Ideogram 4.5, Explained: The Edit That Copies the Pixels It Does Not Touch

What actually shipped

On 30 September 2026 Ideogram announced Ideogram 4.5 as its edit model. The model page describes the failure it is aimed at: each edit on a typical image model adds pixel shifts, color changes, and texture artifacts, and those errors pile up when you edit the same image again. 4.5 is the attempt to stop that pile-up.

It is one model with two jobs. Text-to-image generation, and image-to-image editing. The editing path is the one with a concrete mechanism, and it is the reason to look at this release at all. The model is live in the Ideogram app and on the API. Open weights were promised in the launch video and in Ideogram's own post. They are not out.

It did not make the 1 October briefing. That day went to chip leases, a distillation dispute, and a state law about automated firing. A model you can open in a browser the same week is the more useful thing to write down.

What Precise Edit actually guarantees

The API overview is plainer than the marketing. On POST /v2/image/precise-edit/ideogram-4-5, pixels the edit does not touch are copied exactly from the image you sent, and the result is returned at that image's own width and height. A mask limits the change to one region. Up to four reference images can steer a product, a material, or a style. A masked request can include at most three of those references. Quality is one of four values: very_low, low, medium, high.

That is a different claim from "the model is good at editing." Copying the untouched pixels means the tenth edit cannot slowly repaint the background, because the background was never regenerated. The model page demonstrates the size half of this on a source of 4,016 by 6,016 pixels, 24.2 megapixels: crop a region, edit it, and the crop edges are preserved so the patch stitches back into the original.

Generation is the other endpoint, POST /v2/image/generate/ideogram-4-5. Same model id, a prompt, no source image. Ideogram still talks about the thing the older models were known for, prompt fidelity and crisp type, but 4.5 is not being sold as a typography breakthrough. It is being sold as an edit you can repeat.

Where it actually sits

Every rank below is from Artificial Analysis, checked 2 October 2026. The editing board and the text-to-image board are scored separately. A rank on one says nothing about the other.

On AA-Image-Editing v2.0, Ideogram 4.5 High scores 1,064 Elo from 2,382 votes. That is 23rd, inside a band of 18th to 25th, so the point rank will move. The leader is GPT Image 2.5 Sunburst (max) at 1,182 Elo, then GPT Image 2.5 Flare (max) at 1,162, then Microsoft MAI-Image-2.6 at 1,137. 4.5 is a long way from that top. It is also well ahead of being unranked: the board has the model, the vote count is already in the thousands, and the confidence interval is published.

On AA-Image-T2I v2.0 the same High tier scores 1,014 Elo from 2,247 votes, 34th, inside a band of 25th to 41st. Ideogram 4.0 Quality, the June 2026 model this directory never listed on its own, sits at 1,011. Generation did not move. Editing is the release.

Artificial Analysis prices that High tier at $220 per 1,000 edited images and $100 per 1,000 generated images. Ideogram's launch video says one price per image at any size, starting at $0.008, across four quality modes. Those two statements can both be true: $0.008 is the floor, and the High edit is the ceiling the leaderboard is measuring. The public API price page still stops at Ideogram 4.0, last revised before this model existed, so treat the 4.5 rate card as the launch video plus the leaderboard's High-tier figure until Ideogram reprints the table.

When to use it

What not to believe yet

Secondary write-ups say early testers confirm the consistency, and they name launch partners. ComfyUI has said 4.5 is live in ComfyUI. Recraft has said it is on the way to Recraft Studio, which is not the same as live. The Decoder named Picsart, Runway, Pika, and Leonardo. None of that is a quality measurement. The measurement is the editing board, and it says 4.5 is a competent edit model that does not lead.

"Eliminates artifact buildup" is Ideogram's sentence. The mechanism that can be checked from the outside is narrower and better: untouched pixels are copied, and the file comes back at the size you sent. If a region you asked it to change looks wrong, that is still a generated region. The guarantee covers the rest of the picture.

The listing is at Ideogram 4.5. The rest of the family is on the Ideogram models page. For the editing leader it is being compared with, see GPT Image 2.5. The day's news, which did not include this release, is the 1 October briefing.