Images 2.6 vs Ideogram 4.5: The Specialist Costs More

Usually the specialist is the cheap option. It does one thing, it does it well, and it undercuts the generalist. Ideogram 4.5 inverts that. At its High tier it lists around $0.22 per image — $220 per thousand on Artificial Analysis's editing board — while OpenAI's Images 2.6, released on 4 November 2026 as GPT-Image-2.6 Flare and GPT-Image-2.6 Sunburst, lands near ten cents for a comparable 2K high-quality render on its $30-per-million output token rate. The model built specifically for precise editing charges roughly twice what the frontier lab charges for the same job.

That inversion is the whole comparison. Ideogram 4.5 has spent two generations arguing that a generalist cannot be trusted to change one thing without changing everything else. Images 2.6 is OpenAI's most direct attempt yet to prove that argument wrong — and for the first time, both models are selling almost exactly the same promise.

What Ideogram 4.5 is

Ideogram 4.5 shipped on 30 September 2026 as a closed, hosted model with two jobs: text-to-image and precise local editing. The editing pitch is specific and testable — edit a region and the surrounding pixels come back unchanged, and they keep coming back unchanged across repeated multi-turn passes, so drift does not accumulate. Ideogram demonstrated it at source resolution on a 4,016 × 6,016 image, 24.2 megapixels, which is a claim about pixel preservation rather than about generating a pretty new picture.

Native output is 2K. Pricing runs on four quality tiers rather than resolution — roughly $0.03 at Low, $0.06 at Medium and $0.22 at High on the vendor's own rate card, with resellers quoting $0.04 / $0.07 / $0.26. Availability is Ideogram's own app, API and MCP server, plus partners including Picsart, Runway, Pika, Leonardo and a handful of inference platforms. Open weights were announced as planned and have not been published.

The thesis each one is selling

Ideogram's position is that editing is a different discipline from generation, and that a model optimised for composition will always be tempted to recompose. Its answer is a conservative edit: touch the marked region, leave the rest bit-identical, repeat indefinitely.

OpenAI's position with Sunburst is subtler. It does not promise pixel-identical surroundings; it promises that the subject stays recognisable across edits — the face, the logo, the product detail. That is a semantic guarantee rather than a pixel guarantee, and it matters for a different failure: not "the background changed" but "the person no longer looks like the same person".

Images 2.6 adds two mechanisms aimed at exactly this. The complexity parameter lets the model plan more when a scene carries several constraints. Markup-guided editing — a transparent overlay drawn on the source, mutually exclusive with a mask — replaces a prose description of where to work with a picture of where to work. Ideogram's answer to the same problem is architectural rather than parametric: build the model so it does not want to touch the rest of the canvas.

Which guarantee you need depends on what you are editing. Product photography with a clean background wants Ideogram's pixel promise. A recurring character across a campaign wants OpenAI's identity promise. Most production pipelines need both and have been choosing one.

Text rendering: where the specialist earns its premium

This is Ideogram's strongest ground and the numbers are unusually concrete for an image-model comparison. Independent typography tests around launch found it setting 110 of 111 characters correctly on a poster's small print, 172 of 172 menu prices — with spacing artifacts at the Medium tier — and 300 of 300 characters on a dense ingredients label. A 36-model poster test found it placing all three lines correctly at every quality tier.

It also fails in instructive ways. A nutrition-panel test broke: values shifted rows, one word was misspelled and a row was invented. Ideogram's own documentation is candid that the model is for headlines and short blocks, is best in English, does not support font naming, and that body copy belongs in a design tool.

Images 2.6's text story is less documented. OpenAI says Flare improves layout quality for posters and product imagery, and the previous generation tested well on short English headlines — GPT Image 2.5 matched 111 of 111 characters on the same poster test where Ideogram scored 110. But there is no equivalent published evidence for dense, multi-block typography on 2.6, and no CJK story comparable to what the Chinese labs publish. If typography is the product, the specialist still has the receipts.

What each costs for the same job

• Ideogram 4.5 — around $0.03 per image at Low, $0.06 at Medium, $0.22 at High; Artificial Analysis measures $100 per thousand for text-to-image and $220 per thousand for editing at the High configuration.

• Images 2.6 — $5 per million text input tokens ($1.25 cached), $8 per million image input tokens ($2 cached), $30 per million image output tokens. A high-quality 2K render models to roughly ten cents; a low-quality 1K render to about half a cent.

• A typical edit — 900 text tokens in, 1,200 image tokens in, 4,000 image tokens out — is $0.1341 on Images 2.6, whatever the model does with it.

The per-image and per-token structures behave differently under iteration, and that is where this gets interesting. Ideogram bills the same $0.22 whether the edit changes a button colour or rebuilds the scene. OpenAI bills by the pixels it paints — a small edit with a small output costs less than a full re-render. For a workflow that makes many small corrections to one hero image, the token model is structurally cheaper, and the gap widens with every turn.

For a workflow that generates at scale and edits rarely, Ideogram's Low tier at three cents is competitive with anything, and its predictability is worth something a token estimate is not.

What the boards say

Artificial Analysis places Ideogram 4.5 (High) at 1,016 Elo on text-to-image, in a wide rank band around the high twenties, and 1,069 on editing, in the mid-twenties. Those are mid-table numbers, and they are lower than Ideogram's own launch framing implied — its 15th-place editing claim put it in the same band as MAI-Image-2.6 and Gemini 3 Pro Image, which the current board does not support.

For context, GPT Image 2.5 Sunburst sits at 1,224 and 1,199 on those same boards. So the previous OpenAI generation leads Ideogram 4.5 by more than 200 Elo on text-to-image and 130 on editing. Images 2.6 itself is unscored: we checked both boards on 5 November, the day after launch, and neither Flare nor Sunburst has a row. OpenAI's only ranking claim is that Sunburst took first place on Design Arena ahead of its predecessor — vendor-cited, and without a published Elo we could verify.

Read against each other, the boards say something uncomfortable for both: the specialist is mid-table on general preference, and the generalist's new model has no measurement at all. The place Ideogram actually wins is not the arena. It is the narrow, repeatable, pixel-preserving edit — which is precisely the case an Elo score is worst at capturing.

When to pick which

Pick Ideogram 4.5 if your work is retouching: one asset, many approved corrections, a background that must not move, a client who will compare the before and after at 100%. The $0.22 is a premium for a guarantee, and if you have ever had a brand asset come back with a subtly different logo, you know what the guarantee is worth. Pick it too if dense English typography is the deliverable and you have tested it on your own copy — the published character-accuracy numbers are the best evidence in this category.

Pick Images 2.6 if you need motion, vector output or resolution headroom — Ideogram 4.5 offers none of the three — or if your edits are many and small, where the token model is cheaper per correction. Pick it as well if you are already in OpenAI's stack: same endpoints, same token rates as 2.5, and the migration is a model string.

Neither is on OrcaRouter at the time of writing. What our image line routes today is GPT-Image-2.5 Flare and Sunburst, Google's Gemini and Imagen models and Grok Imagine Image, all behind one key at provider list price with 0% markup and automatic failover. If your pipeline ends up using both a specialist and a generalist, that is the shape of integration worth planning for: one credential, one bill, and the model choice a routing decision rather than a procurement project.

The thing that would settle this comparison is a pixel-diff benchmark: same source image, same instruction, measure how much of the untouched canvas moved. Nobody publishes that. Until someone does, the honest test is your own worst-performing asset — run it through both, and look at what changed that you did not ask to change.