Krea 3 vs MAI Image 2.6 Preview: 16,400 Votes Against a Deadline

Krea 3 publishes a complete price list and cannot tell you whether the model is any good. MAI Image 2.6 Preview, Microsoft AI's image model, has 16,400 blind human comparisons behind its editing score and cannot tell you what an edit costs. That is the swap at the centre of this matchup, and it does not resolve into a winner — it resolves into which kind of unknown you can absorb. Krea shipped Krea 3 on 15 October 2026; Microsoft put MAI-Image-2.6 into public preview on 10 August 2026 and refreshed that notice on 4 September. One is one day old and priced to the cent. The other is two months old, ranked on an independent board, and sold only through Microsoft Foundry.

What each side can actually prove

Start with the evidence, because on this pair it is wildly lopsided and the lopsidedness runs the opposite way from the pricing.

• MAI Image 2.6 Preview, measured. On Artificial Analysis's image-editing leaderboard, read on 16 October 2026, MAI-Image-2.6 sits at 1,151 Elo from 16,400 samples — fourth on the board, with a 95% interval of 1,139 to 1,163. On the text-to-image board the same model is sixth at 1,165 Elo from 7,993 votes. Its cheaper sibling MAI-Image-2.6-Flash is fifth on editing at 1,144 from 10,505 samples. Those are blind pairwise human preference votes, collected by a third party Microsoft does not control.

• Krea 3, unmeasured. No entry on the editing board, no entry on the text-to-image board, no third-party latency run. What exists is Krea's own claim — under 8 seconds end to end at native 2K — measured on Krea's own edit traffic between 24 September and 1 October 2026, with the competitor timings in the same table gathered the same way.

• The price column, where the two boards are kinder to the Microsoft model than Microsoft is. Artificial Analysis lists MAI-Image-2.6 at $38.90 per 1,000 images and MAI-Image-2.6-Flash at $19.50. Krea prices Krea 3 at $0.06 for a 2K edit before references, and references are billed separately: +$0.02 each at 1K, +$0.07 each at 2K, plus $0.02 for high/ thinking mode.

One of these models has been voted on 16,400 times by people who could not see the label. The other has been timed by its own maker. If your question is "is it good," only one of them has an answer today, and it is not the new one.

The 1,336 that Microsoft quotes is not the number the board shows

Microsoft's launch post leads with second place. It should, and it is not the whole picture.

• Microsoft's own chart plots MAI-Image-2.6 at 1,336, ranked No. 2 behind GPT Image 2 Medium's 1,381 and ahead of Grok Imagine Image 2.0's 1,316, sourced by Microsoft to arena.ai, measured on text-to-image. • The Artificial Analysis text-to-image board we read on 16 October 2026 puts MAI-Image-2.6 at 1,165 — sixth, behind GPT Image 2.5 Sunburst at 1,218, GPT Image 2.5 Flare at 1,208, GPT Image 2 (high) at 1,186, Nano Banana 2.1 at 1,172 and Grok Imagine Image 2.0 at 1,167.

Those are two different boards with different Elo scales, so 1,336 and 1,165 are not in conflict and neither is wrong — but they support different sentences. "Second" belongs to the Arena board Microsoft cites. "Sixth of the top tier with a 95% interval of 1,153 to 1,177" belongs to the board that lets you look at the votes. On the editing board, which is the board that matters for a model you would use to change an existing image, MAI-Image-2.6 is fourth at 1,151, with Nano Banana 2.1 third at 1,152 — a one-point gap between third and fourth, comfortably inside both models' margins.

Read that as what it is: the Microsoft model is genuinely among the best editing models on the independent board, and it is genuinely not the second-best model in the world. Both statements come from public data.

The pricing asymmetry, worked through in both directions

Run the arithmetic on Krea's meter and the two products stop sitting in different price brackets.

• A single 2K edit, one reference: $0.13.

• Four 2K references — the count Krea's own docs call the full-detail ceiling — $0.34 an edit, or $34,000 per 100,000 edits.

• Six 2K references, the count the launch post advertises: $0.48 an edit, $48,000 per 100,000. Krea's own chart gives 100,000 Krea 3 edits at 2K as $6,000 per month, explicitly "without image references."

• MAI-Image-2.6 Preview at the independent board's listed rate: $38.90 per 1,000 images, or $3,890 per 100,000 — and that comparison is unfair to Microsoft in one direction and to Krea in the other, because the Microsoft figure has no accompanying note about batches, references or retries, and there is no way to verify it against a vendor rate card, because there is no vendor rate card.

The one thing you can say without hedging: Krea 3 is cheap per image and expensive per composition, and the multi-reference composition is the reason to use it. The model's headline price and its characteristic use case pull in opposite directions, and only one of them appears in the launch post's cost chart.

Reference limits: one number versus three

Microsoft publishes one answer. Krea publishes three, and the difference between them is the difference between a plan and a surprise.

• MAI-Image-2.6 Preview accepts text plus up to five images for image-editing workflows, per Microsoft's Foundry model card, which also sets the output constraint: a minimum of 768 × 768 pixels and a maximum total pixel count of 2,359,296, equivalent to 1536 × 1536. One number, one page, no asterisk.

• Krea 3 takes up to 6 references per the launch post and its FAQ, up to 8 per the product page, and up to 8 in the API's own image_urls/ parameter row — with the additional rule that from five references onward each one is resized to at most 1536 px on its longest side, and the documentation's own advice to "send four or fewer" when fine detail matters. Enable Zero Data Retention and the cap drops to 4 references, because every reference must then travel as a base64 data URI.

For anyone whose use case is the multi-reference one — a character, a garment and a location in a single frame — the effective Krea 3 number is four, and the four-reference edit is $0.34, not $0.06. That is a reasonable price. It is not the launch price.

Where each of these belongs in a production pipeline

Krea 3's proposition is throughput: render only the region that changes, chain edits without accumulating artifacts, and put a whole catalogue through in the time a competitor spends on a hero shot. Its regional_edit/ flag is the mechanism — Krea's docs warn that whole-image edits degrade when you keep editing the model's own output, which is the practical argument for a model that regenerates only what moved. If your workload is "change the background on 4,000 product photos," that architecture is the point, and the independent score you are giving up is a score about something else.

MAI-Image-2.6 Preview's proposition is measured quality behind a Microsoft contract. It is the model to reach for when the output is judged by a person — campaign imagery, brand work — and when the buyer already has Foundry in the stack. What it is not is plannable at volume: with no published rate card, its cost per image is a number from someone else's table, and the resolution ceiling of 2,359,296 total pixels sits under Krea's native 2K output in every aspect that is not nearly square.

Building on either one without betting the quarter on it

Neither of these models is in OrcaRouter's catalogue today, and we are not going to imply otherwise: OrcaRouter routes neither Krea 3 nor MAI-Image-2.6 Preview. Krea 3 is reached through Krea's own early-access API, which needs a workspace in the beta; the Microsoft model is reachable only through Foundry and MAI Playground.

That is exactly the situation a router is built for, though not in the way a launch post usually frames it. When one of two candidate models has no published price and the other has three published reference limits, the sane move is to build the pipeline so that swapping the model is a configuration change. A single OpenAI-compatible endpoint across 200-plus models — with provider list prices passed straight through and no markup on top, so a vendor price change is live the same day rather than at the next contract renewal — makes the swap cheap. Automatic failover that reroutes a failed call to a fallback instead of into your queue makes the beta survivable. What it does not do is make either of these two image models callable today, and the pieces we do route, including the GPT Image and Gemini image families, are the ones to keep under the workloads that have to ship while the two models above sort themselves out.

If you need a model this week and you cannot get an invitation, the decision is made for you. If you can get both, the trade is legible: Krea 3 gives you speed, native 2K and a meter you can forecast, and asks you to trust a consistency claim nobody has audited. MAI Image 2.6 Preview gives you 16,400 human comparisons and a hard resolution ceiling, and asks you to plan a budget around a number that lives on someone else's leaderboard. Pick the unknown you can price.

None of this is a reason to avoid the model, only a reason to know the history: the documented AI incidents and agent failures behind models in this class are collected in the OrcaRouter AI Incident Archive, where each entry carries its own sources, severity, confidence grade and disputed flag.