OpenAI Updates Images 2.6: ChatGPT Now Supports One-Sentence GIF Generation, API Opens SVG Export

On the morning of November 5, many people discovered that ChatGPT had quietly gained a new capability.

By typing a sentence or uploading a photo, users can now have it generate a looping GIF. After generation, they can also adjust the playback speed, view each frame, and even modify a single frame individually.

We gave it a quick try, and some results were quite good, such as those leaning toward animals or animation.

Image | Cat jumping, bear shivering, fur

But some generations were rather rough. For example, our prompt was to generate a GIF of Musk doing a flash-step, and this animated image was neither "Musk" nor a "flash-step."

I believe everyone still remembers vividly that when Images 2.0 first arrived earlier this year, many people lamented that the era of "seeing is believing" had truly passed.

"Image generation" has since remained one of OpenAI's strongest capabilities, but this change is somewhat different.

In the past, AI image generation tools solved the problem of going from nothing to something. Users provide a piece of text, and the model turns it into an image. But for an image to truly enter a workflow, it often still has to go through editing, layout, format conversion, and other steps. For ordinary users, these operations mean opening a few more pieces of software; for designers and content teams, they mean a not-insignificant amount of post-production work.

OpenAI's Images 2.6 update this time has begun to bring these tasks inside the model.

In addition to GIF generation, the new model also brings faster image generation speeds, more stable multi-turn editing, and SVG vector export and image annotation editing interfaces for developers. And OpenAI has not raised prices.

But when model vendors begin to fill in these steps themselves, the existing businesses are no longer so secure.

For OpenAI, this may just be an image product upgrade; for a group of AI tool companies that survive on single-point features, the change may only be beginning.

One upgrade with more capability at no extra cost: the work of three types of tools packed into the model

Images 2.6 is not a simple speed-up version.

From a product design perspective, OpenAI did two things this time: first, make image generation faster; second, make post-generation modification and use more convenient.

First, speed.

According to figures released by OpenAI, Images 2.6 is 50% faster overall than the previous generation. On the API side, the company launched two versions: GPT-Image-2.6 Flare and Sunburst.

Among them, Flare emphasizes response speed and efficiency, suitable for tasks requiring rapid iteration such as posters and product images; Sunburst focuses more on detail and consistency in multi-turn editing, suitable for work with higher requirements for image precision.

Under low-quality settings, Flare can generate an image in as fast as about 6 seconds. For content producers who need to repeatedly adjust materials, faster generation means more rounds of attempts in the same amount of time.

Here's a fun fact: Kuaishou was originally a GIF animation creation tool, and its product name was "GIF Kuaishou." It was only later that it grew from a tool software into one of China's largest short video platforms, undergoing an important product transformation.

I bring this up to say that the demand for GIF animations is very clear.

In ChatGPT, users can now directly generate animated images through text descriptions or by uploading images. For example, making a photo of a person move, or turning a static product image into looping display material. After generation, users can adjust the playback speed and also view and modify individual frames.

This capability has practical significance for social media operations, advertising and marketing, and content creators. In the past, turning a static image into a GIF usually required steps such as image processing, frame-by-frame production, and animation compositing. For ordinary users who use it occasionally, this process is not worth learning specifically; for teams with large material needs, it means extra time and labor costs.

OpenAI mentioned in its official blog that AI office assistant company Town has already used the new model to create typing-wait animations for virtual characters in its product.

Another change occurs in the output of design files.

Images 2.6 has begun to support SVG format export, though this capability is currently mainly aimed at API developers. SVG differs from common PNG and JPEG formats in that it stores vector paths, so it does not become pixelated when enlarged, and it is also convenient for further adjusting shapes, colors, and details.

This means that what the model generates is no longer just images that can be viewed, but also design materials that can enter web pages, applications, and printing workflows.

For designers, this point is important.

Local editing is also advancing in this direction.

ChatGPT had previously provided an image annotation feature, allowing users to circle the areas that need adjustment on an image and then describe the modification requirements through text. In this update, OpenAI has further opened this capability to API developers.

For developers, this means they can embed image modification into their own products without having to rely entirely on users re-entering prompts or specifying modification areas through complex mask operations.

In addition, OpenAI has added a complexity parameter, allowing developers to control how much planning computation the model invests before generation. Simple tasks can reduce planning time and prioritize speed; complex compositions can invest more computing resources.

This is somewhat similar to the idea of adjusting reasoning effort in text models. Instead of offering only a single fixed generation mode, the model now allows developers to decide how resources are allocated based on the complexity of the task.

On pricing, the image output price for both Flare and Sunburst is $30 per million tokens, with image input at $8 and text input at $5, consistent with the corresponding previous-generation models.

When capabilities become interfaces, the position of middlemen changes

Some of them do AI portraits, some do product images, and some focus on posters, avatars, and social media assets. Their product forms differ, but their business logic is similar: call the underlying model at the bottom layer, provide an interactive interface and specific features on top, and then charge through subscriptions, credits, or enterprise services.

On the one hand, the barrier to using foundation models is relatively high; ordinary users do not know how to write prompts, nor do they understand how to control the image. On the other hand, the models themselves can often only complete a single task, and users still need other software to process the generated results.

Startups happen to fill this gap.

They may not have their own models, but through product design they can turn complex technical capabilities into services that ordinary people can use. Users pay for convenience, and startups earn commercial revenue from it.

Meitu, as a product of the internet era, has come back to life in the AI era, and this is exactly the logic: it does not build its own foundation models, but by building the connection between models and users it has made a fortune. In the first half of 2026, its revenue increased by 22.1%, and net profit surged by 39.5%.

Now the problem is here: top models such as OpenAI's have begun to build functions that were previously handled by the application layer directly into the model, and the competitive advantage of startups will be directly affected.

Take AI posters as an example. In the past, a product might offer dozens of templates; users choose a style, enter text, and then the model generates an image. Now, if the underlying model can already understand poster layout, accurately handle text, and directly complete local modifications, then products that merely provide templates and prompts will find it hard to maintain their original differentiation.

If users only need to change a background, adjust a character's clothing, or modify a detail in the image, doing it directly in ChatGPT may be more convenient than opening a standalone app. Especially when these features are already included in the user's existing subscription, third-party tools are put in an awkward position: why would users pay again for the same capability?

From this perspective, OpenAI's ambition is also enormous: it controls the underlying technology and also owns a traffic gateway like ChatGPT, enabling it to push new features directly to a huge user base.

Not just Meitu—Canva and Adobe are also affected. What they need to watch out for is the change in users' demand for tools.

Adobe Firefly currently does not compete purely on model image quality, but emphasizes safety for commercial use. According to Adobe's previously announced enterprise policy, early Firefly models were mainly trained on Adobe Stock, openly licensed content, and public domain content, and some enterprise customers can also obtain contractual intellectual property indemnification.

For brands and enterprise customers, commercial design must consider copyright risk, source of materials, and ultimate responsibility for use. This is also where Adobe is currently relatively solid.

There are also some application providers that have proactively integrated the new model into their products.

Figma Weave and Picsart have become early partners for Images 2.6. OpenAI disclosed in its release information that Figma Weave will use the new model for reusable creative workflows, while Picsart hopes to use faster generation and more stable editing capabilities to make it easier for users to create serialized visual content.

Ultimately, the stronger the underlying model, the more the application layer needs to re-prove its value.

If a tool company offers only functions that the model itself already has, it will be difficult for it to maintain a premium over the long term; but if it can integrate the model into complex business processes and accumulate user data, industry experience, and collaborative relationships, then improvements in model capabilities may instead become its own growth driver.

There is much more business OpenAI wants

In September this year, when OpenAI launched Images 2.5, it had already begun emphasizing more precise editing, reference images, and faster generation speed. According to figures released at the time, users generated more than 3 billion images per week through ChatGPT and the API.

Just two months later, that number rose by another 1 billion. For a company that started with a conversational product, this means images are becoming a fairly sizable use case within ChatGPT.

In particular, if image generation enters more specific workflows, the commercial value may become different. What model vendors are fighting for is no longer just a single API call, but which platform users use throughout the entire work process.

OpenAI's strategic adjustments in recent months also provide another angle for observing this image update.

In March this year, OpenAI announced the discontinuation of the standalone Sora product. Reuters reported at the time that the decision was related to the company reallocating resources, as OpenAI began to focus more attention on areas with clearer commercial prospects, such as coding tools and enterprise customers. Sora-related products subsequently stopped service in April, and the API was also shut down in September.

This once surprised the outside world. Sora is one of OpenAI's most closely watched products in the video generation field; it once sparked a great deal of discussion and was at one point seen as an important gateway for generative AI to enter the content production industry.

But judging from subsequent moves, OpenAI is adjusting its assessment of product value.

Compared with a standalone video generation app, capabilities that can be embedded in ChatGPT, Codex, and enterprise workflows may be more likely to generate sustained usage and commercial revenue.

The product design of Images 2.6 also reflects this trade-off to some extent.

It did not focus on building a brand-new standalone creation community, nor did it try to create another dedicated GIF app; instead, it put GIF generation directly into ChatGPT and opened up capabilities such as SVG and markup to API developers. Users can accomplish more things within the original product, and developers can also integrate these capabilities into their own applications.

This gives OpenAI opportunities in two directions at once: increasing ChatGPT usage scenarios on the consumer side, and on the business side entering more professional software and business processes through the API.

Among these, the significance of the API is especially worth noting.

If ChatGPT directly faces users, then the API faces companies that already have users, businesses, and scenarios. OpenAI does not need to personally develop every vertical tool; by providing underlying models, it can also let other companies embed its capabilities into their own products.

For domestic entrepreneurs, some opportunities may already be quietly slipping away.

But as model capabilities continue to move downward, relying solely on prompt optimization, interface design, and simple feature combinations will make it hard to build a long-term moat.

The truly valuable part may lie outside the model.

For example, can an e-commerce image tool understand the product characteristics of different categories, ensure that product details are not distorted, and connect to merchants' product databases and advertising delivery systems? Can an enterprise design tool manage brand assets, unify visual standards, and integrate with internal approval and collaboration processes?

For tool-type startups, the question they need to answer anew is this: when underlying models can already accomplish more and more things, what exactly do they still control?

Even more, as the capabilities of top-tier models grow stronger and stronger, should humans think about what is left for themselves?

This article comes from the WeChat public account "Phoenix Technology"; author: Phoenix Technology; published by 36Kr with authorization.