OpenAI Releases GPT-Image 2.6: Two Models, GIF and SVG Export via API

Image generation via API has long been a compromise: either fast and approximate, or beautiful and slow, with everything extra then fixed by hand in a graphics editor. With GPT-Image 2.6, OpenAI is trying to cut through this dilemma: the API now offers two specialized models, along with GIF and vector export.

What exactly is new

Flare is the fast model. On default settings it returns a result in six seconds. It suits interfaces where the user is waiting for an answer right now: previews, avatars, draft covers.

Sunburst is built for high demands on detail and visual consistency. A typical task for it is a series of illustrations in one style or a character whose face doesn't change from frame to frame.

Besides the split into two models, version 2.6 includes several other changes:

• GIF animation generation from text or a sketch: mascots, storyboards, reactions for social media;
• graphic markup in the API, which lets you specify precisely which area of an image to edit;
• SVG export;
• a complexity parameter that divides the model's resources between quick edits and calculating scenes with many constraints.

GPT-Image 2.6 is available to all ChatGPT and Codex users regardless of plan.

What it looks like in practice

Imagine an app that makes stickers for small businesses. While the user is settling on an idea, you run drafts through Flare: six seconds is a tolerable wait. Once a variant is chosen, Sunburst builds the final version so the mascot matches the style of the rest of the set.

Next, markup comes into play. Previously, a request like "change the color of the hat" could redraw half the picture. Now you can select an area and limit the edit to it. For such small edits, low complexity is enough, while for a scene with a dozen conditions ("three characters, logo on the left, sunset, no text") the value should be raised.

SVG is useful because logos and icons scale without losing quality and can be edited in code. But the announcement has no figures on vector quality or model pricing, so test on your own tasks before building it into production.

Context: the AI bill is becoming an argument

The complexity parameter didn't appear by chance: everyone is counting money right now. According to Bloomberg, Apple lowered the daily cap on engineers' AI spending from 300 to 200 dollars. For some teams it may drop to 100 dollars, and 25 is being discussed for non-technical divisions. Those with complex tasks will have to ask management for an extension, while the rest are advised to use cheap or local models. Microsoft has revised its internal quotas, Tesla capped spending at 200 dollars a week, and Uber at 1,500 dollars a month on coding tools.

Cognition follows the same logic: its SWE-2 Flash is built for routine agent work, code review, and documentation. It is a Mixture-of-Experts model with 276 billion total and 12 billion active parameters, which the company says is three times faster and five times cheaper than the larger SWE-2. There are no plans to publish the weights because of dual-use risks. On the Pro, Max, and Teams plans, the model is free in Devin Desktop and the console until the end of November 2026.

What to do about it

OpenAI, Cognition, and Apple's budgets share a common principle: don't run every task through the most expensive model. If you are building a product with graphics generation, set up routing by request type.

• Send drafts and interactive previews to Flare, and final assets to Sunburst.
• For local edits, use markup and low complexity.
• For logos and icons, test SVG export on real examples.
• Measure cost and latency on your own scenarios before launch.

The main takeaway: the market is moving from the question "which model is best" to "which model is good enough for this money." Whoever learns to pick a model for the task will spend less for the same result.