GPT-Image-2.5 vs MAI-Image-2.6 Preview: The Editing Leader You Can't Buy vs the Newcomer You Can

September 2026 is the first month in a year when the two most interesting image models are both brand new and you can only call one of them. On September 8 OpenAI made GPT-Image-2.5 generally available — the ChatGPT Images 2.5 engine, sold through the API as GPT-Image-2.5 Flare and GPT-Image-2.5 Sunburst — and anyone with an OpenAI key can hit it immediately. Microsoft's MAI-Image-2.6 Preview, which currently sits at the top of the independent image-editing leaderboards, has been in private preview since August 19 with invites gated through the MAI Playground and Microsoft Foundry, and still has no public price card. The editing leader is not for sale. The newcomer you can actually buy has not yet been scored by anyone except its maker. That is the strange shape of the image-model market this comparison lives in.

Worth stating once so the rest of the piece reads correctly: every capability claim for both models is vendor-reported and unreproduced, and every ranking below comes from independent human-preference boards — Artificial Analysis and LMArena — which neither OpenAI nor Microsoft controls. GPT-Image-2.5 has no independent score yet: it appeared on no leaderboard in its first 24 hours. MAI-Image-2.6 Preview has plenty, because it has been testable in preview for three weeks.

What the independent boards actually say right now

On the Artificial Analysis image-editing board, MAI-Image-2.6 holds the top position in the September 2026 snapshot, narrowly ahead of GPT Image 2 (high), with Meta Muse Image and Nano Banana 2 close behind. That #1 slot is not a fluke of one snapshot — it took the top spot from Microsoft's own MAI-Image-2.5-Pro in late August and has held it through the board's rebuild. On the separate text-to-image board the picture is different: GPT Image 2 (high) — the model GPT-Image-2.5 is meant to supersede — is #1, and MAI-Image-2.6 sits at #2. On LMArena's single-image-edit leaderboard the preview model climbed to second or third behind GPT Image 2 (medium), with its largest gains in text rendering and in product/branding/commercial design.

The caveats matter as much as the ranks. These are votes on a "preview" build, and Microsoft has said the version it sells at general availability may differ from what testers are judging. Vote counts for a three-week-old model are far lower than for incumbents, so the Elo numbers are noisy. And the two boards use different scales and different voting pools — MAI-Image-2.6 being #1 on AA editing while only #2 or #3 on LMArena is normal, not a contradiction.

The gated leader: what MAI-Image-2.6 Preview actually is

MAI-Image-2.6 Preview is Microsoft AI's unified image model — one system that does both text-to-image generation and image editing — built by Mustafa Suleyman's Microsoft AI organization and released August 19, 2026 into a private preview on the MAI Playground and Microsoft Foundry. "Preview" is doing real work in that name: access is by invitation, there is no public pricing, and the model is not purchasable through a normal API key. Its sibling MAI-Image-2.6-Flash, a faster variant that also appears on the AA board just below the main model, shipped in September.

What makes the preview worth watching despite the gate is that the independent signal points the same direction as Microsoft's claims. Microsoft says MAI-Image-2.6 is built for editing precision and text rendering; the arena data backs the text-rendering story specifically — the biggest single improvement over MAI-Image-2.5 is in text rendering, up 43 points on LMArena. The gap between "tops the editing board" and "you cannot give Microsoft money for it yet" is the entire tension of this matchup.

The buyable newcomer: what GPT-Image-2.5 claims

GPT-Image-2.5 is OpenAI's answer to the same problem Microsoft aimed MAI-Image-2.6 at: image models that edit as well as they generate. OpenAI's own framing splits the model into two endpoints — GPT-Image-2.5 Flare, the fast default that OpenAI says generates up to 50% faster than GPT-Image-2, and GPT-Image-2.5 Sunburst, which trades speed for tighter edit control and is aimed at production creative work. The claims that matter for an editing comparison are reference fidelity (a subject stays recognizable through style and setting changes), multi-turn editing that alters only what was asked, and text rendering that no longer garbles Chinese characters.

What GPT-Image-2.5 does not yet have is the one thing MAI-Image-2.6 Preview does: an independent number. Its predecessor GPT Image 2 (high) still tops the AA text-to-image board and ranks #2 on editing, so OpenAI's new model has to prove it beats its own predecessor before it can credibly challenge the model ranked above that predecessor in editing. OpenAI reports that third-party tests by Manus measured 2–4× faster generation and improved transparent-background output, but speed was never the contested axis in this matchup — editing quality is, and on that axis the only evidence for 2.5 so far is OpenAI's word.

How the two compare on the axes that decide a purchase

• Status — GPT-Image-2.5: generally available, September 8, 2026. MAI-Image-2.6 Preview: private preview, invite-gated since August 19.

• Access — GPT-Image-2.5: any OpenAI API key, two endpoints. MAI-Image-2.6 Preview: MAI Playground and Microsoft Foundry, by invitation, no public self-serve API.

• Price — GPT-Image-2.5: $8 per million image-input tokens / $30 per million image-output tokens (same rate card as GPT-Image-2); per-image cost unpublished. MAI-Image-2.6 Preview: no public price card; Artificial Analysis lists roughly $38.9 per 1,000 images for the preview build.

• Editing evidence — GPT-Image-2.5: OpenAI-reported only; no independent board entry as of September 9. MAI-Image-2.6 Preview: #1 on the AA image-editing board (September 2026 snapshot); #2–3 on LMArena single-image edit.

• Text-to-image evidence — GPT-Image-2.5: none independent. MAI-Image-2.6 Preview: #2 on AA text-to-image, behind GPT Image 2 (high).

• Product shape — GPT-Image-2.5: a fast endpoint and a precision endpoint, chosen per job. MAI-Image-2.6 Preview: one unified model, plus a faster MAI-Image-2.6-Flash sibling.

Both models have a "preview"-grade honesty problem in opposite directions: OpenAI sells a GA model with zero independent validation, and Microsoft has independent validation for a model you cannot buy.

Who should act, and who should wait

If you are a Microsoft Foundry or MAI Playground customer who can get an invite, MAI-Image-2.6 Preview is the highest-scoring editing model on an independent board and you should be testing it against your own edit workloads now — that is the entire point of a preview. Everyone else faces a real choice with no perfect option. GPT-Image-2.5 is the only genuinely new flagship image model you can call today, and for edit-heavy production it is a reasonable bet on OpenAI's track record — but it is a bet, and you are paying token-metered prices while OpenAI has not published what an image costs. The disciplined move for a team that needs independent validation is to keep GPT Image 2 (high) — the proven predecessor, still #1 on text-to-image and #2 on editing, and available through OrcaRouter at provider list price with no markup — as the production baseline while you run side-by-side tests of GPT-Image-2.5 against it, and to re-check the boards the moment MAI-Image-2.6 opens up. Neither GPT-Image-2.5 nor MAI-Image-2.6 Preview is on OrcaRouter's catalog as of September 9, 2026, so for now you hold two vendor keys either way.

The near-term tells are simple. Watch whether MAI-Image-2.6 keeps its #1 editing slot as its vote count grows and whether Microsoft publishes pricing — that turns "you can't buy it" into "here is what it actually costs." Watch whether GPT-Image-2.5 appears on the AA boards at all and where it lands relative to GPT Image 2 (high): if OpenAI's new model cannot beat its own predecessor in independent editing votes, the editing-leader story is moot. Until one of those two things happens, the honest summary is that the best independently-scored editing model you can't buy is dueling the best new model you can — and the scoreboard that would settle it has no GPT-Image-2.5 entry on it.