Generated title card headed 'Ideogram 4.6', subtitle 'Native 4K edits, and the one number the vendor can be checked on'. Two cards. Left, 'What shipped on 19 November 2026': Output ceiling 4096 by 4096 pixels; Edit endpoint Precise Edit, optional mask; Reference images up to five through Generate; Quality modes Very low, Low, Medium, High; Price per image $0.008 to $0.60. Right, 'What the vendor measured': Rapidata multi-step edit 1st, 1,892; Reported interval 32 points; Gap over Ideogram 4.5 26 points; Masked edits, usable 90% against 82%; Independent boards no 4.6 row yet. Footer reads that the Rapidata figures are the vendor's snapshot dated 17 November 2026, that usable-result rates are vendor-reported at High quality, and that prices are per Ideogram's own API rate card. The OrcaRouter logo sits bottom-right.
Guides & Insights

Ideogram 4.6: Native 4K Edits and the Drift Numbers

Author

jinhao.song

Date Published

Latest models · 20View all models →
Benchmarks: Artificial Analysis · updated daily
Back to all posts

Ideogram 4.6 went live on 19 November 2026, and the first thing worth saying about it is that the vendor put a number on the claim. Most image-model releases this year shipped with a gallery and an adjective. This one shipped a leaderboard position: first on Rapidata's Multi-step Image Edit board at 1,892, with Ideogram 4.5 second at 1,866. That is a gap of 26 points on a Bradley–Terry fit rescaled to Elo, and it is the only part of the launch a reader can check without trusting the company that made the model. Everything else on the page — native 4K, less pixel drift, sharper colour, more texture — is asserted. Two of those four are now partly evidenced. Two are not evidenced at all.

What actually shipped

Generated comparison scoreboard headed 'Ideogram 4.6 - the launch, line by line', with columns for Ideogram 4.6 and Ideogram 4.5, for comparison. Rows: Released 19 November 2026 against 30 September 2026; Output up to 4096 by 4096 against native 2K; Quality modes Very low, Low, Medium, High on both; Generate at High $0.10 at 2K and $0.20 at 4K against $0.10 fixed; Edit at High $0.22 at 2K and $0.60 at 4K against $0.22 fixed; Rapidata multi-step edit 1st at 1,892 with a 32-point interval against 2nd at 1,866 with a 30-point interval. Footer reads that prices are per Ideogram's own API rate card and that the Rapidata scores come from the vendor's snapshot dated 17 November 2026, with 4.6 configured as Precise Edit at High Quality.

The vendor's own specification block is unusually specific, which makes it easy to separate what is new from what is inherited. This is the release, in its own terms.

• Resolution — up to 4K, defined as output up to 4,096 by 4,096 pixels; Very low quality is capped at 2K; Precise Edit preserves the source image's dimensions rather than re-cropping

• Quality modes — Very low, Low, Medium and High, available on both endpoints

• Endpoints — POST /v2/image/generate/ideogram-4-6/ for new images, and POST /v2/image/precise-edit/ideogram-4-6/ for a focused change to an existing one

• Reference images — up to five through Generate; when references are attached, the first image is the one being edited and the rest are treated as references

• Local edits — an optional mask through Precise Edit, so a change can be confined to a painted region

• Price — from $0.008 per image at Very low / 2K, up to $0.60 per image for a High-quality 4K edit

None of that is a breakthrough on its own. What is notable is the combination: a fixed set of four quality modes carried across both endpoints, a hard resolution ceiling of 4,096, and masking available on the edit endpoint rather than as a separate product. That is a workflow specification, not a demo specification, and it is the shape you want if the model is going into a pipeline instead of a queue of one-off generations.

One thing that did not ship is worth recording, because it was expected: there is no launch post on the Ideogram blog. The blog index carries articles up to earlier models and a navigation link to 4.6, but no 4.6 announcement. The model page and the API pricing page are the vendor's primary sources for this release. Anyone citing a launch essay is citing something that does not exist.

The four claims, and what backs each

Screenshot of the benchmark.ai page 'Multi-Step Image Editing Benchmark: Identity Drift Judged by Humans', published by Rapidata. A leaderboard table with columns Model, Organization, Score, 95% CI Interval, Responses, Matches and Winrate: Ideogram 4.6 first at 1,892 with a 86.3% winrate, Ideogram 4.5 second at 1,866, Seedream5.5 third at 1,805, FLUX 3 Image fourth at 1,697, Muse Image fifth at 1,406, MAI-Image-2.6 sixth at 1,299, MAI-Image-2.1-Pro (Thinking) seventh at 1,291, Seedream 5.0 Pro eighth at 1,174, MAI-Image-2.5-Pro ninth at 1,145, Qwen-image-3.0-Pro (Thinking) tenth at 1,014, GPT-Image 2.6 eleventh at 1,001 and Nano Banana Pro twelfth at 987, with a 'Show all 23 models' control below. The caption under the table states that scores come from a Bradley-Terry fit rescaled to Elo with 95% Wald intervals. The page's abstract states 199 photographs and 413,667 responses collected across 76,714 matches, and a model filter in the left sidebar reads 'Ideogram 4.6 (Precise Edit, ...)'.

The change log against 4.5 is four sentences long: less pixel drift outside the edited region, more accurate colour, native 4K generation, and more realism and texture detail than any prior Ideogram model. Those four claims have very different evidentiary status, and the page does not always make the difference clear.

Drift is the one with third-party backing. Rapidata's board is built specifically to measure it: 199 photographs of real people, each put through a five-prompt chain of everyday edits to clothing, objects or text in the scene, with none of the prompts touching the face. Each model applies the chain to its own previous output, every step is automatically checked to confirm the requested edit was made, and human annotators then compare the face after the fifth edit against the original. The aggregate is 413,667 responses across 76,714 matches, fitted with a Bradley–Terry model so that a higher rating means less identity drift. Ideogram 4.6 sits at the top of that board.

Colour accuracy, 4K and texture are vendor-reported. The model page says 4.6 "leaves far less pixel drift outside your edits and holds colour more accurately," and that it "generates in native 4K, with more realism and texture detail than any Ideogram model before it." The 4K half of that is checkable in the least interesting way — the API will return a 4,096-pixel image or it will not, and the pricing table implies it will. "More realism" is not checkable at all from the page, because the page's evidence is a gallery of sequences the vendor chose.

The five usable-result rates

The most concrete numbers in the launch are hidden in a comparison chart rather than in the prose. Ideogram reports how often each model produced a usable result on the same requests at High quality, across five editing tasks, and prints the previous model's rate beside the new one.

• Character consistency — 85% on 4.6 against 69% on 4.5

• Exact color — 83% against 46%

• Product detail — 86% against 61%

• Product shots — 88% against 66%

• Masked edits — 90% against 82%

Read that as a set and the pattern is informative. The smallest gain, eight points, is on masked edits — the task where a mask already constrains where the model is allowed to work, so there is less room for it to go wrong. The largest, 37 points, is exact colour, which is the task where a model either lands on the hex you asked for or does not. Those two facts are consistent with a model that got materially better at holding a specified value rather than one that got better at everything evenly.

They are also, unambiguously, vendor figures. There is no published methodology behind "usable result," no annotator count, no statement of who judged usability or whether they saw both outputs side by side. Treat the direction as credible and the magnitude as marketing until someone else reproduces it.

Where the independent record is thin

Here is the part of the release that the launch page does not say out loud. Artificial Analysis, the other board that people use to compare image models, has no Ideogram 4.6 entry. Its text-to-image leaderboard scores Ideogram 4.0, 4.0 Fast, 4.0 Instant, 4.5 at High quality and several P-Image variants, and its editing leaderboard scores 4.5 at 1,065 Elo from 4,003 samples. There is no row for 4.6 on either as of this writing. So the single independent data point for this model is the Rapidata board, and a single board is a single view.

That board also needs one caveat attached. The two configurations being compared are not identical: Ideogram 4.6 is listed as Precise Edit at High Quality, and 4.5 is listed as High Quality with High Precision. Both are the top settings for their respective models, but "High Precision" is not a mode that appears in the 4.6 API, and the vendor page itself notes that the leaderboard scores are separate from the task-specific percentages in the comparison chart. A 26-point gap between two top settings of two models in the same family is a real result. It is not a controlled experiment on the effect of the upgrade.

The intervals make the same point more sharply. 4.6 is reported at 1,892 with a 32-point interval and 4.5 at 1,866 with a 30-point interval. Those intervals overlap — 4.6's floor of 1,860 is below 4.5's ceiling of 1,896. The point estimate favours 4.6 and the board correctly ranks it first. Whether the underlying difference is distinguishable from noise at 95% confidence is a separate question, and on these numbers the honest answer is that it is close.

What the 4K tier costs

Screenshot of Ideogram's own Ideogram 4.6 model page. The hero reads 'Ideogram 4.6 is the new frontier of precise editing.' Below it an 'Ideogram 4.6 at a glance' block lists Release date November 19, 2026; Resolution Up to 2K or 4K, with Very low quality supporting up to 2K; Quality modes Very low, Low, Medium, High; Reference images up to five through Generate; Local edits optional mask through Precise Edit; and API endpoints Generate and Precise Edit. To the right sits a 'Meet Ideogram 4.5' promo card. Lower down a 'Multi-step Image Edit - benchmark.ai' panel shows a horizontal bar chart headed '#1 in multi-step image editing' with Ideogram 4.6 first, Ideogram 4.5 at 1866, Seedream5.5 at 1805, Flux 3 Image at 1697, Muse Image, MAI-Image-2.6 at 1299 and MAI-Image-2.1-Pro (Thinking) at 1291, sourced to Rapidata. No per-image API rate table is visible in this capture.

The price sheet is where this release changes the arithmetic, and it changes it more than the benchmark does. Ideogram 4.6 bills by quality and by resolution, which is a departure from how 4.5 is priced. The full grid:

• Very low — $0.008 at 2K for both Generate and Edit; no 4K option

• Low — $0.03 at 2K, $0.06 at 4K, on both endpoints

• Medium — $0.06 at 2K for Generate and Edit; $0.12 for a 4K generation, $0.20 for a 4K edit

• High — $0.10 for a 2K generation, $0.20 for a 4K generation, $0.22 for a 2K edit, $0.60 for a 4K edit

The last cell is the one to sit with. A High-quality 4K edit costs six times the 2K generation at the same quality, and 2.7 times the 2K edit. Anything at 2K or below is billed at the 2K price, so the cliff is exactly at the point where you cross into 4,096. If your deliverable is print or a 4K display, that is the number your per-asset cost becomes. If it is not, you are paying a premium for pixels nothing in your pipeline will display.

Worth stating plainly for anyone comparing this against a routing layer: Ideogram is not on OrcaRouter's model list today, so there is no per-call route to these endpoints from us and no reason to expect our price to differ from Ideogram's own. Where a single key does help is on the other side of a mixed workload — the Goog​le image models, the G​PT-Image line and Grok Imagine Image are all on the same key with provider list price passed through at 0% markup, so a workflow that uses Ideogram for edits and something else for generation does not need a second contract for the second half.

Who should move now

This is a precision-edit release wearing a resolution headline, and the two audiences are different. If your work is multi-turn editing — the same source image, one change at a time, twenty turns deep — 4.6 is the model the evidence actually supports, and the masked-edit and colour numbers suggest the improvements concentrate exactly where a long edit chain compounds error. Move, and test on your own worst sequence rather than the vendor's gallery.

If your work is single-shot generation at 2K, the case is much weaker. You would be paying the same $0.10 at High that 4.5 charges, for a drift improvement that mostly matters on the second edit onward, and the 4K tier that justifies the release's headline would go unused. Ideogram's own model guide makes a version of this argument: it recommends 4.6 for precise edits and output up to 4K, and suggests comparing against 4.5 before migrating an existing workflow. That is an unusually candid thing for a launch page to say, and it is the right advice.

The open question is the one Artificial Analysis will answer when it scores the model. Until then, the defensible position is narrow: 4.6 is first on the one board that measures what it claims to improve, the intervals around that first place are wide, and the price of the feature that made the headline is $0.60 an edit.