# ChatGPT Images 2.6: GIFs, reference images and two new API models

By Ivo Donker · 5 November 2026 — compiled with AI assistance (DeepSeek V4.1 Flash)
On 4 November 2026 OpenAI published the blog post “Introducing Images 2.6”, with the subtitle “Generate images faster and with more control over editing rounds”. Images 2.6 is coming to ChatGPT, Codex and the API. You can bring images to life with GIFs, for example by animating a brand mascot or character, or by turning a product shot into a loop. You can also refine visuals with greater precision and keep characters, products and brand details consistent as you edit.

For creators, the main change is in the daily workflow in ChatGPT: GIFs are added, and you can save and reuse reference images. For developers, there are two API models: GPT-Image-2.6 Flare and GPT-Image-2.6 Sunburst. According to OpenAI, people create 4 billion images per week across ChatGPT, Codex and the API, and on average more than half a billion images per day. This article explains what Images 2.6 adds in practice, how Flare and Sunburst differ and how to choose without assuming a price difference. Other AI tools for image generation are listed in the [directory of AI tools for image generation](https://directory.llmnet.nl/en/ai-tools-beeldgeneratie), which is useful if you want to compare Images 2.6 with existing options.

## What ChatGPT Images 2.6 is

ChatGPT Images 2.6 is the new version of OpenAI's image generation in ChatGPT, Codex and the API. The emphasis is on two practical points: faster generation and more control over successive edits. OpenAI states the speed difference directly: “Images 2.6 generates images 50% faster than before.” That is a speed claim from OpenAI; the sources do not mention additional benchmark data.

The second line is consistency. If you work on the same image across several rounds, you do not want a character, product or brand detail to change with every edit. According to OpenAI, Images 2.6 is designed to hold on to those details better as you continue editing. That makes the feature relevant for work where a series of images needs to fit together, such as campaign material, product visuals or characters that return in different expressions. The promise is not that every image automatically remains identical, but that you have more control over the editing rounds.

Images 2.6 is not a separate product outside ChatGPT. OpenAI explicitly names ChatGPT, Codex and the API as the places where the new version becomes available. That means the release affects both creators working in the chat environment and teams that include image generation in their own application or workflow. The API gets separate model names and settings, while ChatGPT gets the user-facing features including GIFs and reference images.

## New in ChatGPT: GIFs and reference images

GIFs are the most visible addition in ChatGPT. OpenArt calls them “the native language of social media” in OpenAI's announcement, but apart from that positioning, the examples are concrete: marketers and creators can animate a brand mascot for a campaign, create reactions and memes for social media, or use animated storyboards to explore character movement. According to OpenAI, even a simple sketch can become an animated character. That opens a different workflow than generating still images only: you can first set down an idea or sketch and then see how motion turns out.

“GIFs are the native language of social media, and GPT-Image-2.6 lets every OpenArt creator speak it fluently from a single prompt.”

— Coco Mao, CEO & co-founder OpenArt

In plain terms, Mao says that GIFs already belong to social media and that, in her view, Images 2.6 enables OpenArt creators to make that motion from a single prompt. That is a partner quote, not an independent benchmark; it mainly says something about how OpenAI positions the feature. Town, a startup building an AI work assistant, used Images 2.6 in the API to explore animated characters. That shows OpenAI is presenting GIF-like output not only as something that plays in ChatGPT, but also as something that can be explored in API workflows.

Reference images get a separate place in ChatGPT. According to OpenAI's X thread of 4 November 2026, you can save reference images in ChatGPT and reuse them, so you do not need to upload them again each time. That may look like a small change, but it does change how you set up a series of images. If you work with a fixed mascot, a product shot or a brand style, you can keep that reference and use it again in later conversations. You then do not have to find and upload the same file again with every iteration.

Images 2.6 is also in Codex. For background on that environment, you can follow the [news about Codex](https://apps.llmnet.nl/en/codex-nieuws), where Images 2.6 is also available. The combination of ChatGPT and Codex means the same image generation appears in different forms of work: in a chat conversation, in a coding or development context and through the API. The exact interface differences between those environments are not elaborated in the sources; the availability is explicitly mentioned.

## New in the API: GPT-Image-2.6 Flare and Sunburst

In the API you choose between two models: GPT-Image-2.6 Flare and GPT-Image-2.6 Sunburst. The corresponding model ids are gpt-image-2.6-flare and gpt-image-2.6-sunburst. OpenAI does not position them as two completely different products, but as two models within the same Images 2.6 family. Both are available in the API. The difference is in what they focus on: Flare on speed and layout quality, Sunburst on higher image quality and consistency across edits.

When quality is set to low, Flare can generate images in as little as six seconds. That is a fast mode for situations where you want to do many iterations or first determine the direction. Flare also improves layout quality for posters, product imagery and other professional visuals. That makes the model interesting for work where composition and placement of elements matter, such as a poster or a product visual in which text, image and proportions need to be right. The six-second speed claim applies specifically at quality low; it is not a general promise for every setting.

Sunburst focuses on higher-quality image generation and even greater image consistency across edits when quality is set to medium or high. If you work on the same character, product or brand image across several rounds, that can help. The model is therefore not only a quality level, but also a choice for the kind of work where successive edits need to keep fitting together. OpenAI explicitly names quality medium or high for Sunburst in combination with consistency across edits.

If you want to compare Flare and Sunburst with other image generation models, you can use the [overview of image generation models](https://hub.llmnet.nl/en/beeldgeneratie-modellen) to compare the models and their positioning. Pay attention to the quality setting: OpenAI links the differences between Flare and Sunburst to quality low and quality medium or high respectively. Without that context, a comparison would be too rough.

## Flare or Sunburst: a selection guide

The choice between Flare and Sunburst starts with your workflow, not with price. Flare and Sunburst have the same standard rates. You therefore do not need to assume a price difference when choosing a model. First look at the kind of image and the number of editing rounds. If you work with many fast iterations where quality low is acceptable and layout quality for posters, product imagery or other professional visuals is central, Flare is the obvious choice. If you work with quality medium or high and want characters, products or brand details to remain consistent across multiple edits, Sunburst is the logical candidate.

Model | 
Model id | 
Emphasis according to OpenAI | 
Quality setting | 
Rates | 

GPT-Image-2.6 Flare | 
gpt-image-2.6-flare | 
Layout quality for posters, product imagery and other professional visuals; in as little as six seconds at quality low | 
quality low: in as little as six seconds | 
Same standard rates as Sunburst | 

GPT-Image-2.6 Sunburst | 
gpt-image-2.6-sunburst | 
Higher image quality and even greater image consistency across edits | 
quality medium or high: greater consistency across edits | 
Same standard rates as Flare | 

A practical rule of thumb is therefore: use Flare when speed and layout at lower quality are your most important constraints, and use Sunburst when you need consistency across edits at medium or high quality. That is not an absolute rule for every conceivable image, but it does follow the way OpenAI positions the two models in the announcement. Test with your own images and your own prompts, especially if brand or product details matter.

The selection guide also depends on what you want to automate in the API. If you are building a pipeline in which a user first makes a rough composition and then makes only detail changes, a fast first step with Flare can be useful and a later, more consistent step with Sunburst can follow. If you need a carefully built final image directly, you can choose Sunburst from the start. In both cases, quality and complexity remain separate settings; the model choice does not replace those settings.

## Markup, SVG and the complexity parameter

In addition to the model choice, the API gets more control over targeted edits. Image markup, already available in ChatGPT, is coming to the API. With it, teams can pinpoint exactly where a specific change should be made. That is useful when you do not want to regenerate the whole image but want to adjust one part: a detail in a product shot, an element in a poster or a part of a character. Markup helps keep the edit targeted instead of making a completely new variant every time.

SVG exports make designs ready for websites, apps and print. That is a different form of usability than delivering only an image file. If you want to use a design in a web page, app or print work, an SVG export can be a step in the workflow. The sources do not say that every output is automatically available as SVG; they name SVG exports as a possibility within the API controls. You will therefore need to determine in your own implementation when you use that export and how you follow it up.

The new complexity parameter has low, medium and high settings and adjusts how much planning the model does. Lower settings suit quick edits. Image quality controls visual detail and fidelity; complexity guides how much goes into the scene. Those are two different controls. You can therefore want an image with a lot of detail but edit it with little planning, or have a simple scene that requires more planning because it contains many elements or constraints.

OpenAI recommends low complexity for most editing tasks, paired with the quality your workflow requires. For scenes with many elements or constraints, you can try medium or high complexity. That recommendation is practical: do not start on the highest setting by default if you only want to change a small detail. Use low complexity as a quick edit, and switch to medium or high when the scene contains many elements or when you notice that the planning falls short. Keep the quality setting separate: quality is about visual detail and fidelity, complexity about the amount of planning.

In practice, this means you can build a workflow in layers. First determine whether you need a GIF or a still image. Then choose the model: Flare for speed and layout at low quality, Sunburst for consistency at medium or high quality. Next, set quality and complexity based on the task. With image markup you pass on targeted changes, and with SVG export you can prepare a design for web, app or print. That order prevents you from changing every setting at once and then no longer knowing which adjustment had which effect.

## What it costs

The rates below come from the OpenAI API pricing page. Flare and Sunburst have the same standard rates; there is therefore no separate price difference between the two models. Always consult the [OpenAI API pricing page](https://developers.openai.com/api/docs/pricing) for current rates, because the pricing page is the source for the amounts and the X posts do not mention them.

- Text input: $5 per 1 million tokens.

- Image input: $8 per 1 million tokens.

- Image output: $30 per 1 million tokens.

- Cached text input: $1.25 per 1 million tokens.

- Cached image input: $2 per 1 million tokens.

Because Flare and Sunburst have the same standard rates, the model choice is primarily a choice based on quality, consistency and speed, not on a lower or higher token rate. That does not change the fact that costs in an API integration can add up through your own use. Keep an eye on your usage and your settings with a view to [rate limits and costs in your API integration](https://api.llmnet.nl/en/rate-limits-en-kosten). That page helps you understand which aspects to monitor when you include image generation in a product or workflow.

For earlier price developments of model APIs, you can look at the overview of [earlier price changes of model APIs](https://radar.llmnet.nl/en/model-api-prijswijzigingen-wie-verlaagde-wat). That is useful as background, but do not turn the rates from the OpenAI pricing page into your own calculation example with numbers of tokens per image. OpenAI does not provide an example of the number of tokens an image costs, so such a calculation would not belong here. Limit yourself to the published rates per 1 million tokens and to monitoring your own usage.

A practical approach is to make agreements in advance about which quality setting and complexity you use by default, rather than measuring costs only afterwards. If you work with Flare at quality low for fast iterations, keep track of how often you regenerate. If you work with Sunburst at medium or high for consistent edits, keep track of how many editing rounds you need. The rates are the same, but your usage is not necessarily. The pricing page and your own records are leading in that respect.

## Availability and what is not known

“Images 2.6 is available globally today in ChatGPT and Codex across all subscription tiers on desktop, mobile, and web.”

In plain terms: Images 2.6 is available worldwide in ChatGPT and Codex, across all subscription tiers, on desktop, mobile and web. GPT-Image-2.6 Sunburst and Flare are available in the API. OpenAI also published a system card with safety evaluations and safeguards for Images 2.6. For background on subscriptions and API credit, there is the earlier [news about ChatGPT subscriptions and API credit](https://nieuws.llmnet.nl/en/chatgpt-pro-business-api-tegoed).

What is not known: OpenAI does not publish further subscription-specific details in these sources. There are also no different token prices for Flare and Sunburst: the standard rates are the same. The motion feature in this announcement is GIF; the sources do not mention any other motion or video export than GIFs. More specific details about separate conditions per subscription are not in the sources and should therefore not be assumed.

Availability across all subscription tiers is stated separately from API access. If you build an API integration, check your own environment and the current pricing and documentation pages before planning volume. The system card and the pricing page are the places to look for the information that OpenAI itself publishes.

## What this means for creators and developers

For creators, the biggest change is that the workflow in ChatGPT looks more like a sequence than like isolated generations. You can save a reference image and reuse it, so you do not start over with every prompt. You can animate a character or mascot as a GIF and then return to still visuals for other expressions. And you can work on details more precisely with image markup and the quality settings. That makes it sensible to document your prompts and references. If an image works, record which reference, which quality and which complexity you used. For reusable prompts, you can turn to [prompts for image generation that you can reuse](https://community.llmnet.nl/en/prompts-voor-beeldgeneratie).

With brand images, consistency is often more important than one isolated image. A mascot that looks different in a GIF than on a poster, or a product shot in which a detail shifts with every edit, costs time to repair. Images 2.6 focuses on that problem by keeping characters, products and brand details more consistent during editing. In practice, that does not yet mean you can publish every output without checking. You will still need to verify whether colours, proportions and details are right. The difference is that you can build that check into a tighter workflow: first record a reference, then edit in small steps and export only when the image is right.

For developers, the most important question is not only which model looks better, but which model fits the step in the pipeline. Flare is attractive when you want to iterate quickly at quality low and layout quality for professional visuals matters. Sunburst is attractive when you need consistency across edits at quality medium or high. Because the standard rates are the same, you can make the first choice based on output and workflow. After that, you can measure how your own usage develops and whether the chosen settings are sustainable.

The complexity parameter requires a deliberate default setting. OpenAI recommends low complexity for most editing tasks, paired with the quality your workflow requires. For scenes with many elements or constraints, you can try medium or high. That does not mean high is always better: more planning can be useful for a complex scene, but is not needed for a simple detail change. Therefore build a default path with low complexity and an exception path for complex scenes. Record in your documentation when you deviate from the default path, so that a colleague can reproduce the same choice.

Monitoring costs is part of the design, not only of financial administration. Because Flare and Sunburst have the same standard rates, the difference lies in your usage: how often you generate, which quality you choose and how many editing rounds you need. Measure that in your own integration and compare it with the rates on the OpenAI pricing page. If you want to explore an alternative route, you can also look at [running image generation locally as an alternative](https://gids.llmnet.nl/en/lokale-beeldgeneratie). That is a different trade-off than the API, but it can help determine which images you want to do in the API and which you do not.

The arrival of GIFs also means you need to think in advance about the final format. A GIF is suitable for short movements and social contexts, but requires different checks than a still image. Check whether the motion is right, whether the character remains recognisable and whether the loop does not visually disturb. Use an animated sketch or storyboard as an intermediate step when you are not yet sure how a character should move. That way you avoid having to redo a complete animation as soon as one detail is wrong.

## Practical checklist

- First determine whether you need a still image or a GIF; the motion feature in this release is GIF.

- Save a reference image in ChatGPT and reuse it, so you do not upload it again every round.

- In the API, choose gpt-image-2.6-flare when speed at quality low and layout quality are the priority.

- Choose gpt-image-2.6-sunburst when you need consistency across edits at quality medium or high.

- Use image markup for targeted changes and SVG export when a design needs to go to web, app or print.

- Start with low complexity for most editing tasks and try medium or high for scenes with many elements or constraints.

- Keep quality and complexity separate: quality is about visual detail and fidelity, complexity about how much planning the model does.

- Check the rates on the OpenAI API pricing page and monitor your own usage; Flare and Sunburst have the same standard rates.

ChatGPT Images 2.6 mainly adds practical control to a workflow that creators already know. GIFs make it possible to add motion to a mascot, character or product image, while reference images in ChatGPT reduce the need to repeat the same upload. In the API, Flare and Sunburst provide two models focused on different quality settings, with the same standard rates. For developers, the choice is therefore primarily a workflow choice: fast iteration and layout, or consistency at medium or high quality.

The most sustainable approach is to start small. Test one campaign image or one character, record your references and settings and see how the edits relate to your expectations. Switch to a heavier setting only when the scene calls for it. Keep an eye on availability and pricing through the sources that OpenAI itself publishes, and do not assume terms or rates that are not listed there. That keeps the move to Images 2.6 manageable, both in ChatGPT and in the API.

Checked on 2026-11-05.

This article was compiled with AI assistance and editorially reviewed; sources are listed below.

Read also: [OpenAI pipeline found 948 Linux root bugs in one run](https://nieuws.llmnet.nl/en/openai-cyber-pipeline-948-linux)

## Sources

- OpenAI blog “Introducing Images 2.6”, with the announcement, GIFs, API models and pricing context: https://openai.com/index/introducing-images-2-6/

- OpenAI on X, thread of 4 November 2026 about, among other things, reference images in ChatGPT: https://x.com/OpenAI/status/2118124944928432327

- OpenAI on X, additional thread about Images 2.6: https://x.com/OpenAI/status/2118124966055223755

- OpenAI Developers on X about the API release: https://x.com/OpenAIDevs/status/2118124753340756055

- OpenAI API pricing page with the standard rates for Flare and Sunburst: https://developers.openai.com/api/docs/pricing