MAI Image 2.7 Preview vs Nano Banana 2.1: What 42 Elo Actually Buys

Almost every number in the MAI Image 2.7 Preview launch is impossible to compare with anything. The model debuted on October 27, 2026 at 1,193 Elo on Artificial Analysis's AA-Image-Editing v2.1 board, and most of the models you would want to place against it were measured weeks earlier on a board that has since changed. Nano Banana 2.1 is the exception. Google's model sits on the same version of the same board at 1,151, from the same 22 October 2026 evaluation — so 1,193 minus 1,151 is a subtraction you are allowed to make. The gap is 42 Elo, and it is the only clean head-to-head this launch gives you.

That makes the two models the natural pair for the question everyone actually has: is the best editor on the board worth using over the cheapest good one that ships with masks, 4K output, fourteen reference images and a published price? The answer turns out to depend on a piece of arithmetic nobody has run on Nano Banana 2.1's price cut.

Why This Comparison Is the Only Fair One on the Board

Leaderboards are not thermometers. They are moving populations — models enter, models get deprecated, votes accumulate — and a score is only meaningful relative to the other numbers recorded in the same sweep. This is the trap in most coverage of the 2.7 launch. Microsoft's own chart puts 2.7 at 1,193, GPT Image 2.5 Sunburst at 1,199, GPT Image 2.5 Flare at 1,182, MAI-Image-2.6 at 1,151, Krea 3 Ultra at 1,151, Nano Banana 2.1 at 1,151, MAI-Image-2.6-Flash at 1,144 and GPT Image 2 at 1,136 — all from the same AA-Image-Editing v2.1 run, which is what makes the whole chart quotable. Six Elo separates 2.7 from the top and that six points is within the noise a blind-vote board can move in a week.

What is not noise is the shape of the distribution. Between 1,136 and 1,199 sit eight models from four labs, all within 63 points. In editing, the frontier is crowded, and 42 points — the 2.7-to-Nano-Banana-2.1 gap — is about two thirds of the entire spread of the top eight. It is a real separation inside a compressed field, and it is also small enough that a specific workload can invert it.

Spec Lines, With the Gaps Marked

• Release — Nano Banana 2.1 arrived October 6, 2026, replacing Nano Banana 2 and sitting below Nano Banana Pro in Google's line; MAI Image 2.7 Preview arrived October 27, 2026.

• Scope — Nano Banana 2.1 generates and edits, built on Gemini 3.6 Flash; 2.7 is editing only, with text-to-image promised but not shipped.

• Output — Nano Banana 2.1 offers 1K, 2K and 4K with aspect ratios from 1:8 to 8:1; Microsoft has published no resolution for 2.7.

• References — up to fourteen for Nano Banana 2.1, split as four characters and ten objects kept consistent; unpublished for 2.7.

• Control — Nano Banana 2.1 supports mask-based local editing, where you paint a region and only that region changes, plus three thinking levels; Microsoft describes 2.7's strengths by use case — restyling, relighting, text and symbol edits — with no published mask interface.

• Price — Nano Banana 2.1 is $0.0336 per 1K image, $0.0504 at 2K and $0.113 at 4K through the Gemini API, with batch jobs half price again; 2.7 has no published rate.

• Access — Nano Banana 2.1 is in the Gemini app, AI Studio, the Gemini API, Search's AI Mode, Google Ads, Flow and Stitch; 2.7 is in MAI Playground, Arena and Artificial Analysis, with no API and no Foundry listing.

The table you would draw from that is not close on anything except the score. So the score had better be worth a lot.

The Famous 50% Price Cut Is Smaller Than It Sounds

Google's headline for Nano Banana 2.1 was that image output costs half what it did: $30 per million output tokens, down from $60. That is true, and for pure text-to-image it is roughly a halving, because most of the bill is output.

Editing is a different shape of request. An edit carries the source image in as input tokens as well as producing an image out, and Google raised input prices at the same time — from $0.50 to $1.50 per million tokens, a tripling. Text and reasoning output went from $3 to $7.50 per million. So the actual saving on an edit is not 50%; it is the difference between two linear costs, and it flips sign when the input side gets large enough.

Put it as an inequality. Let I be input tokens and O be output tokens for one request. The old request cost 0.5I + 60O and the new one costs 1.5I + 30O. The new price is cheaper exactly while 1.5I + 30O < 0.5I + 60O, which simplifies to I < 30O: input has to stay below thirty times output. That is a generous margin for a 1K edit and a much thinner one for a 4K edit with a large reference set, because reference images are input and a 4K output is not thirty times the token count of the references that produced it.

The practical consequence: if your workload is "one source image in, one similar-size image out" — which describes most product-photo and retouching pipelines — Nano Banana 2.1's edit pricing is materially better than its predecessor's, but not by half, and a 4K pipeline with four or five references may see almost none of the cut. If your workload is prompt-only generation, you get the full halving, and batch mode halves it again.

Where the 42 Elo Would Show Up, If It Does

Microsoft's category detail is more useful than its overall score, because it says where the gain came from. Between MAI-Image-2.6 and 2.7, reasoning-based edits climbed from seventh to first, scene and style edits from seventh to first, and text or symbol edits from sixth to first. The largest single lead is 24 Elo in scene and style over the next-best model. Those are vendor statements about Artificial Analysis's categories, and the pattern is coherent: 2.7 got better at the edits that require a judgement about the whole image rather than a local repair.

Nano Banana 2.1's published weaknesses are almost the photographic negative of that. Google's own model card admits blurry small text at 1K, occasional left/right confusion, imperfect character likeness, and limited 3D reasoning and factuality. Left/right confusion is a spatial-reasoning failure; blurry small text at 1K is a resolution-dependent one. Neither is the kind of error a scene-relight improves.

So the honest reading of the 42 Elo is that it is concentrated where Microsoft says it is: whole-image restyling, relighting, text integrated into an existing composition, and edits that need the model to understand how objects and light relate. If your work looks like that, the gap is likely to be visible. If your work is masking one object out of a product shot and dropping in another, you are not testing the axis on which 2.7 leads, and a mask you can actually draw is worth more than a score.

The Mask Question Is the One That Decides It

This is the least glamorous difference and the one that most often settles a real integration. Mask-based editing means a human has already decided what changes; the model's job is to make the change look native. That division of labour is why editing tools in production pipelines are reviewable: a reviewer can see the painted region and judge only that region.

MAI Image 2.7 Preview, as documented, takes an instruction and an image. Microsoft's description of the capability set — restyle the scene, change the text, restore the asset — is instruction-driven throughout, with no published mask parameter, no bounding-box control, and no documented reference-image maximum. It may well accept masks; there is simply no published interface to program against, which for a developer is the same problem.

For a creative director working image by image, instruction-driven is faster and often better. For a pipeline that has to produce an audit trail, a mask is not a nice-to-have.

Where Each One Lives, and What That Costs You to Reach

Nano Banana 2.1 is one of the most widely distributed image models in the world: the Gemini API, AI Studio, the Gemini app, Search's AI Mode, Google Ads, Flow and Stitch. Distribution is not a spec, but it determines how much integration work you do before you learn whether the model fits.

MAI Image 2.7 Preview is in three places, two of which are evaluation surfaces. That is a deliberate choice for a preview, and it is also a wall: a team that wants to test the 42 Elo on its own assets has to do it by hand in a browser, one image at a time, and cannot script the experiment.

On the pricing side there is a concrete piece of plumbing worth knowing. OrcaRouter carries Google's image line at provider list price with 0% markup — Nano Banana 2 at $0.15 per request and Nano Banana Pro at $0.24 today — so a vendor price move lands on our side the same day it lands on Google's rather than waiting for a renegotiation. Whether 2.1 itself is added is a catalogue decision, not a pricing one; when it is, the halved output rate is what you will be charged, because there is nothing in between you and the list price.

The Decision Rule

If your edits are whole-image judgements — relight this scene, restyle this room, put new text on this sign and make it look printed — use MAI Image 2.7 Preview in Arena or Playground this week and compare it against Nano Banana 2.1 on ten of your own images. That is the exact axis where 2.7 claims its 24-point lead, and blind comparison is the only test that transfers.

If your edits are local and specified — replace this label, remove this cable, extend this background — Nano Banana 2.1 wins on the ability to say precisely what should change, at $0.0336 for a 1K output, with 4K available when the destination needs it and fourteen reference images when brand consistency needs it. The 42 Elo never gets a chance to matter, because you are not asking the question it answers.

If you need an editor in a pipeline today and neither of those describes you, the models on that board are the ones with published interfaces. MAI Image 2.7 Preview will be a genuinely interesting option the day it has one; the leaderboard has already told you what to expect from it, and the only thing missing is the ability to find out for yourself.