What changed in AI image tools this quarter

admin
What changed in AI image tools this quarter cover graphic

AI image tools had a quieter quarter than assistants or coding tools, at least in the sources we could verify. Between August and early November 2026 the clearest changes were OpenAI’s ChatGPT Images 2.5 in September, a new Gemini image model with an adjustable “thinking” setting in October, an early Midjourney experiment with the same idea, and a further OpenAI Images update in early November.

This article lists those changes with dates and sources, explains what each means for someone choosing an image tool, and is honest about what we could not confirm. We have not tried to cover every image generator, because we could only verify a handful of dated releases in the window.

How we compiled this: we did not run hands-on tests. This is based on vendor pages and public reporting, checked on 11 November 2026. Vendor claims, such as speed improvements, are the vendors’ own and have not been independently verified by us.

Key takeaways

  • OpenAI released ChatGPT Images 2.5 on 8 September 2026, adding a Sketch tool, templates, image comments and, it says, generation up to 50 percent faster than Images 2.0.
  • A secondary source reports Google deployed Nano Banana 2.1 on 6 October 2026, with a thinking level you can set before the model draws and up to 14 reference images.
  • The same source says Midjourney began testing a “Rerun (Thinking)” feature on 8 October, limited to its alpha site, and that V8.2 has been its default model since 24 July.
  • A timeline of OpenAI release notes lists Images 2.6 on 4 November with faster generation, GIF creation and API editing controls; we could not open OpenAI’s own page for it.
  • The common theme is control: editing only what you ask for, setting how much the model reasons, and iterating on a result rather than starting again.

Summary

Three things changed in practice.

  • Editing precision. OpenAI says Images 2.5 changes only what you request and stays consistent across several editing turns. That addresses a common frustration, where a small edit alters the whole picture.
  • Reasoning before or after drawing. Google’s Nano Banana 2.1 exposes a thinking level, and Midjourney is testing a feature that reruns a finished image with extra reasoning to fix missed objects, layout, anatomy and text.
  • Speed. OpenAI claims faster generation in both Images 2.5 and, per a timeline, Images 2.6.

Dates are those given by each cited source.

OpenAI: Images 2.5 and 2.6

ChatGPT Images 2.5, 8 September 2026

OpenAI’s announcement page is dated 8 September 2026. It lists the following changes:

  • Sharper details and more natural lighting and textures.
  • Generation up to 50 percent faster than Images 2.0, which is OpenAI’s own figure.
  • More precise editing that changes only what you request, and better consistency across multiple editing turns.
  • Sketch: draw a rough idea in ChatGPT, using “@Sketch”, and have it turned into a finished image.
  • Templates for formats such as posters and merchandise, image comments that let you place notes directly on an image to guide edits, and prompt sharing so others can adapt the prompt behind an image.

On availability, the page says the update is rolling out to all tiers of ChatGPT, ChatGPT Work and Codex on desktop, mobile and web. For developers it names two new API models: GPT-Image-2.5 Flare, described as the default for most applications with about 50 percent lower latency than GPT-Image-2, and GPT-Image-2.5 Sunburst, which offers greater precision with longer generation times. Pricing is listed on OpenAI’s site, and we did not record figures, so check the page.

Images 2.6, 4 November 2026

A third-party timeline of OpenAI release notes lists “Images 2.6” on 4 November 2026, adding faster image generation, GIF creation and API editing controls. We could not open OpenAI’s own release note, so we cannot say which plans get it or how the API controls work. Treat this as unconfirmed until you see it on OpenAI’s site.

What it means: if you create images inside ChatGPT, the Sketch and image comment features change how you can direct edits. If you build with the API, the two model names suggest a choice between speed and precision, but you should read OpenAI’s current documentation for limits and prices before switching.

Google: Nano Banana 2.1

An article from Atlas Cloud dated 19 October 2026 says Google deployed Nano Banana 2.1 on 6 October 2026, with a thinking level you can set before the model draws. It also reports that the Gemini API lists the model as generally available from 6 October, with 1K, 2K and 4K output, support for up to 14 reference images, and minimal, medium and high thinking levels, with medium as the default.

This is a secondary source, and we could not open Google’s own release notes. Google’s documentation is where to confirm the model identifier, pricing and which apps include it.

What it means: a thinking-level setting gives you a trade-off to test between cost, latency and how well the model follows complex prompts, such as text inside images or multi-object layouts. The sources do not tell us how large that trade-off is, so you would need to measure it with your own prompts.

Midjourney: V8.2 and Thinking

The same Atlas Cloud article reports two Midjourney facts. First, V8.2 has been Midjourney’s default model since 24 July 2026, and the version page describes it as focused on aesthetics, image quality and personalisation. It includes a new Edit Model that replaces Omni Reference, Character Reference and the Retexture tool. Those points fall before our window but explain the current state.

Second, Midjourney began testing a Thinking feature two days after Nano Banana 2.1’s release, with the article giving the announcement date as 8 October 2026 on Midjourney’s updates site. It appears as a “Rerun (Thinking)” button under the lightbox, reruns a completed V8.2 task rather than planning before the first image, and targets missed objects, layout, anatomy and text. It is limited to the alpha site for subscribers, described as an early version, and has no endpoint available. We could not open Midjourney’s updates page, so these details are second-hand.

What it means: for now this is an experiment, not a general feature. If you rely on Midjourney, watch for it to leave alpha before planning a workflow around it.

Context from just before the window

A “Q3 2026” image generator round-up from DataNorth lists several releases that fall just before the period we cover. We include them only to explain the starting point.

Date (2026) Item As reported
30 June Nano Banana 2 Lite Google released a high-volume variant of Nano Banana 2
7 July Muse Image Launched by Meta’s Superintelligence Labs, free for everyday use in Meta AI, Instagram and WhatsApp
Late July FLUX 3 Black Forest Labs unveiled a multimodal image, video and audio model, in limited early access

The same page’s newest dated item is 7 July, so it does not tell us what happened later with these products. We found no source in the window that confirms FLUX 3 general availability.

Timeline at a glance

Date (2026) Vendor Change Source type
8 September OpenAI ChatGPT Images 2.5 and two new API models Vendor announcement
6 October Google Nano Banana 2.1 with thinking levels Secondary report
8 October Midjourney Thinking rerun feature in alpha testing Secondary report
4 November OpenAI Images 2.6 listed: faster generation, GIF creation, API editing controls Third-party timeline

What it means for tool buyers

  • Test editing, not just first drafts. Both OpenAI and Midjourney emphasise changing part of an image without disturbing the rest. A fair test is to take one image and request three small edits in a row. Our guide on comparing AI tools side by side shows how to keep that fair.
  • Look at the unit of cost. The Gemini API lists output sizes and thinking levels, and OpenAI lists two API models with different speed and precision. Costs may differ by setting, so read how to read an AI tool’s pricing page before assuming a headline price applies to your use.
  • Check where the tool sits in your plan. OpenAI says Images 2.5 reaches all ChatGPT tiers. A feature in an alpha site, as Midjourney’s Thinking is, is not something to depend on.
  • Review output for errors. Text in images, hands and layout are the areas Midjourney’s own feature targets, which tells you they remain weak spots. See how to check AI output for mistakes.
  • Keep a record of your own results. Because we found few independent measurements in this window, your own side-by-side trial is the best evidence you will get. Our trial guide explains a simple approach.

What we could not confirm

We believe in saying what is missing. For this round-up we could not confirm, from sources we opened:

  • Any dated August 2026 image-tool release. The only list we opened for the window had its newest item in July.
  • Updates from Adobe Firefly, Ideogram, Recraft, Stability AI or Black Forest Labs in the window. Several vendor sites were not available to us.
  • Pricing or plan changes for any of the tools above.
  • Independent quality comparisons. The pages we read were vendor announcements or secondary reports, and a comparison page we opened gave dates but no hands-on method.

If one of these matters to you, check the vendor’s own news page directly.

What to watch next

These are questions raised by the sources, not predictions.

  • Whether Midjourney’s Thinking feature moves out of alpha and gets an endpoint.
  • Whether OpenAI’s own pages confirm the details of Images 2.6.
  • Whether Black Forest Labs’ FLUX 3 reaches general availability.
  • Whether independent evaluations of “thinking” image models appear, so claims can be compared.

Frequently asked questions

What is new in ChatGPT Images 2.5?

According to OpenAI’s 8 September 2026 page: sharper details, faster generation, more precise edits, Sketch, templates, image comments and prompt sharing, plus two API models.

What is the Nano Banana 2.1 thinking level?

A reported setting, minimal, medium or high, that controls how much the model reasons before it draws. Medium is the reported default.

Is Midjourney’s Thinking feature available to everyone?

Not according to the source we read. It was described as limited to the alpha site for subscribers.

Did Google or OpenAI change image prices?

We did not confirm any price changes. Check each vendor’s pricing page.

Which image generator is best?

We cannot say. We have not tested them, and the sources we read do not compare them independently. Run your own trial on your own prompts.

Are AI images labelled or watermarked?

We did not find a source in this research that addresses image labelling for these releases, so we make no claim. Check each vendor’s terms and policies.

Conclusion

The documented changes in image tools this quarter point the same way: more control over edits and more attention to reasoning, with speed claims on top. The evidence is thinner than for some other categories, and several details come from secondary sources. Before you switch, read how to choose an AI tool, check how to protect company data if you upload brand or client assets, and run a side-by-side trial.

Sources

Keep reading