
MAI Image 2.7 Flash: Canvas's Engine at 29% of the Cost
- OrcaNEWOrcaCyber: Hy4 Flash Abliterated / Uncensored2026-11-18$0.07 / $0.28 per 1M tokens · 187 tok/s
- tencentNEWTencent: Hy4 Flash2026-11-16$0.02 / $0.07 per 1M tokens · 508 tok/s
- GoogleNEWGoogle: Gemini 4 Carbon2026-11-10$2.00 / $10.00 per 1M tokens · 87 tok/s
- AnthropicNEWAnthropic: Claude Fable 5.22026-11-09$8.00 / $40.00 per 1M tokens · 97 tok/s
- orcaNEWOrca: OrcaC2 1.02026-11-079 tok/s
- OrcaOrcaCyber: GLM 5.3 Abliterated / Uncensored2026-11-04$3.00 / $5.00 per 1M tokens · 89 tok/s
- OrcaOrcaCyber: Hy4 Abliterated / Uncensored2026-11-04$0.80 / $1.60 per 1M tokens · 133 tok/s
- OrcaOrcaCyber: Kimi K3.1 Pro Abliterated / Uncensored2026-11-04$2.00 / $5.00 per 1M tokens · 77 tok/s
- tencentTencent: Hy42026-11-04$0.83 / $2.50 per 1M tokens · 117 tok/s
- OrcaOrcaCyber: MiniMax M3.1 Flash Abliterated / Uncensored2026-10-29$0.80 / $1.60 per 1M tokens · 114 tok/s
- OrcaOrcaCyber: GLM 5.3 Flash Abliterated / Uncensored2026-10-27$0.80 / $1.60 per 1M tokens · 66 tok/s
- OrcaOrcaCyber: Qwen3.8 Flash Next Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 92 tok/s
- OrcaOrcaCyber: MiMo V2.6 Flash Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 41 tok/s
- MoonshotMoonshotAI: Kimi K3.1 Pro2026-10-2447Intelligence
- OrcaOrcaCyber: DeepSeek V4.1 Flash Abliterated / Uncensored2026-10-23$0.80 / $1.60 per 1M tokens · 212 tok/s
- minimaxMiniMax: M3.1 Flash2026-10-1929Intelligence
- OrcaOrca: OrcaCyber Ultra 1.02026-10-17$5.00 / $10.00 per 1M tokens · 111 tok/s
- deepseekDeepSeek: DeepSeek V4.1 Flash 10172026-10-1733Intelligence
- DeepSeekDeepSeek: DeepSeek V4.1 Flash 09102026-10-17$0.15 / $0.60 per 1M tokens · 168 tok/s
- DeepSeekDeepSeek: DeepSeek V4.1 Flash2026-10-1740Intelligence
MAI Image 2.7 Flash is the model behind Canvas, the sketch-and-edit workspace Microsoft turned on at labs.microsoft.ai this week — and it is not the model that took No. 1 on Artificial Analysis. MAI Image 2.7, the flagship, sits top of the editing board at 1,215 Elo and fourth on text-to-image. The Flash variant that actually runs the thing you can click sits seventh and eleventh, at $10.24 per 1,000 images against the flagship's $34.82. Canvas is the most concrete product Microsoft has built on the cheaper model, and for most people it will be the first place they meet either one.
That matters because the launch coverage of 17 November was almost entirely about the flagship: the No. 1 placement, the four-point lead over GPT Image 2.6 Sunburst, the one-fifth-of-the-cost claim. All of that is true and none of it describes the model you get when you open Canvas. This piece is about the cheap one — what it scores, what Microsoft has and has not documented about it three days after launch, and where a 36-point editing gap stops being worth 3.4 times the price.
What Canvas is, and what it will not keep
Canvas opens as a two-panel workspace: a prompt column on the left, an image on the right, and a mode switch between New image and Diagram. Microsoft's own launch copy describes the loop as turning sketches into polished diagrams, choosing a portion of an image to "replace, move, or restyle", and refining "without starting over". The page's inline instructions are more specific than the marketing: "Describe it in the panel, or start from a photo. Each run makes one image in about 10 to 20 seconds," and for the edit pass, "Change a word, or select an object in the image."
Two things about the implementation are worth knowing before you build a habit on it.
• Nothing persists. The workspace keeps a This session list of versions, but the page states plainly that "Results here aren't saved to a canvas" — close the tab and the branch you spent twenty edits building is gone. There is a Download button and nothing resembling a project, a share link, or a saved document.
• The published surface is thin. Default aspect ratio is square 1:1, the prompt box counts against a 1,000-character limit, and no output resolution is stated anywhere on the page. Microsoft's model page for MAI Image 2.7 does not give one either.
One naming correction, since it will confuse anyone following the coverage: the product is MAI Labs — the page title is "MAI Labs | Microsoft AI", and it carries the disclaimer "a limited preview of new and experimental AI models." playground.microsoft.ai resolves to the same application, which is presumably why the workspace is being described as an MAI Playground feature. Same surface, two names.

The model under it is not the one on the board
Microsoft's launch post is explicit that Canvas is "powered by our Flash model", and the workspace's own shipped configuration confirms it — with a wrinkle worth knowing. The application sends its model list to the browser, and the entry with "id":"mai-image-2-7-flash" carries "sidebarLabel":"MAI-Image-2.7". So the picker a visitor sees says MAI-Image-2.7 while the model actually answering is the Flash variant. Artificial Analysis rates the two separately, and they are far enough apart that nobody should treat them as one product with a knob turned down.
On AA-Image-Editing v2.1, read on 20 November 2026: MAI-Image-2.7 is first at 1,215 Elo from 5,478 votes, with a 95% interval of 1,201 to 1,229. MAI-Image-2.7-Flash is seventh at 1,179 from 5,891 votes, interval 1,166 to 1,192. On AA-Image-T2I v2.1 the gap widens: 1,214 for the flagship at fourth, 1,158 for Flash at eleventh. Those are Artificial Analysis's own measurements with sample counts attached, not Microsoft's.
Thirty-six points on editing is not a rounding artefact. Each row carries a band of roughly ±13 to ±14, and the difference between two independent estimates is wider than either band alone — around ±19. A 36-point step clears that with room to spare, which is the opposite of the situation on the top of the board, where the flagship's four-point lead over GPT Image 2.6 Sunburst sits well inside the noise and should not be read as a settled ranking.

Flash has a score and almost no paperwork
Here is where the three-day-old release gets odd. Open microsoft.ai/models and there is one image model listed: MAI Image 2.7, at /models/mai-image-2-7/. That page describes Flash in a paragraph-sized card — "Built for fast, cost-efficient image generation without compromising on quality" and "less than half the cost of our flagship model" — and its own generator dropdown offers both entries. There is no /models/mai-image-2-7-flash/ page, no datasheet, no resolution, no latency figure.
The Azure side is thinner still. Microsoft's "Foundry Models sold by Azure" reference lists MAI-Image-2.5, MAI-Image-2.5-Flash, MAI-Image-2.5-Pro, MAI-Image-2.6 and MAI-Image-2.6-Flash, each marked Preview with a documented 32,000-token context, PNG output and "up to five images" as input. Neither MAI-Image-2.7 nor MAI-Image-2.7-Flash appears in that table, even though the launch post says both are in public preview on Microsoft Foundry. Three days is not long; the 2.6 family's own documentation did not arrive instantly either. But it means that as of today the only public numbers for Flash are somebody else's measurements.
What the Labs page does publish, unintentionally, is its own rate policy — the application ships its model configuration to the browser, and it can be read there:
• Per signed-in user — Flash: 4 requests a minute, 40 a day. Flagship: 2 a minute, 20 a day.
• Per IP — Flash: 6 a minute, 120 a day. Flagship: 3 a minute, 60 a day.
• Per model, site-wide — Flash: 30 a minute. Flagship: 2 a second with a 120-a-minute ceiling.
Two things follow. First, Microsoft has given the cheap model exactly double the per-user quota of the expensive one, which is what a high-volume engine looks like rather than a demo. Second, the same configuration caps image input at four per request for both models, while Foundry's documentation for the 2.6 family states five. That is a documentation mismatch, not a measured regression — Microsoft publishes no 2.7 datasheet to contradict either number — but anyone porting a five-reference editing workflow onto the new family should confirm the cap before assuming it carried over.
Where a 36-point gap stops mattering
The price ratio is the whole argument, so it is worth stating precisely rather than repeating the vendor's framing. Microsoft says Flash costs "less than half" the flagship. Artificial Analysis's measured rates on both boards are $10.24 and $34.82 per 1,000 images — Flash at 29.4%, or a little under three-tenths. Microsoft's claim is directionally right and conservative against the independent numbers; the two agree closely enough that neither should be treated as spin.
Put a volume on it. At 10,000 images a month, the flagship is $348.20 and Flash is $102.40 — a difference of $245.80 a month, or roughly $2,950 a year, for 36 Elo. Whether that is a bargain depends entirely on what the images are for. A diagram redrawn six times before it reads correctly is a different purchase from a hero image that ships once.
Against the rest of the board, Flash's position is more interesting than "the cheap Microsoft model" suggests. On editing it sits seventh, between Kling Image 4.0 at 1,183 and Seedream 5.5 at 1,160 — ahead of Krea 3 Ultra, Nano Banana 2.1 and both earlier MAI generations, and behind only the five GPT Image tiers plus Kling. At $10.24 per 1,000 images that is:
• versus GPT Image 2.6 Sunburst — 32 Elo behind (1,179 against 1,211) at 4.9% of its $210.72 rate.
• versus Muse Image — 50 Elo ahead (1,179 against 1,129) for 2.4% more than its $10.00 rate.
• versus Seedream 5.5 — 19 Elo ahead (1,179 against 1,160) at under half its $28.00 editing rate.
Read those three lines together and the Flash model is not a compromise purchase. It is the cheapest row on the board that clears 1,175 Elo, and the model above it costs four times as much for eight more points per 1,000 images.
Neither MAI model is in the OrcaRouter catalogue as of 20 November 2026, and this piece will not pretend otherwise: what Microsoft exposes today is a limited, non-commercial Labs preview, and the production route runs through Foundry rather than through us. What is routable is the comparison set — GPT Image 2.5 Sunburst and GPT Image 2.5 Flare are callable on one key at the provider's list price with 0% markup, which is the standing arrangement for every model we carry: no uplift on top of what the provider charges, so a vendor price cut lands here the same day rather than at the next contract renewal. If Microsoft ever publishes a Foundry rate for MAI Image 2.7 Flash, that is the arrangement it would be measured under.

What would turn this into a decision
Three things are missing, and each of them is a number somebody could publish next week.
A resolution. Canvas states none, the model page states none, and the Foundry reference does not cover the family. Without it, "production-ready visuals" is a phrase rather than a specification, and the 4,194,304-pixel ceiling Microsoft documented for the flagship has no confirmed counterpart on Flash.
A rate card. AA's $10.24 is a measurement of what Microsoft is charging through the surfaces it can see, not a published tier. Until a Foundry price sheet lists MAI-Image-2.7-Flash the way it lists the 2.6 family, the number is a good estimate and not a contract.
An independent head-to-head on the task Canvas is actually for. Every figure above comes from the general editing and text-to-image boards, and neither board isolates the thing the Diagram mode claims — turning a rough sketch into a legible, correctly-labelled diagram. That is a narrow capability, and it is the one Microsoft is selling. Until someone runs it, the honest position is that Canvas is a well-built workspace on a well-scored but thinly documented model, free to try, non-commercial, and not yet a thing to build a pipeline on.
For anyone deciding today: if the job is a handful of diagrams a week and you are inside Microsoft's ecosystem, Canvas is the cheapest way to see what the 2.7 generation does with a sketch, and its rate limits — 40 images a day per account — are generous enough to form a real opinion. If the job is volume, the number to watch is not the Elo. It is whether $10.24 survives contact with a published Foundry rate.
