State of AI image, September 2026

September 2026 stills are a three-name problem. GPT Image 2 launched 21 April 2026 and sits first on text-to-image and editing arenas. Nano Banana 2 launched 26 February 2026. Grok Imagine Image 2.0 arrived 7–8 August 2026, second on both arenas, with region-level editing, multi-image compositing, templates, and Quality Mode as the default in Grok (kingy.ai; orcarouter; invideo). This is a dated letter, not a tour of every image row in the catalog.

The video letter next door is about audio-native models and a Sora clock. This one is about which still you make before that video, and which still you ship as a still. Mixing those jobs is how a team uses a 4K image model to feed a 15-second generate and then wonders why the pack shot looks like a painting.

Three launches, three jobs

GPT Image 2 (OpenAI, 21 April 2026). Arena leader on generation and on edits in the sources above. The Versely row to link is GPT Image 2 text-to-image. Use it when prompt adherence and a "this is the object I named" still is the job: product-on-table, a specific composition, type-in-image when it actually holds, an edit that must follow instructions. It is not automatically the right first frame for a UGC bathroom. It will happily make a catalogue.

Nano Banana 2 (Google, 26 February 2026). Fast Google stills. The Versely row is Nano Banana 2 at a low credit debit relative to cinematic video rows. The Omni Flash pipeline write-up pairs Nano Banana 2 Lite (gemini-3.1-flash-lite-image, ~4 seconds, a third-party per-image USD figure in The Next Web's June 2026 piece) with Omni Flash video editing (The Next Web). On this site, link the published Nano Banana 2 row rather than a Lite slug that is not in the catalog file. Job: iteration, thumbnails, the still you will throw away, the still you will feed to video. Not always the hero pack shot.

Grok Imagine Image 2.0 (xAI, 7–8 August 2026). Second on the arenas, region-level editing, multi-image compositing, templates, Quality Mode default in Grok. Versely rows include Grok Imagine Image Quality and Grok Imagine. Job: region edits (change this part, leave the rest), composites from several stills, a look that is not trying to be a studio pack shot. Quality Mode is a Grok-app default in that reporting; in a multi-model studio you pick the Quality row when you mean it, not because a consumer app did.

Flux 2, Seedream, Ideogram, Recraft, Qwen Image Edit, Midjourney v7, Imagen 4 — they are in the catalog. They are not this letter's three. A September state-of that named twenty still models would be a sitemap. The three above are the ones whose dates define the last six months of stills news.

Arena rank is a public scoreboard, not a brief. First on an arena can still lose a SKU label. Second with region edit can win a "change the lid, keep the bottle" job. Fast-and-cheap can win the twenty-thumbnail morning. If your meeting is "who is number one," you are not in production yet.

Stills as video fuel, stills as stills

The Omni Flash path made a still-then-edit loop respectable: generate or shoot a frame, then converse the video. Seedance 2.0 takes up to 12 references. Veo has reference-to-video and first-last-frame rows. That is the September-normal fuel job. Requirements:

The still must be a frame you would accept on screen for a beat, not a mood board.
Product labels must be real photography if the video will be an ad or a listing. Generated labels warp in stills and then warp harder in motion.
Region edits (Grok Imagine 2.0's selling point in the August coverage) are how you avoid regenerating a whole pack shot to move a reflection.

The still as still job is different: main images, A+ modules, thumbnails, ads that never become video, carousel posts. GPT Image 2 for adherence, Ideogram-class rows when on-image text is the point, Nano Banana 2 when you need ten options before lunch. Do not run a video model to get a poster. Do not run a poster model and then be surprised when image-to-video invents motion the still never implied.

Credits: GPT Image 2 text-to-image and Nano Banana 2 sit at a few credits per still on their catalog rows; Grok Imagine Quality is in the same small band; video rows are an order of magnitude more. A morning of still iteration is a cheap morning. A morning of iterating in video because nobody picked a still model is an expensive morning. This page will not print USD.

Thumbnails and text-on-image remain the honesty test. Arena wins do not mean a title-safe YouTube thumbnail. Test the actual letters at the actual size. If the letters fail, use a text overlay in the editor rather than asking the image model to typeset your brand.

What did not happen, and what not to say

No, GPT Image 2 is not a video model. No, Astra is not an image model — GPT-6 Astra (3 September 2026) is computer use, not pixels. No, Grok Imagine Image 2.0 did not make region edit unique forever; it is the reason Grok is in this letter in August. No, Nano Banana 2 did not "replace" photography.

Sora's shutdown is a video event. It does not retire DALL-E-era muscle memory, and it does not make OpenAI's stills row disappear. Keep GPT Image 2 even if you are migrating Sora video jobs to Seedance and Veo.

Watermarks and C2PA on stills are the same disclosure story as video when the still is a realistic person in an ad. Article 50 has applied since 2 August 2026. A catalogue still of a generated face used as "a customer" is a UGC-shaped problem in a JPEG. Label it. Do not assume a still is exempt because it does not talk.

Midjourney remaining in Meta Vibes' early model mix is a distribution fact for generated video feeds, not a stills ranking. If you make stills that will be remixed in an AI-only feed, stylised and character-forward will travel. Pack shots will not. That is a channel choice, not an arena choice.

This letter lists only releases with a dated source, and there is no September 2026 "GPT Image 3" or "Nano Banana 3" among them.

A stills brief you can reuse until the next letter

Write the job first:

Hero still that must match a SKU. Photograph it, or generate only the scene around a cut-out you own. GPT Image 2 if you are generating the scene and need obedience.
Twenty options, one will live. Nano Banana 2. Kill hard.
Edit one region. Grok Imagine Image 2.0 Quality row. Do not regenerate the bottle to move a shadow.
Fuel for video. Still must survive motion. No tiny text you expect the video model to keep. No warped logo.
Type-led still. Do not trust any of the three blindly. Overlay.

Then pick the row in the catalog, not the arena tweet. The AI video generator is the wrong door if you have not made the still yet; start on the image row, then move.

September 2026, dated: GPT Image 2 (21 Apr) for obedience, Nano Banana 2 (26 Feb) for speed, Grok Imagine Image 2.0 (7–8 Aug) for region edit and composites. Everything else is a specialist row or an older habit.

Sign up if those three rows should sit next to the video rows instead of in three vendor tabs. The next letter will only be worth writing if one of the dates changes.

FAQ

Is Grok Imagine Image 2.0 "better than" GPT Image 2 because it is newer?

Newer is August versus April. Arena reporting still had GPT Image 2 first and Grok second. Better is the job: region edit and composites versus obedience on a named object. Run both on the same brief if the still matters.

Should every Omni Flash clip start in Nano Banana 2?

That is the pipeline Google-adjacent write-ups recommend (Nano Banana 2 Lite + Omni Flash). Start in a fast stills row, then edit video. If you already have photography, skip the generate and go to the edit. Do not generate a fake product to feed Omni.

Do I need a new stills vendor because Sora is ending?

No. Sora is video. GPT Image 2 is the OpenAI stills line in this letter. Migrate video jobs; keep the stills row that fits the stills job.