Ideogram 4.6 Brings Native 4K to Precise Editing, but Its 26-Point Lead Is Inside the Error Bars
TL;DR: the short version
On November 19, 2026, Ideogram shipped Ideogram 4.6. The story here is not "a prettier image model." It is the engineering spec for editing work:
- Native 4K. Output ceiling rises from 2K to 4096×4096 pixels.
- Two separate endpoints.
generate(new images, or generation with reference images) andprecise-edit(one focused change to an existing image), each with four quality tiers. - Masks live inside the edit endpoint. You no longer need a separate inpainting product to confine a change to a painted region.
Three things worth knowing up front:
- It wins on multi-step editing, not on generation quality. Ideogram publishes a Rapidata Multi-step Image Edit score of 1,892 for 4.6 against 1,866 for 4.5, ranked first and second. But 4.6 carries a ±32 interval and 4.5 carries ±30 — the intervals overlap, so that 26-point gap is not statistically clean.
- 2K pricing did not move. 4K is the new bill. Every tier costs the same as 4.5 at 2K. Once you ask for 4K, a High-quality edit jumps from $0.22 to $0.60 — 2.7× the 2K edit and six times a 2K generation at the same quality.
- There is no launch blog post. The model page and the official X account are the only first-party sources; the blog index has no 4.6 article. That is unusual, and it means every secondary report is reading the same spec sheet.
This article is compiled from multiple media reports, together with Ideogram's official model page and its official X post.
What this actually is
Ideogram built its reputation on text that renders correctly inside an image; the 4.0 line targets posters, packaging and stickers — work with real typography requirements. 4.5, released September 30, was its "precise editing" release, priced flat: one price per image regardless of how many reference images you attached or what resolution you asked for.
4.6 pushes that into a shape you can wire into a production pipeline. Four quality tiers (Very low, Low, Medium, High) apply across both endpoints, resolution is capped at a hard 4096, and masking sits in the edit endpoint rather than in a separate product page.
Ideogram's own description of the change from 4.5 is four sentences long: less pixel drift outside the edited region, more accurate color, native 4K generation, and more realism and texture detail than any prior Ideogram model.
Those four claims have very different levels of evidence behind them. Below, they are separated.
The three genuinely new things
One: a hard 4096-pixel ceiling. "4K" means output up to 4096×4096. The Very low tier is still capped at 2K. Precise Edit preserves the source image's dimensions instead of re-cropping to hit a resolution target — which matters if you produce product shots and e-commerce detail pages, because an edit that silently changes the aspect ratio wrecks the surrounding layout.
Two: the endpoints are split. Generation goes to POST /v2/image/generate/ideogram-4-6; editing goes to POST /v2/image/precise-edit/ideogram-4-6. Generate accepts up to five reference images; when references are attached, the first image is the one being edited and the rest are treated as style or content references. It is not glamorous, but it removes a recurring ambiguity: previously "the reference images" and "the image to edit" were often crammed into one parameter and you inferred the order by convention.
Three: the mask is an optional parameter on the edit endpoint, not a separate product. You paint a region and the change is confined to it. When this capability is sold separately, it usually means another SDK to integrate and another billing line to maintain.
How to read that 26-point lead
This is the part that deserves the most care.
The leaderboard Ideogram cites is Rapidata's Multi-step Image Edit. The method: 199 photographs of real people, each put through a five-prompt chain of everyday edits (clothing, objects, 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 actually happened, and human annotators then compare the face after the fifth edit against the original to see whether identity drifted.
The aggregate is 413,667 human responses across 76,714 pairwise matches, fitted with a Bradley–Terry model and rescaled to something Elo-like, where a higher number means less identity drift.
On the snapshot shown on the official page (November 17, 2026):
- 1st: Ideogram 4.6, 1,892, ±32
- 2nd: Ideogram 4.5, 1,866, ±30
- 3rd: Seedream 5.5, 1,805, ±29
- 4th: FLUX 3 Image, 1,697, ±27
It looks like a 26-point win. Spread the intervals out, though: 4.6's lower bound lands near 1,860 and 4.5's upper bound near 1,896, so they overlap. The board is right to print 4.6 first, but at 95% confidence the two models are not cleanly separated on this measure.
There is a second detail not to skip: those two rows are not the same configuration. 4.6 is evaluated as Precise Edit at High Quality; 4.5 is evaluated as High Quality with High Precision — a 4.5-era label that does not correspond to any mode in the 4.6 API. Comparing each generation at its own top setting is fair as a product ranking. It is not a measurement of "who is stronger with the same switch flipped."
The five numbers Ideogram reports for itself
The official page includes a comparison chart giving the share of requests that produced a usable result at High quality, 4.6 against 4.5:
- Character consistency: 85% vs 69%
- Exact color: 83% vs 46%
- Product detail: 86% vs 61%
- Product shots: 88% vs 66%
- Masked edits: 90% vs 82%
Read as a set, the information is in the shape. The smallest gain, eight points, is on masked edits — where the mask already constrains where the model may act, leaving less room to go wrong. The largest, 37 points, is exact color — the task where you either land on the hex you asked for or you do not.
That shape is consistent with a model that got materially better at holding a specified value, rather than one that improved evenly across everything.
It must be said plainly: these five figures are vendor-reported. There is no published methodology behind "usable result," no annotator count, and no statement of whether annotators saw both models' outputs side by side. Treat the direction as credible and the magnitude as marketing until someone reproduces it.
One gap on the independent side is also worth recording: Artificial Analysis currently has no 4.6 entry on its image leaderboards — only the 4.0 family and 4.5. So apart from the board Ideogram itself cites, there is no second public ranking to check against today.
Cost: 2K unchanged, 4K steep
This section is for anyone already paying Ideogram. At 2K, 4.6 is priced identically to 4.5 tier by tier. Every difference sits in the 4K column:
- Very low: $0.008 at 2K, no 4K option
- Low: $0.03 at 2K, $0.06 at 4K
- Medium: $0.06 generate/edit at 2K; $0.12 for 4K generation and $0.20 for a 4K edit
- High: $0.10 generate / $0.22 edit at 2K; $0.20 for 4K generation and $0.60 for a 4K edit
Put another way: the model did not get more expensive; the pixels did. A High-quality 4K edit at $0.60 is 2.7× the 2K edit and six times a 2K generation at the same setting. The Medium row is more awkward still — a 4K edit at $0.20 costs almost as much as a 2K High-quality edit at $0.22 while sitting a tier lower. If your deliverable has to be 4K, you are choosing between a cheaper low-tier large image and a costlier same-tier large image, with no comfortable middle.
For a rough budget: 100 4K product images with two edits each lands around $120 at High quality; the same workflow at 2K lands near $44. What the difference buys is not having to upscale for print or large displays. If your endpoint is a web page or a phone screen, that money is probably unnecessary.
Who it's for
- E-commerce and product-image pipelines. One hero shot needs a dozen color, material and scene variants while the subject stays pixel-identical. 4.6's mask plus Precise Edit plus source-dimension preservation maps directly onto that.
- Brand assets. Logo, primary color and whitespace are hard constraints, and changing copy must not reflow the whole layout. Ideogram's exact-color improvement from 46% to 83% describes exactly this work.
- Iterative design workflows. Change one thing, look, change the next — rather than restarting from scratch each round. That is what the leaderboard it wins actually measures.
- Print or large-format delivery. 4096 is a real ceiling increase, not an interpolation trick.
Who it's not for: limitations
- People who just want a beautiful concept image from scratch. Composition and aesthetics dominate that job; 4.6's selling points (less drift, color fidelity) barely register, while its 4K pricing is very real.
- Budget-sensitive teams delivering to screens. The 2K tiers did not get cheaper, and the same money may buy a higher blind-preference score elsewhere.
- Anyone who needs a third-party leaderboard before signing off. Artificial Analysis has no 4.6 entry; the only number available comes from a board the vendor cites, with overlapping intervals.
- Technical selection processes that require statistical significance. Twenty-six points against ±32 and ±30 does not support "significantly better than 4.5."
Pitfalls
- Do not read 4.6 as a clean sweep over 4.5. The leaderboard intervals overlap and the configurations differ. What is unambiguously different: whether 4K exists, whether masking sits in the edit endpoint, and how the price sheet is computed.
- Do not go looking for a blog post. There is none this time. Anything you read claiming "the official blog says…" is paraphrasing the model page.
- Do not miss how 4K is billed. Anything at 2K or below bills at the 2K price; only 4K moves tiers. And 4K edits and 4K generations are two different prices ($0.60 vs $0.20 at High quality) — do not copy one into the other when you wire up the API.
- Do not assume 4K is faster. The page gives no latency figures for 4K; the spec block covers resolution and quality tiers only.
- Do not use the usable-result rates as an acceptance test. They are vendor figures. Run ten of your own images.
How to use it: next steps
- Start with your own task, not the leaderboard. Pick ten real images you need to edit and run the same prompts through 4.5 and 4.6 Precise Edit at High quality, 2K. Judge one thing only: did anything outside the requested change move. The whole test costs under a dollar.
- Decide whether you actually need 4K. Web and phone delivery stays at 2K. For print or large displays, open the 4K column — and prefer generation over editing for the big output ($0.20 vs $0.60).
- Wire the endpoints correctly. Generation to
generate, edits toprecise-edit; up to five reference images, with the first being the subject under edit. Do not stuff reference images into the edit endpoint as style control. - Try the mask before you ask for a full re-render. If a change can be painted, do not let the model decide the boundary itself.
- Keep the original and every prompt. The classic failure in multi-step editing is drifting by round three with no record of which step caused it.
Sources
- Ideogram 4.6 official model page: https://ideogram.ai/models/4.6
- Ideogram official X launch post: https://x.com/ideogram_ai/status/2123461101224624402
- OrcaRouter: what shipped in Ideogram 4.6, and what the drift score does and does not prove: https://www.orcarouter.ai/blog/ideogram-4-6-native-4k-precise-edit
- OrcaRouter: Ideogram 4.6 vs 4.5 pricing comparison: https://www.orcarouter.ai/blog/ideogram-4-6-vs-ideogram-4-5
- AI Compass daily edition, November 19, 2026: https://ashutoshai.com/ai-compass/2026/11/19/
- Promptowy hands-on test (Polish): https://promptowy.com/ideogram-4-6-premiera/