OpenAI releases Images 2.6: 50% faster generation, GIFs, and SVG export

If you build products that generate images, you have probably encountered a familiar pain point: the quality is good, but generation is slow and expensive, while faces drift from one frame to the next. OpenAI has released Images 2.6, and the update targets precisely those issues.

What's new

The headline feature for mainstream users is GIF animation. You can change the playback speed and edit individual frames instead of regenerating everything from scratch. The second major addition is saved references. You upload an image of yourself, a pet, or a character once and then reuse that identity. According to early feedback, the model finally handles faces properly.

Another new feature is annotation directly on an image, along with SVG export. This is a major step forward for icons and simple illustrations: the vector file can go straight into a design system or website. The model also lays out posters and product images more cleanly, meaning composition and text behave more predictably.

Generation speed is up 50% compared with version 2.5.

What's in the API

As before, there are two models:

• gpt-image-2.6-flare — optimized for speed, starting at 6 seconds per image at low quality;
• gpt-image-2.6-sunburst — optimized for quality and consistency across edits.

Prices have not changed: roughly 0.5 to 5 cents per image. That matters when calculating unit economics: the upgrade is effectively free.

Pricing and benchmarks

The author of the original post, Igor Akimov, highlights an interesting detail. Most benchmarks, including Artificial Analysis, Arena, and DesignArena, compare the most expensive HQ mode. But the API also offers medium and low settings, which are almost 10 times cheaper. When his team selected a generator through blind testing, 80% chose OpenAI, and that option also proved to be the price leader. The takeaway is that models should be compared at the quality setting you actually use, not by the top rows of leaderboards.

One caveat: this is the perspective of a single author and his team, not an independent study. GPT-Image's lead in rankings should be validated against your own workloads, especially if you need specific styles or strict control over text.

What to do in practice

1. If you are using version 2.5, switch to 2.6: the price is the same, speed is higher, and migration usually comes down to changing the model name.
2. For interactive scenarios where the user is waiting for a result, choose flare. For iterative editing and characters that need to remain recognizable, choose sunburst.
3. Run your prompt set at low and medium quality, not just high. Users often will not notice the difference in the final product, while the bill can be several times lower.
4. Try SVG export for icons and illustrations if you build websites or design tools.
5. Use saved references wherever consistency matters: mascots, avatars, and personalized content.

The models are available in ChatGPT and Codex, so you can prototype without writing code. The author, for example, uses them to develop website design ideas and generate illustrations and icons.

Bottom line: this is not a revolutionary update, but it is a practical one. It is faster at the same price and adds features that address real pain points: character consistency and vector output.