ChatGPT Images 2.6: What's New and How to Spot It
- 6 views
- 0 likes
- 0 comments
- All languages combined

Sign in and earn +30 points by reading this
Read to the end, wait a few seconds, then tap the claim button below.
In short
ChatGPT Images 2.6, released on 4 November 2026, adds GIFs, saved references and two API models, Flare and Sunburst [1][2]. OpenAI claims 50% faster generation, but one independent test measured about 6% [1][7]. Images carry C2PA metadata and a SynthID watermark, which screenshots and compression can strip [3][5].
Key takeaways
- OpenAI released Images 2.6 on 4 November 2026 in ChatGPT, Codex and the API, across all ChatGPT tiers [1].
- New features include GIFs, saved reference images, markup editing in the API and API-only SVG output [1][2][3].
- OpenAI claims 50% faster generation; PhotoWorkout measured about 6% for Sunburst and no gain for Flare [1][7].
- OpenAI's own limits for GPT Image models include slow complex prompts, imperfect text and imprecise layouts [4].
- OpenAI says GIF motion violations remain hard to detect [3].
- C2PA metadata and SynthID watermarks help identify OpenAI images, but a negative check is not proof an image is real [3][5].
On this page 11 sections
- What is ChatGPT Images 2.6?
- What changed compared with Images 2.5?
- Is it really 50% faster? What testers found
- What are the limits of Images 2.6?
- How does OpenAI handle safety for Images 2.6?
- How can you spot an image made with ChatGPT Images 2.6?
- Should you use it, and what should you watch?
- Common mistakes
- Frequently asked questions
- Glossary
- Sources
What is ChatGPT Images 2.6?
Images 2.6 is the image generator OpenAI released on 4 November 2026 for ChatGPT, its Codex coding tool and its developer API [1]. OpenAI says it is available globally across all ChatGPT subscription tiers on desktop, mobile and web [1]; TechRadar reported that this includes free accounts [6]. In the API, it comes as two models: GPT-Image-2.6 Flare, tuned for speed, and GPT-Image-2.6 Sunburst, tuned for detail and consistency across edits [1].
OpenAI's release notes summarise the update as faster image generation that is better at keeping details consistent across edits, with GIF creation rolling out and saved reference images in ChatGPT [2]. OpenAI also says people now create 4 billion images a week across its products [1].
What changed compared with Images 2.5?
According to OpenAI, the main changes are [1][2][3][4]:
- Speed: OpenAI says images generate 50% faster than before, and that Flare can produce an image in as little as six seconds at low quality [1].
- Animated GIFs: short looping animations from a prompt or a sketch, rolling out in ChatGPT [1][2]. In the API, a GIF can have 2 to 16 frames (9 by default) at 20 to 2,000 milliseconds per frame [4].
- Consistency and references: better at keeping characters, products and brand details the same across edits, and ChatGPT can save reference images of people, pets or characters for reuse [1][2].
- Markup editing in the API: developers can mark the area to change on a separate transparent annotation layer [4]; OpenAI's system card lists photo markup in both ChatGPT and the API [3].
- SVG output: vector (SVG) images, in the API only [3][4].
- Complexity setting: a new API control, low, medium or high, for how much the model plans a scene; it has no separate fee [4].
The table below separates what OpenAI claims from what independent testers reported.
Is it really 50% faster? What testers found
The speed claim is OpenAI's, and early tests do not fully bear it out. PhotoWorkout, a photography site, generated the same 1024 by 1024 image three times on each model through a ChatGPT-login setup. Sunburst at high quality averaged 30.1 seconds on 2.6 against 31.9 seconds on 2.5, about 6% faster; Flare at low quality averaged 31.2 seconds on 2.6 and 29.3 on 2.5, nowhere near six seconds [7]. That is one small test with one setup, not a benchmark, but it shows the headline figure should be read as a best case.
Other findings from PhotoWorkout's hands-on test, which are impressions rather than measurements [7]:
- Faces and buildings in real photos held up well during edits, but the model rewrote handwriting in a notebook it had been told not to touch.
- Text the user specified rendered without errors; text the model invented, such as camera names, was made up.
- Asked to animate a real photo in ChatGPT, it pasted AI-drawn overlays onto the unchanged original using code, instead of animating the scene.
- PhotoWorkout reports that both 2.6 models cost the same as their 2.5 predecessors on OpenAI's pricing page, at $8 per million image input tokens, $30 per million image output tokens and $5 per million text input tokens; the pricing page itself was not checked here.
TechRadar's reviewer counted only two genuinely new features in ChatGPT, GIFs and persistent reference images. A first GIF of his cat was just "ok", and a follow-up prompt asking for more movement gave a much better result, a reminder that prompt wording matters [6].
What are the limits of Images 2.6?
OpenAI's developer guide lists limits that apply to its GPT Image models generally, not specifically to 2.6 [4]:
- Latency: complex prompts can take up to two minutes.
- Text: the model can still struggle with precise text placement and clarity.
- Consistency: it may still struggle to keep recurring characters or brand elements identical across generations.
- Layout: placing elements exactly in structured compositions can be hard.
There are also format limits. The image edit endpoint does not support SVG output, and an SVG is not a supported input for edits; SVG output does not support streaming, and GIF output does not support streaming or partial images [4]. Masks used for editing are guidance only: the model may not follow a mask's exact shape [4]. Higher complexity can increase latency and introduce unwanted changes, so OpenAI suggests comparing results [4].
How does OpenAI handle safety for Images 2.6?
OpenAI's system card says Images 2.6 is based on the same safety foundations as Images 2.5, with additional safeguards for GIF creation [3]. In an automated test of 3,077 deliberately challenging prompts, unsafe images that were not blocked made up 1.72% of Sunburst 2.6 outputs, against 1.92% for its 2.5 baseline, and 2.14% for Flare 2.6, against 1.49% for its baseline [3]. OpenAI marks only some differences in that table as statistically significant; the unsafe-not-blocked figures for both models are not among them [3]. These prompts were chosen to be hard, so the rates are not representative of normal use [3].
GIFs add new risks, the card says, such as sexualised movement or harmful meaning that appears only across frames. OpenAI says motion-related violations remain challenging to detect and its safeguards may not catch all of them [3].
On severe-risk testing, OpenAI says neither model crosses its High thresholds for biological or cyber capability, but it still applies the mitigations for a model treated as High in biology, as a precaution [3].
How can you spot an image made with ChatGPT Images 2.6?
OpenAI adds two kinds of provenance signal to images from Images 2.6. One is C2PA metadata, an industry standard for recording where a file came from, which OpenAI supports through the C2PA Conformance Program [3]. The other is an invisible SynthID watermark from Google DeepMind, embedded in the image itself, applied in ChatGPT, Codex and the API [3].
To check a file, OpenAI offers a Verify web tool on its website and a content provenance API [5]. Use them carefully:
- Use the original file. Editing, converting or sharing a file can remove its metadata; compression, cropping and screenshots can erase or weaken signals [5].
- Read a negative result correctly. A not_detected result means no supported signal was found, not that the image is human-made [5].
- Remember the scope. OpenAI's checker looks only for OpenAI signals; it is not a general AI detector and does not identify content from every AI system [5].
- Check the issuer. OpenAI advises checking which provider issued C2PA credentials before attributing an image [5].
- Add human judgement. For anything high-stakes, such as news photos or evidence, pair automated checks with human review [5].
Visual clues help too, such as invented text in the background, as PhotoWorkout found [7]. The guide How to Tell if an Image or Video Is AI-Generated covers those signs in detail. Images shared during protests, crises or elections deserve extra care, because a stripped file looks the same whether it is real or generated.
Should you use it, and what should you watch?
For quick social graphics, mock-ups and simple animations, Images 2.6 adds useful tools, especially GIFs and saved references [1][2]. For photo editing where every detail must stay untouched, test first: PhotoWorkout's notebook example shows the model can change things it was told to leave alone [7]. For anything that will be published as fact, do not use generated images as if they were photos.
Points to watch: whether the GIF rollout reaches all accounts [2], whether independent tests confirm the speed gain [7], and how well the provenance signals survive as images are shared across platforms [5].
| Feature | What OpenAI says | What testers reported |
|---|---|---|
| Speed | 50% faster; Flare as little as 6 s at low quality [1] | Sunburst about 6% faster; Flare slower than 2.5 in one test [7] |
| GIFs | Rolling out in ChatGPT; 2 to 16 frames in the API [2][4] | Cartoon GIFs drawn frame by frame; real photo animated with code overlays [7]; cat GIF improved with a better prompt [6] |
| Edit consistency | Better at keeping details across edits [1][2] | Faces and buildings held up; handwriting was rewritten [7] |
| Availability | All ChatGPT tiers, globally [1] | Released to all levels including Free [6] |
| API price | Not covered here; see OpenAI's pricing page | Same as 2.5 on OpenAI's pricing page [7] |
Common mistakes
Repeating '50% faster' as a measured fact
Fix: That is OpenAI's claim [1]. One independent test found about 6% for Sunburst and no gain for Flare [7].
Treating a not_detected result as proof an image is real
Fix: It only means no OpenAI signal was found; metadata may have been stripped or the watermark degraded [5].
Using OpenAI's checker on images from other AI tools
Fix: It looks only for OpenAI signals and is not a general AI detector [5].
Assuming SVG export works in ChatGPT
Fix: The system card lists SVG output as an API feature [3].
Expecting a photo to be animated like a video
Fix: In PhotoWorkout's test, ChatGPT added code-drawn overlays to a real photo rather than animating it [7].
Frequently asked questions
Is ChatGPT Images 2.6 free?
OpenAI says it is available across all ChatGPT subscription tiers [1], and TechRadar reported that this includes Free accounts [6]. OpenAI's release notes describe GIF creation as rolling out, so it may not appear on every account at once [2].
Can ChatGPT make GIFs now?
Yes. GIF creation is rolling out with Images 2.6 [2]. In the API, GIFs can have 2 to 16 frames [4].
What is the difference between Flare and Sunburst?
Flare is the faster API model; Sunburst offers higher quality and better consistency across edits at medium or high quality [1].
Do Images 2.6 pictures have a watermark?
They carry C2PA metadata and an invisible SynthID watermark [3]. Neither is a visible label, and metadata can be removed [5].
How do I check if an image came from ChatGPT?
Upload the original file to OpenAI's Verify web tool; screenshots and compressed copies may lose the signals [5].
Glossary
- C2PA
- An industry standard for attaching provenance information to media files, describing how and by whom they were made [3].
- SynthID
- Google DeepMind's invisible watermark embedded in the content itself; OpenAI adds it to Images 2.6 output [3].
- Flare and Sunburst
- The two GPT-Image-2.6 API models, one for speed and one for detail and edit consistency [1].
- Markup editing
- Marking the area to change on a separate transparent layer and describing the change in a prompt [4].
- System card
- A document in which an AI developer describes a model's safety testing, safeguards and limits [3].
Sources
- Introducing Images 2.6 OpenAI · 2026-11-04
- Release notes (entry of 4 November 2026: ChatGPT Images 2.6 and GPT Image 2.6) OpenAI · 2026-11-04
- ChatGPT Images 2.6 System Card OpenAI Deployment Safety Hub · 2026-11-04
- Image generation (API guide) OpenAI for Developers
- Content provenance (API guide) OpenAI for Developers
- ChatGPT Images 2.6 is here, and it can turn your photos into animated GIFs, so I tried it on my cat TechRadar · 2026-11-05
- OpenAI GPT-Image-2.6 Tested on Real Photos: Perfect Text, Convincing Edits, No Real Speedup PhotoWorkout · 2026-11-07
Comments
Top comments