MAI Image 2.7 Preview vs Krea 3: The One Elo Gap in This Series You Can Actually Trust

Of the four editing models MAI Image 2.7 Preview is being measured against this week, Krea 3 is the only one whose score sits on the same ruler. Both appear on Artificial Analysis's AA-Image-Editing v2.1 board, and on that board Krea 3 Ultra holds 1,151 Elo — the same figure as MAI-Image-2.6 and Nano Banana 2.1, three models tied at the top of the second tier. MAI Image 2.7 Preview debuted at 1,193, which puts 42 points between the two.

That is a real number in a way the other comparisons are not. It is the same board, the same scoring run, and the same snapshot. It is also small enough to be fragile: a 42-point Elo difference over a leaderboard with confidence intervals is a genuine advantage in aggregate, not a guarantee on your specific asset.

Which makes the interesting question not "which is better" but "what does Krea 3 do that a model with 42 more points does not," because the answer turns out to be a long list.

Same Board, and What 42 Points Buys

Artificial Analysis scores 2.7 at 1,193, second overall, six behind GPT Image 2.5 Sunburst at its max tier and eleven ahead of GPT Image 2.5 Flare. Krea 3 Ultra's 1,151 places it in the group immediately below, level with Microsoft's previous generation.

Category detail sharpens it. Microsoft says 2.7 improved across all ten use cases and all seven editing actions relative to 2.6, and Artificial Analysis has it first on four of the seven actions — enhancement and restoration, reasoning-based edits, scene and style, and text or symbol editing — with the widest margin in scene and style at 24 points. The three it does not lead, composition, individual-object edits and identity preservation, are held by OpenAI.

Read against Krea 3, that profile suggests the 42 points are concentrated rather than spread. A Krea 3 user moving to 2.7 would expect the largest gains on whole-frame work — restyle this scene, relight this photograph, restore this scan — and much less on the regional, masked, precisely-bounded edits that Krea 3 is explicitly built for.

Eight Seconds Against Whenever You Get a Turn

Krea 3's defining number is latency. The company claims end-to-end edits in under eight seconds at native 2K, with the fastest configurations landing around four to five, and describes the model as not yet real-time but close enough to feel interactive. Whether the number holds under your own load is a separate question, but the claim is specific and checkable in a way most launch claims are not.

MAI Image 2.7 Preview has no latency figure because it has no endpoint. Its interaction model is a browser tab in MAI Playground, one image at a time, and Microsoft has published no throughput, queue or rate information for the preview. The earlier MAI-Image-2 playground was reported to allow roughly fifteen generations a day with a wait between them; that figure was never confirmed for 2.7 and should not be assumed.

The consequence is structural rather than incremental. An eight-second edit is fast enough to sit inside a design tool, which is why Krea positions 3 as an editor for commercial teams rather than a generator with an edit mode. A playground round trip is fast enough to evaluate and far too slow to build on. Comparing the two on speed is not comparing two speeds; it is comparing a latency figure to an absent one.

LoRA Training Is the Capability With No Counterpart

Krea 3 supports training a LoRA on a specific product, person or style, and the company claims it is the first editing model trainable this way. Whether or not the "first" holds up, the capability is the most consequential thing in either spec sheet.

An editing model that generalises is useful for one-off work. An editing model you can tune on your own catalogue is a different kind of asset: it learns what your product looks like from angles the prompt cannot describe, and it stops needing a reference image with every request. For a studio with a fixed product line, that converts a per-call cost into a one-time training cost plus cheaper calls.

MAI Image 2.7 Preview offers nothing comparable, and could not in its current form — there is no training interface, no weights, and no API. Microsoft's model is a capability you rent in a browser; Krea's is a capability you can adapt.

Style references are the lighter version of the same idea, and Krea 3 supports those too, alongside regional editing and bounding-box control. The pattern is consistent: Krea ships controls, Microsoft ships judgment.

Reference Counting Is a Mess on Both Sides

Krea 3's documentation disagrees with itself about how many reference images it accepts. The launch post says six. The product page and API documentation say eight. The API downsizes references beyond the fifth and recommends four or fewer for full detail, and a Zero Data Retention mode caps the count at four. All of those are Krea's own numbers, published at different times for different surfaces.

That is worth flagging not as a deception but as a reminder about how spec sheets age. A reference ceiling is a product of resolution, memory and billing model, and it changes as a vendor tunes the serving stack.

For MAI Image 2.7 Preview the equivalent number does not exist at all. Microsoft has published no reference maximum, no resolution, no quality tiers and no mask convention. The predecessor MAI-Image-2.6 accepted up to five references and capped output near 1536×1536; assuming 2.7 inherits either limit is inference. On this dimension Krea 3 is not better documented — it is documented, and the ambiguity is visible rather than absent.

Thinking Modes and the Cost of Being Careful

Krea 3 exposes fast and high thinking modes, and the pricing reflects the difference: $0.06 per image at 2K on the standard setting and $0.08 on high thinking at 2K, with $0.05 and $0.03 at 1K respectively. Input images bill separately at $0.02 per megapixel. Output goes to native 2K with a 4K option for print and large-format work.

The two-mode design is a good fit for editing, because edit difficulty is wildly bimodal. Straightening a horizon and rebuilding a reflection are not the same job, and paying the same rate for both overcharges the easy case. A team that routes simple edits to the fast mode and complex ones to high thinking can cut its bill substantially without a quality decision on every request.

MAI Image 2.7 Preview has no tiers, no modes and no price. Its quality is whatever the model does, and its cost is the analyst's afternoon.

CPI-Bench Against a Leaderboard

Krea cites a CPI-Bench score of 4.81 out of 5 for Krea 3 Ultra, placing it behind GPT Image 2.5 and ahead of GPT Image 2, Seedream 5 Pro, and Nano Banana 2 and Pro. That is a vendor-reported benchmark result and should be read as one — it is not an independent ranking, and the comparison set is chosen by the company publishing it.

It does, however, point at something the Elo board does not capture well: user preference on commercial creative work, which is what CPI-Bench is built around. Krea's own pitch is that the model is 10× faster than GPT Image 2 and 4× faster than Nano Banana Pro, and a benchmark that rewards usable output per unit time will favour a fast model even when a slower one scores higher in blind pairwise comparisons.

Microsoft's evidence runs the other way. Its numbers are leaderboard placements — second on the editing board, third on Arena Edit with 1,470 points and a 45-point gain over 2.6, which makes Microsoft the second-ranked lab on that board — plus a set of category claims derived from those runs. Both are legitimate; neither is a head-to-head against the other model.

The Verdict Splits by Job, Not by Score

If you are building an editing feature into a product this quarter, Krea 3 is the answer by default, and not because of its Elo. It has an API, a published price, a latency figure, reference controls, regional masks, a fast mode for cheap edits, and LoRA training for your own catalogue. MAI Image 2.7 Preview has none of those things, and its 42-point lead does not survive contact with a pipeline that cannot call it.

If you are choosing an editor for work a human will do by hand, the ranking is worth something. Scene restyling, relighting and restoration are where 2.7's margin is concentrated, and a designer working in Playground can spend that advantage directly.

If you are trying to decide whether the frontier has moved far enough to switch, the honest read is that it has moved by 42 points on one board, which is meaningful at population scale and small at the scale of a single difficult image. The practical way to settle it is to run your five hardest edits through both — the Arena comparison for 2.7 costs nothing, and Krea 3 costs about thirty cents for the same five at 2K.

OrcaRouter does not host either model, so this is not a routing decision today. It is a decision about what to build your editing layer on, and the layers that are routable right now — the OpenAI image tiers that bracket 2.7 on that board, at provider list price with no markup added — are the ones you can put behind a fallback chain while a preview-grade model proves itself.