AI image, October 2026: GPT Image 2.5 arrives
October 2026 stills letter: GPT Image 2.5 Flare and Sunburst, Hy Image 3.5 Preview, Qwen-Image-2.1's research licence, Imagen 4's exit, what to run.
Dated 1 October 2026. Last month's letter argued that stills came down to three names: GPT Image 2 for obedience, Nano Banana 2 for speed, Grok Imagine Image 2.0 for region edits. Four weeks later one of those three has a successor, two labs shipped models with strings attached, and one Google model is already off Google's developer API.
This letter covers only the changes. If a model is missing, it did not show up in our launch tracking for 20 August to 24 September.
OpenAI replaced its own leader
ChatGPT Images 2.5 shipped on 8 September as two API models, GPT-Image-2.5 Flare and GPT-Image-2.5 Sunburst. OpenAI's pitch: up to 50% lower latency, 2K output, better multi-turn editing, and better subject preservation across edits. Flare is the fast default. Sunburst is the precise tier that takes longer (9to5Mac).
Both are Versely rows, with text-to-image and edit variants on the GPT Image 2.5 Flare page. The practical surprise is the credit column. Flare and Sunburst start at 4 credits, and the 4K tier is 7. GPT Image 2 starts at 9. The newer OpenAI model is the cheaper row here, so the default for fresh work flips to 2.5.
How to split the two: Flare for drafts, batches and anything where the first acceptable frame wins. Sunburst when an edit has to keep a face, a label or a layout exactly and you can wait for it. The Flare vs Sunburst breakdown has the side-by-side.
Keep GPT Image 2 where a campaign is already approved on it. Switching models mid-campaign changes skin tones, type weight and lighting in ways a client notices before you do. Start the next campaign on 2.5 instead.
Tencent previewed Hy Image 3.5, which is not on Versely
Tencent released Hy Image 3.5 Preview on 22 September. One model does generation and multi-round editing. It goes up to 4096 by 4096, takes 20 reference images, and is listed at $0.024 per image at 1K to 2K on Tencent's international API. There are no weights and no technical report (CellCog).
Hy Image 3.5 is not on Versely. The Tencent rows here are HunyuanImage 3.0 Instruct, text-to-image and edit, at 8 credits per megapixel; both live on the HunyuanImage 3.0 Instruct page. That is the closest Versely row, and it is a generation behind.
The word "Preview" is the whole story. A preview with no report and no weights can change behaviour between one week and the next. Test it on Tencent's API if the 20-reference stack matters to you. Do not build a recurring template on it yet.
Qwen-Image-2.1 ships open weights with a research-only licence
Alibaba's Qwen team released Qwen-Image-2.1 on 20 September: 7B open weights, native 2K, RGBA output with a real alpha channel, and up to 10 reference images. The catch is the licence. It is now research-only, with no commercial use (DataNorth).
That means an agency or a brand cannot ship 2.1 outputs in paid work on the open-weights terms. The licence explainer spells out what the Qwen Research License bars. If you downloaded the weights to a studio box expecting the old terms, read it before a client file leaves that machine.
Qwen-Image-2.1 is not on Versely. The Qwen rows here are Qwen Image 3 text-to-image and image-to-image at 3 credits per megapixel, with Pro at 6, both up to 2K. For the transparent-PNG job that RGBA output promises, generate on a stills row and cut the subject out afterwards.
Recraft announced V4.1 Flash with no specs
Recraft posted V4.1 Flash on 23 September and called it "the fastest image model on the market". That is all there is. No resolution, no latency figure, no price (Recraft announcements). One source, no numbers, so we treat it as an announcement rather than a spec sheet.
V4.1 Flash is not on Versely. The Recraft rows are Recraft 4.1 at 3 credits, 4.1 Utility at 3, 4.1 Pro at 17, and the vector rows at 7 and 24. For brand graphics and vectors today, standard 4.1 is already a low-credit row; speed is rarely what holds a logo job up.
Google: Imagen 4 is off the Gemini API
Google shut Imagen 4 down on the Gemini API on 17 August (Virtual Arc Software). Our Google sunsets roundup lists what else Google retired this summer and what replaced it.
If a pipeline called Imagen 4 through the Gemini API, it has been failing since mid-August, so look at the logs. For new Google-family stills, Nano Banana 2 is the row: 4 credits. Nano Banana 2 Lite is 3 credits at a fixed 1K for throwaway iterations; Nano Banana Pro starts at 12 credits and goes to 4K.
Nano Banana 2.5 is an unconfirmed arena model.
Grok Imagine Image 2.0 got its own rows
Last month's letter linked the older Grok Imagine rows. The 2.0 generation now has its own Versely row: Grok Imagine Image 2.0. It offers thirteen aspect ratios from 2:1 to 1:2, 1K or 2K output, and up to four images per request, from 2 credits. A separate 2.0 Edit row starts at 5. Four images per request suits thumbnail work: one call returns four candidates to choose between.
The job has not changed since September: region edits and composites, where you change one part of a frame and keep the rest. What changed is that you no longer have to settle for the previous generation to do it here.
Midjourney: nothing new
Midjourney shipped no new model in the window. V8.2 has been the default since 24 July, and it still is. Versely serves Midjourney V7 at 8 credits, not V8.2. If a brief depends on something only V8.2 does, run it on Midjourney's own site.
Routing stills on 1 October
| Job | Row | From |
|---|---|---|
| Fast default, most new work | GPT Image 2.5 Flare | 4 credits |
| Edit that must keep a face or label | GPT Image 2.5 Sunburst Edit | 4 credits |
| Approved campaign already on it | GPT Image 2 | 9 credits |
| Twenty throwaway options | Nano Banana 2 Lite | 3 credits |
| Google look, iterate and edit | Nano Banana 2 | 4 credits |
| Change one region, keep the rest | Grok Imagine Image 2.0 Edit | 5 credits |
| Brand graphic or vector | Recraft 4.1 / 4.1 vector | 3 / 7 credits |
| Chinese or English type in the frame | Qwen Image 3 | 3 credits per megapixel |
| Tencent instruction-following | HunyuanImage 3.0 Instruct | 8 credits per megapixel |
Credits scale with resolution tier where a row has tiers. Pick the tier the placement needs. A 4K still for a 1080-pixel feed tile is paying for pixels nobody sees.
Decisions for the month
Move new work to GPT Image 2.5. It is the newer model and the cheaper row. Keep GPT Image 2 only for work already signed off on it.
Check the licence before you check the output. Qwen-Image-2.1 is research-only. Hy Image 3.5 is a preview with no report. Neither is a safe base for client deliverables this month.
Clear Imagen 4 out of templates. Rebuild them on Nano Banana 2.
Re-test your longest edit chain. OpenAI's claims for 2.5 are about multi-turn editing and keeping the subject intact. That is where a stills workflow breaks quietly: the fourth edit drifts the face or softens the label. Run a real chain on Sunburst Edit and judge the last frame, not the first.
Ignore arena rumours in planning. An unconfirmed model is not a line item.
The November letter depends on whether Recraft publishes numbers, whether Tencent drops the "Preview" label, and whether Qwen revisits the 2.1 licence. Until then, the rows above are the ones to plan on, in the web app or the mobile app.
FAQ
Is GPT Image 2.5 better than GPT Image 2?
OpenAI claims lower latency, 2K output and better multi-turn editing and subject preservation. On Versely it is also the cheaper row. Run your own brief on both before you move an approved campaign.
Can I use Qwen-Image-2.1 commercially?
Not under the open-weights terms. The 2.1 licence is research-only with no commercial use. For commercial stills from the Qwen family, the Versely rows are Qwen Image 3.
Where do I go now that Imagen 4 is gone?
Nano Banana 2 for general Google-family stills, Nano Banana 2 Lite for cheap iteration, Nano Banana Pro when a 4K Google still is the job.