OpenAI Images 2.6 Is Here: GIFs, SVG Export, and the New Flare and Sunburst API Models

OpenAI shipped Images 2.6 on November 4, 2026: animated GIF creation inside ChatGPT, reusable reference images, and two new API models called GPT-Image-2.6 Flare and Sunburst. Flare is the speed tier, generating images in as little as six seconds. Sunburst is the quality tier, with better edit consistency, image markup, and native SVG export. OpenAI says the whole stack runs 50% faster than before.

That's the pitch. Here's what actually matters once you start using it.

What's new in ChatGPT Images 2.6

The headline consumer feature is GIFs. You give it a prompt, a photo, or a sketch, and it turns that into a short looping animation. You can adjust playback speed and edit individual frames after the fact. OpenAI is pitching this at marketers: animate a brand mascot, turn a product shot into a social reaction, block out character movement for a storyboard. It's rolling out gradually across all ChatGPT tiers plus Codex, on web, desktop, and mobile.

The second feature is reference images. You save a photo of yourself, your dog, or a recurring character once, then reuse it in new generations without re-uploading every time. Small quality-of-life thing, but if you've ever re-uploaded the same headshot into ChatGPT for the tenth time, you'll get why this exists.

OpenAI also dropped a usage number alongside the launch: people are now generating over 4 billion images a week across its products, up from 3 billion in September. That's a real jump in two months, and it's probably why OpenAI felt the need to ship a speed tier at all. Flare exists because the quality bar for a quick meme or a mascot sticker is a lot lower than for a hero image, and serving all of that at Sunburst-level cost would get expensive fast.

Flare vs Sunburst: what's actually different

Flare is built for the back-and-forth of interactive editing: you're iterating in a chat window and don't want to wait 30 seconds between tweaks. Sunburst is for when you need the same character or product to look the same across five edits in a row, plus it's the only one of the two that does SVG export and markup-guided edits.

If you're building a product around this, the choice is basically: pick Flare for anything users will iterate on live, pick Sunburst for anything that goes into a final asset.

SVG export isn't quite what it sounds like

This is the part I'd slow down on before you build a pipeline around it. SVG export doesn't mean the model reasons natively in vector space. Under the hood it's a raster-to-vector conversion: the API generates a reference raster image first, then runs it through a separate authoring step to trace it into SVG. You call it with output_format: "svg" and n set to 1, and you get back static vector content: no scripts, no external resources, no embedded raster images baked in.

Two things to know before you budget for this. First, it needs an extra svg.model parameter that wasn't obvious from the initial docs. Second, and this one actually matters for your bill: you're charged twice, once for the image model rendering the reference, and again at the authoring model's rate for the SVG tracing tokens. It's a genuinely useful feature if you need clean vector output for a web asset or a print job, but don't assume it's free just because the per-token rate didn't change.

GIFs work the same conceptual way. You set output_format: "gif" with frame_count and frame_duration_ms, defaulting to 9 frames at 200ms. Worth noting: OpenAI's own docs disagree with each other on the frame limit: the guide says 2 to 64 frames, the API reference caps it at 2 to 16. I'd test against the API reference number and not trust the higher one until OpenAI fixes the docs.

The pricing story OpenAI didn't put in the headline

OpenAI's framing is that Flare and Sunburst bill at the same per-token rate as 2.5: $30 per million output tokens. Technically true. But per-token parity isn't per-image parity, and this is where it gets annoying.

At 1024x1024, token usage per image went up a lot between 2.5 and 2.6:

That's roughly 4x more tokens at low quality and 3.3x more at medium quality for the same resolution. Same sticker price per token, 3-4x the actual bill per image. People on Hacker News called this out fast, and "deceptive marketing" came up more than once in that thread. I don't think OpenAI is lying exactly (the per-token rate genuinely didn't move), but leading with "same pricing" when the real cost per image jumped this much is the kind of framing that erodes trust the second someone runs the numbers, which someone always does within a day of launch.

There's also a gap worth knowing if you're running anything at scale: Batch API support is gone for 2.6. GPT-Image-2.5 Flare still gets Batch at $15 per million output tokens, half the standard rate. If your workload is the kind of thing that can tolerate async batch processing to save money, 2.6 doesn't give you that option yet.

Is Images 2.6 actually better than 2.5?

Depends who you ask, and that split shows up clearly in the HN thread. Some people are genuinely happy with 2.6: less garbled text in generated images is the most common praise. Others say 2.5 still wins on benchmarks and makes fewer compositional errors (objects in the wrong place, wrong counts, that kind of thing). The GIF feature in particular got dinged for visible flicker and inconsistency between frames, which tracks with what you'd expect from a model that wasn't trained end-to-end for temporal coherence. The "grid trick" approach to keeping frames consistent isn't new, people have been doing variations of it for a while, OpenAI just made it a first-class feature.

My honest read: this is an incremental release dressed up as a bigger one. GIFs and SVG export are genuinely useful additions if you need them. The underlying image quality isn't a clean win over 2.5, and the pricing move is worse than it looks at first glance. If your workflow is batch-heavy, stay on 2.5 Flare for now. If you need SVG assets or quick GIFs for social content and can absorb the real cost increase, 2.6 is worth trying.

There's a broader thing simmering under this launch too: a few HN commenters pointed out that image generation progress has felt flat lately compared to the pace of improvement in video models, and that the "AI look" (that slightly-too-smooth, slightly-off quality everyone can spot now) hasn't really gone away across any of these releases. Images 2.6 doesn't change that. It's a feature update, not a quality leap.

If you're tracking how the image-gen field is shaking out generally, it's worth comparing this against where Nano Banana 2.1 and the mystery Mona Lisa 1 model landed: different trade-offs, same underlying question of whether anyone has actually cracked photorealism without the tell-tale artifacts. And if you want the wider context of what's shipping across the AI tools space this year, our 2026 roundup covers where image gen fits next to everything else.

Watch the announcements