Ideogram 4.5: What's New, Pricing and Alternatives

It is the successor to Ideogram 4.0, but positioned as an edit model first. The model page headline reads "the most precise edit model", and the copy says it "reduces this drift, preserving details across multi-turn edits". Ideogram has long been known for legible text inside images; this release carries that strength into editing images that already exist.

Availability on launch day, per The Decoder: the Ideogram web platform and the API, with partners including Picsart, Runway, Pika and Leonardo AI. Recraft has announced it is coming to Recraft Studio without giving a date. Ideogram also has an iPhone app.

Three changes stand out for anyone doing multi-turn editing, all described on Ideogram's own page and in launch coverage:

1. 1

   Multi-turn stability

   The model is tuned so that edit five looks as clean as edit one. Ideogram shows the same sequence of edits run on several models and says competitors become unusable "within a few edits".

2. 2

   Editing at full resolution

   You can edit a crop of a very large image and stitch it back without downsizing the original. Ideogram's example is a 4,016 x 6,016 pixel source (24.2 MP). As SuperPowerDaily points out, that is a crop-and-reinsert workflow, not full-image generation at 24 MP.

3. 3

   Two API workflows

   According to AlphaSignal's summary of the API, "precise edit" returns a result at the input image's own size, while "generate and edit" mixes reference-driven generation with edit instructions. Requests can carry up to four reference images, an optional mask and a high-precision mode.

AlphaSignal also reports four quality modes, prompts of up to 10,000 characters and native 2K output. Ideogram's credit table confirms the quality modes by name (Very Low, Low, Medium, High), so the tier structure is not in doubt even where the other numbers come from coverage.

The model page groups its examples into ten edit types. They are a good map of what the model is meant for:

• Colour and lighting, including new product colourways
• Text modification inside a design
• Product photography
• Interior design
• Architecture
• Restoring old photos
• Sketch to image
• Style reference
• Reframing
• Depth to image

Most of these are jobs where the requested change is small but the image around it is full of detail you want to keep: a label on a bottle, the sofa in a staged room, a scratch on an old print. That is the kind of brief where drift costs the most time, because a fresh regeneration would change everything else too.

The first independent numbers came from Artificial Analysis. Reporting on them, AlphaSignal says the model placed 23rd on the image editing leaderboard and 34th for text-to-image generation, up from 40th for Ideogram 4.0 Quality. It is the first Ideogram model listed on the editing board.

Those ranks sound modest, but AlphaSignal adds that the model comes closest to the category leaders in three areas: text and symbol edits, scene and style edits, and composition and framing. That fits the specialist positioning. Leaderboard positions move every week, so check the live board before relying on a rank.

Hands-on coverage is still thin. MindStudio's first look tried two edits, fixing garbled text in an older Ideogram image and swapping in a reference character, and reported both held up, while stating plainly that a fuller verdict needs more testing.

The model page runs "14 edits with the same prompts" on its own Medium setting, GPT Image 2.5 Sunburst High, Nano Banana Pro and Nano Banana 2, and concludes that the GPT Image and Nano Banana outputs "become unusable within a few edits".

Treat this as a claim, not a verdict. It is Ideogram's choice of image, prompts and settings, and nobody outside the company had published a matched multi-turn test when we checked. It is also a test of a specific workflow: applying fourteen edits in a row. If you usually make one or two changes and then export, drift matters much less.

There are two ways to pay: credits in the Ideogram app, or per image through the API. Both are listed on the pricing page.

API prices are the same at 1K and 2K. Precise Edit uses the same four tiers, from $0.008 at Very Low to $0.22 at High, which is the 27-fold spread AlphaSignal highlighted. In other words, the cheapest edit costs less than a cent and the most careful one costs about a fifth of a dollar.

Subscription plans, billed yearly: Plus at $15 a month with 2,400 priority credits, Pro at $42 with 7,500, and Team at $20 per user with 3,600 per user (minimum two users). Monthly billing is higher ($20, $60 and $30). Paid plans also include slow credits with no monthly cap. Free accounts get weekly credits when they sign in with Google, Apple or Microsoft; Ideogram says "the amount may vary" and offers no free trial of paid plans.

Not yet. Ideogram released open weights for the previous model, Ideogram 4, on 3 June 2026. Its GitHub repository describes it as Ideogram's "first open-weight text-to-image model", with 9.3 billion parameters; the code is under Apache 2.0, but the model weights use an Ideogram 4 non-commercial license, so commercial use of the open model is not covered.

For 4.5, Ideogram has said an open-weight release is coming soon, according to The Decoder and AlphaSignal, but it has not given a date or a license. Until it ships, the only ways to use the new model are Ideogram's own app and API and the partners that integrate it.

ChatGPT can edit images in several rounds, and for many people that is enough. OpenAI's help centre says you can upload an image or pick a generated one, select an area with a selection tool and describe the change, or just describe the edit in chat.

OpenAI is candid about the limits: "Highlights are not always precise, and edits may extend beyond the area you selected." That is the same drift problem Ideogram is targeting. If your job is one or two quick changes, ChatGPT is convenient. If you need many rounds on a print-resolution file where untouched pixels must stay untouched, a dedicated edit model is the safer test.

The right alternative depends on the job. These are options we could verify today, grouped by what they are best at:

On LazyKiwi, two tools cover the most common single edits. Edit text in image changes the wording in a photo or graphic while keeping the design, and the AI photo editor applies a described change to a photo. If you would rather generate a new design than edit one, the image generator lists several models with the credit price shown up front. These are prompt-based editors for a single pass; we make no claim about how they behave over fourteen consecutive edits.

1. 1

   Pick one real image

   Use a photo or design you actually work with, ideally with text and fine texture.

2. 2

   Write five small edits

   Change one thing each time: a colour, a word, an object, the light, a crop.

3. 3

   Apply them in sequence

   Feed each result into the next edit, in each tool, with the same wording.

4. 4

   Compare the untouched areas

   Zoom into regions you never asked to change. Shifts there are drift.

5. 5

   Record cost and keeps

   Note the price per edit and how many results you would use.

Five rounds will not reproduce Ideogram's fourteen, but it is enough to see whether a tool holds your kind of image together, and it costs a few cents on the cheaper tiers.