Images 2.6 vs Krea 3: A Generalist That Now Animates Against an Editor That Only Edits

Search for "Images 2.6 vs Krea 3" and you will get pages about Microsoft's MAI-Image-2.6, because two shipping models now share a number and nobody has bothered to disambiguate them. So let us start there. Images 2.6 is OpenAI's, rolled out across ChatGPT, Codex and the API on 4 November 2026 — one day before this was written — in two API variants, GPT-Image-2.6 Sunburst for quality and GPT-Image-2.6 Flare for speed. Krea 3 is Krea's editing model, live since 15 October 2026, sold as an editor first rather than a generator with an edit box bolted on. They are not really the same product category, and the comparison is more useful once you accept that: one is a platform feature with two API faces, the other is a specialist rate card.

That difference decides almost everything below. Images 2.6 inherits an enormous installed base, adds animated GIF generation and reusable reference images inside ChatGPT, and ships behind token-based pricing that has not moved since the 2.5 generation. Krea 3 does one thing — region-scoped image editing at native 2K — and sells it at a flat per-edit price with surcharges for every reference image you attach. If you are buying a creative assistant, the first is the safer default. If you are buying a production edit step that runs a million times, the second is the one with a rate card you can forecast.

The honest state of the leaderboard

Here is where the two models genuinely differ in verifiability, and it cuts the opposite way from what you might expect.

Krea 3 has independent numbers now. On Artificial Analysis's AA-Image-Editing v2.1 board, checked on 5 November 2026, Krea 3 (Ultra) sits at 1,151 Elo with 5,573 samples — level with Nano Banana 2.1 and MAI-Image-2.6, which post the same 1,151 across ranks four through seven. The cheaper Krea 3 (Fast) mode scores 1,125. Those are third-party, human-preference, and they replaced the vendor-only picture that existed at launch, when the board had no Krea row at all.

Images 2.6 has no independent number of any kind. Not on AA-Image-Editing v2.1, not on AA-Image-T2I v2.1, not on Arena. A full read of both Artificial Analysis image boards on 5 November 2026 returns rows for GPT Image 2.5 Sunburst (1,199 editing, 1,224 text-to-image), GPT Image 2.5 Flare, and GPT Image 2 (high) at 1,137 and 1,188 — and nothing for 2.6. The model is a day old and the boards are Elo systems fed by thousands of pairwise votes, so this is a lag, not a verdict. But it means any sentence of the form "Images 2.6 beats Krea 3" is currently unsupported by any measurement anyone has published.

What you can compare is the parent. If 2.6 merely holds its predecessor's position, the relevant reference point is GPT Image 2.5 Sunburst at 1,199 editing — 48 Elo above Krea 3 Ultra, a gap well outside both models' confidence intervals. OpenAI's own claim is that 2.6 would not go backwards; it says Sunburst is tuned for "greater image consistency across edits" at medium and high quality. That is a claim about a property the Elo boards reward but do not isolate.

What one edit actually costs

This is the part where the two rate cards are least comparable, so it is worth doing the arithmetic on one realistic job: take a 2K product shot, attach two 1K reference images, and ask for a single regional change.

• Krea 3 — $0.06 for the 2K output, plus $0.02 for each 1K reference, so $0.10. Add the "high" thinking mode and it is $0.12. Four 2K references instead of two 1K pushes it to $0.34, which is the number worth remembering when a brief says "match these four angles."

• GPT-Image-2.6 Sunburst — billed in tokens, not images. Text input is $5 per million tokens (cached $1.25), image input $8 per million (cached $2), image output $30 per million. OpenAI publishes no flat per-image price, so the honest figure is an estimate from the output-token budget: roughly $0.053 at high quality for a 1024-class output, $0.094 at xhigh, $0.21 at max. Those token budgets are carried over from the 2.5 generation, which is the best available approximation and not a published 2.6 figure.

• Batch — OpenAI's published batch token rates are roughly half ($2.50 / $4.00 / $15.00 per million), but the model pages for the 2.6 variants list /v1/batch as not supported, so the discount is not currently reachable for this model. Krea publishes no equivalent discount.

The shape of the risk is what matters. At one edit, the two are within a few cents and the comparison is noise. At a hundred thousand edits a month, Krea's price is fixed and known, while the 2.6 invoice floats with the token budget your quality setting implies — and the setting is a per-request parameter, which means a single careless default can move a monthly bill by 4× without anyone changing a line of application code.

That is also where a router earns its place rather than being a talking point. OrcaRouter passes provider list price through at 0% markup, so the per-token OpenAI rate you see is the one on this page and the one on your invoice, with no spread added in between. The useful move with a model this new is to put it behind the same key as the incumbents you already trust — GPT-Image-2.5 Sunburst and Flare, GPT Image 2, Imagen 4 and Gemini's flash-image line are all on the catalogue — and let automatic failover catch the day the new endpoint rate-limits you. You are measuring a day-old model against a known one, and you would rather not do that with a production path that has a single point of failure.

Where Krea 3 is actually the better tool

Three things, and they are specific.

Krea 3 was trained to edit, not adapted to it. The vendor's claim is that it renders only the changed region rather than redrawing the whole frame, which is why region-scoped editing is the default interaction rather than a mask you draw and hope the model respects. For product work that is the difference between a label surviving an edit and a label being quietly re-lettered — and re-lettering is the failure mode that makes teams stop trusting an editor entirely.

It also ships 251 launch presets, which sounds like consumer garnish and is not: day-to-night, blockout, sketch, re-skin, new sky, green screen. A preset is a prompt someone already tuned, and for a studio running the same five transformations across a catalogue, that is a week of prompt engineering you do not do.

And its latency story is a flat per-edit number you can put in a queue design. Krea claims most edits finish under 8 seconds end to end at native 2K. Nobody has independently measured that — it is vendor-reported — but the structure of the claim (a bound at a fixed resolution) is more useful for capacity planning than a percentage-speedup claim against an unnamed baseline.

What Krea 3 does not do is generate. There is no text-to-image row for it on the AA T2I board because there is no text-to-image mode. It needs a source image. If your pipeline starts from a prompt, you need a generator in front of it, and that changes the cost comparison entirely — you are now paying two models.

Where Images 2.6 pulls ahead

Images 2.6 is the only one of the two that can start from nothing, and it is the only one with animation.

The GIF feature is the headline of the 4 November release: ChatGPT can now build a looping animated GIF from a prompt or an uploaded photo, in one step, with playback-speed control and per-frame editing afterwards. Whether the API returns animated files is a separate question and currently unanswered — OpenAI's model pages list image output and point to the image guide for "animated GIF output," but no public test has confirmed a GIF coming back from a /v1/images/generations call. Treat ChatGPT GIF support as shipped and API GIF support as unverified.

The rest of the release is more consequential for developers than the GIF. Images 2.6 adds SVG export, markup-guided edits (send an annotated image and the model edits where you drew), reusable reference images, and a complexity parameter that controls how much planning budget the model spends on a request — OpenAI's guidance is to start at low and raise it when a scene has several interacting constraints. SVG export in particular is a category change: an editor that emits vector paths is usable for logo and icon work that a raster editor cannot touch at any quality level, and Krea 3 has no equivalent.

The volume claim is worth repeating with its label attached: OpenAI says users now generate over 4 billion images a week across ChatGPT and the API, up from 3 billion in September 2026. That is a company figure, with no published split by feature and no guarantee the two counts are defined the same way.

Which one to put in the pipeline

If your work is editing existing images at volume — catalogue retouching, regional changes to product photography, restyling a set that already exists — Krea 3 is the more predictable purchase. Fixed per-edit pricing, a native 2K output, a measured 1,151 Elo on an independent board, and a design that treats preservation of the untouched region as the point rather than a nice-to-have.

If your work is generation with editing attached — concepting, marketing variants, anything that starts from a sentence — Images 2.6 is the only one of the pair that applies, and the SVG and markup-edit features make it more than a chatbot convenience. The absence of an independent benchmark is a reason to pilot it, not a reason to dismiss it; a day-old model with no Elo row is simply a model nobody has voted on yet.

The mistake to avoid is treating the 1,151-versus-nothing comparison as a win for either side. Krea 3 has a number and Images 2.6 does not, which tells you about the measurement calendar, not the models. The comparison that will actually settle this arrives when the AA boards add a 2.6 row — and when they do, the number to watch is whether it clears 1,199, because that is where OpenAI's own predecessor sits and roughly where a specialist editor has to be beaten to justify switching.

Until then, the cheapest way to get a real answer is to run your own twenty edits through both and count the ones you would ship. That costs about two dollars. It is also the only test currently available that is not somebody's marketing.