figure likely belongs to the lowest tier, whose edit quality is undocumented.

The open-weight promise has no date

The Decoder quotes Ideogram saying an open-weight release is coming soon, with no timeline. Ideogram's model page links a GitHub repository named ideogram-oss/ideogram-4 and a Hugging Face collection, but we did not confirm that 4.5 weights are in either. Until they are, teams cannot run reproducible edit-drift tests on their own hardware, and price and terms remain whatever the hosted API sets.

How to test it before it ships

The failure that matters is cumulative drift. Take 50 production images, apply ten sequential masked edits to each, and measure pixel difference outside the mask after every step, plus color shift and texture change. Run the same script against your current model and log cost per accepted edit, not per call. If error in the unmasked region stays flat across steps, the claim holds for your workload; if it compounds, the demo hid the failure.

Verdict

Do not switch on the strength of this launch. The pitch targets a real production pain, drift across repeated edits, but the only evidence is Ideogram's own page and a press summary of it. Pilot the hosted API on your own images, price the tiers you would actually use, and wait for the open weights before committing to a self-hosted design. Revisit once independent drift measurements exist.