ChatGPT Images 2.6: Faster and Makes GIFs, but in the API a Regular Image Now Costs 3 to 4 Times More

November 5, 2026, 1:35 a.m. Israel time

OpenAI posts a short announcement: "Images, upgraded." ChatGPT can now make GIFs and remember your reference images, and the API gets two new models that generate images 50% faster.

The post says nothing about price. So I opened the cost calculator in OpenAI's docs, and there, in a single line of code, sit two numbers: 4.2 and 3.3.

That's how many times more tokens a regular image now uses, at the same price per token.

ChatGPT Images 2.6 is OpenAI's new image model, launched on November 4, 2026 in ChatGPT, Codex and the API. That's the official name in ChatGPT. In the API it comes as two models: GPT-Image-2.6 Flare, the fast one, and GPT-Image-2.6 Sunburst, tuned for quality. We read the launch page, the docs, the pricing page and the safety card. We also checked the independent leaderboard and the first developer tests, published in the hours after launch.

In short

ChatGPT Images 2.6 adds GIFs from text, a sketch or a photo, saved reference images (@name), markup on the image for precise edits, and, in the API, SVG files.
In ChatGPT it's an upgrade at no extra cost, on every plan. GIFs are still rolling out, and not everyone has them.
In the API the price per token hasn't changed, but at low and medium quality an image uses 4.2x and 3.3x the tokens. A medium image costs about 4.4 cents instead of 1.3. And there's no Batch discount yet.
No jump in quality: on the Arena leaderboard, 2.6 Sunburst is second, behind 2.5 Sunburst. If your agent generates images in volume, don't switch automatically.

My take, at a glance

GIFNew and useful in ChatGPTA short animation from a product photo or a character, no software. Still rough, but fine for Stories

×3.3Pricier in the API, at mediumSame price per token, more tokens per image. At low the gap is ×4.2. At high and up, no change

#2Not better than 2.5On Arena, 2.5 Sunburst is still first, in both generation and editing

×1.9Faster, in most measurementsOn Arena: 19.4 seconds vs. 36.9 for Sunburst. In one other test, at medium, 2.6 was slower than 2.5

My take: a good upgrade for ChatGPT, a quiet price hike in the API

If you work in ChatGPT, this is an upgrade worth trying today. GIFs and reference images are exactly what a small business owner needs: take one product photo or your brand character and get a consistent series of images plus a short animation for your Stories. No software, no designer.

If you're connected through the API, meaning an agent that generates images on its own, the picture is less rosy. OpenAI wrote "50% faster" and said nothing about cost. The price per token is the same as 2.5, but at the cheaper settings each image uses 3 to 4 times the tokens. And on the big independent leaderboard, quality didn't go up.

Picture a print shop that didn't raise its price per sheet. It's just that a draft print, which used to take one sheet, now takes four. High-quality prints are exactly the same. If you mostly printed drafts, you'll pay more. If you printed at high quality, you won't notice.

What's new in ChatGPT Images 2.6

GIF

From text, a hand sketch or an existing image
Change the speed and ask for fixes in the chat
Still rolling out in ChatGPT. Uploaded GIFs can't be edited

Saved reference images

Save an image once and give it a name
In later chats, type @name, no re-upload
Good for a brand character, a product, you

Markup and precise edits

Mark on the image where to change it
According to OpenAI, the character, outfit and product hold up better across edits
Was in ChatGPT, now in the API too

SVG, API only

A vector file for web, apps and print
In practice: the model makes a regular image, then converts it to vector
Check any logo before sending it to print

Sources: OpenAI's launch page, the Images in ChatGPT help article and release notes, November 4, 2026; the image generation guide in the API docs.

Two more small things. The API adds a new parameter, complexity, which sets how much the model plans before it draws. OpenAI recommends low for quick edits and medium or high for a scene with lots of detail. Note that the default is high. And "images with thinking," where the model plans before drawing, are available in ChatGPT only on Plus, Pro and Business.

OpenAI's own sample prompt gives away how GIFs work: it asks for a 6-frame "sprite sheet" in a 3-by-2 grid. In the API a GIF is 2 to 64 frames, and the default is 9 frames of 0.2 seconds each, a 1.8-second loop. This isn't video. It's a short animation.

OpenAI on X: a GIF demo in ChatGPT Images 2.6, November 4, 2026

OpenAI's own demo. In user tests, below, the result is less smooth.

What an Images 2.6 image costs in the API

This is the article's key finding. On the pricing page, 2.6 Flare and 2.6 Sunburst have exactly the same prices as 2.5: $30 per million image output tokens, $8 per million image input tokens and $5 per million text input tokens (the prompt). But the cost calculator in OpenAI's docs multiplies the token count, for 2.6 only: 4.2x at low quality and 3.3x at medium. At high, xhigh and max there's no multiplier. We found this in the calculator's own code, and calculated the numbers in the table from it.

Quality, 1024 square image	Images 2.5	Images 2.6	Difference
low	0.6 cents	2.5 cents	4.2x
medium	1.3 cents	4.4 cents	3.3x
high	5.3 cents	5.3 cents	Same price
max	21 cents	21 cents	Same price

What 1,000 square images at medium quality cost

Images 2.5, on Batch

$6.6

Images 2.5

$13.2

Images 2.6

$43.5

Platin Digital's calculation based on OpenAI's cost calculator and pricing page, November 5, 2026. Output cost only, excluding text and input files. 2.6 has no Batch price yet: it isn't listed in the Batch table, and the option is switched off on the model pages.

Who will feel it: an agent that generates images in volume at low or medium quality. Think product images for a store, a banner for every post, or drafts to choose from. Who won't: anyone already working at high and up, and of course anyone using ChatGPT on a subscription, where there's no per-image charge. Still, the difference is cents per image. For a small business generating a few dozen images a month, it won't break the budget. It just isn't "the same price," the way it looks on the pricing page. We explained what a token is and why it's counted in What Is a Token in AI?

Is Images 2.6 better than Images 2.5?

According to the big independent leaderboard, no. On Arena, people see two images for the same request without knowing which model made each one, and pick the better one. Here are the standings after a few hours:

Text-to-image, Arena	Rank	Score
GPT-Image-2.5 Sunburst	1	1427
GPT-Image-2.6 Sunburst (Medium)	2	1409
GPT-Image-2.5 Flare	3	1399
GPT-Image-2.6 Flare (Medium)	4	1384
Google's Nano Banana 2.1	Lower	1325

Image editing tells a similar story: 2.6 Sunburst is third, behind both 2.5 models. 2.6 still has only about 3,000 votes per model. The gap between 1409 and 1427 is 18 points, but 2.6's margin of error is about 13 points either way, so the two models' ranges almost touch. That's why it's fair to call the quality "similar," not to say 2.6 is worse. That's also the word Arena attributes to OpenAI itself: "similar quality to Image 2.5," just faster. And in the comics, anime and fantasy category, according to Arena, 2.6 Sunburst actually comes first.

Arena on X: where GPT Image 2.6 lands on the leaderboard, November 5, 2026

The leaderboard with screenshots. 2.6 Sunburst second, 2.6 Flare fourth.

One contradiction is worth flagging. OpenAI's developer account shared a post from DesignArena, a smaller leaderboard, saying 2.6 Sunburst is "#1." When we checked DesignArena's public leaderboard, 2.6 didn't appear on it at all yet, and 2.5 Sunburst was first. It isn't on Artificial Analysis yet either. The launch page itself has no charts or benchmarks: just "50% faster," without saying compared to what, and not a word on quality against competitors or on price. As one Hacker News commenter put it: "They didn't even compare on cost, speed, or any benchmarks at all."

Real tests: what developers found in the first hours

Five tests, and what came out

Speed, same image2.6 at low: 6.7 seconds, 2.5 cents. 2.5 at medium: 12 seconds, 1.4 cents. Nearly 2x faster, at nearly 2x the price

Same request, 2.5 vs. 2.6Three images at medium: 2.6 was slower (48.7 seconds vs. 39.8) and 2.8x pricier. Text came out accurate in both

A design site rolls back to 2.5After an hour: Flare 2.6 "too simplistic" or "inherently broken," so back to 2.5 Flare. Sunburst 2.6 "seems solid"

A hard prompt with logicMax Woolf, a developer who tests image models with especially hard prompts: one of his tougher ones "now surprisingly passes"

GIF from a photo"Surprised that it could do it at all." But a simple animation came out blotchy, and in another test the loop didn't close smoothly

Green: in 2.6's favor. Red: against it. Orange: depends on what you need. Sources: Hacker News commenters and a Kingy AI test, November 4–5, 2026. These are one-off tests, some run only once, so take them as a hint, not a verdict.

The speed is real, but it isn't free. Arena measured median time to a finished image: 2.6 Sunburst at medium took 19.4 seconds vs. 36.9 for 2.5 Sunburst, and 2.6 Flare took 13.4 seconds vs. 21.7. The only measurement we found that comes close to OpenAI's "in as little as six seconds" is from a developer on Hacker News: 2.6 at low produced an image in 6.7 seconds, vs. 12 seconds for 2.5 at medium, and cost nearly twice as much. In a structured test by Kingy AI, with the same three requests at medium, 2.6 was actually slower. The likely explanation, in our view, is the complexity default, which is high. If that's right, you get the speed by tuning the settings, not automatically.

GIFs are a start, not a finished product. In the same test, the API GIF came out at 512 by 512 pixels, 10 frames, with uneven motion and a mark nobody asked for. In ChatGPT it came out even smaller: 358 pixels. For a Story, a social media reply or a waving character, that's enough. For an ad, not yet.

When to switch to Images 2.6

✓

In ChatGPT: yes, todayNo extra cost

GIFs for stories and social
A brand character via @name
Markup edits, with fewer unwanted product changes

In the API: test SunburstOnly if you need what's new

Automated GIFs or SVGs
Markup edits at medium or high, where, according to OpenAI, Sunburst keeps details better
At high and up, where the price hasn't changed

Don't switch automaticallyIf the agent makes lots at low or medium

Product images in volume
Overnight Batch jobs at half price
A pipeline 2.5 Flare already handles well

My rule: in an existing agent, you don't swap models on launch day. Run 2.6 on 20 to 30 real images from the business, with the same settings, and check three things: quality, time and the cost on the invoice. If there's no improvement you can see with your own eyes, stay on 2.5. And if you need GIFs or SVGs, add 2.6 for those tasks only.

Three practical notes before you start:

Every image is labeled as AI. According to OpenAI's safety card, the images carry C2PA metadata, plus an invisible Google watermark (SynthID). Anyone who checks the file with a verification tool, like the one OpenAI released, can tell it was made with AI. That's fine, just worth knowing.
Reference images of people: OpenAI hasn't published how long saved images are kept or how to delete them. Save a brand character or a product. Don't save photos of customers or employees without their consent.
An SVG isn't a designer's logo. The model draws a regular image and converts it to vector. For a simple icon on your website, that may be enough. For signage and print, have someone go over the file.

What I'm doing with it

I usually build agents on Hermes Agent, an open-source framework from Nous Research, and pick the models and tools that fit each business, from OpenAI, Anthropic, Google and others. Platin Digital is an OpenAI Select Partner, and that doesn't change the choice. In agents that generate images, 2.5 stays my default for volume work. I'm testing 2.6 Sunburst for GIF and SVG tasks, and for edits where keeping product details matters, and price and quality will decide. Markup on the image works in the API with 2.5 too.

We cover what content work with an agent looks like beyond images in Writing Website Content With AI, and what an agent costs beyond the model price in How Much Does an AI Agent Cost? And more from OpenAI in recent weeks: GPT-6.1 Terra, where the price per token didn't tell the whole story either.

Frequently asked questions

What is ChatGPT Images 2.6?

ChatGPT Images 2.6 is OpenAI's image model launched on November 4, 2026, after Images 2.0 in April and Images 2.5 in September. It generates images faster, can make GIFs from text, a sketch or a photo, saves reference images for reuse via @name, and holds details better across edits. In the API it comes as two models: gpt-image-2.6-flare, the fast one, and gpt-image-2.6-sunburst, tuned for quality.

Is ChatGPT Images 2.6 free?

According to OpenAI, Images 2.6 is available on every ChatGPT plan, including Free, on desktop, mobile and web, and in Codex too. GIF generation is still rolling out gradually, and not everyone sees it. "Images with thinking" is reserved for Plus, Pro and Business. OpenAI hasn't published how many GIFs or images each plan allows.

How much does an image cost with GPT-Image-2.6 in the API?

The price per token is the same as 2.5: $30 per million image output tokens. But according to OpenAI's calculator, a 2.6 image uses more tokens at low and medium quality. A square image at low costs about 2.5 cents instead of 0.6, at medium about 4.4 cents instead of 1.3, and at high and up the price hasn't changed. 2.6 has no Batch price yet, which saves half on 2.5.

What's the difference between Flare and Sunburst?

Flare is built for speed. According to OpenAI, at low quality it can generate an image in as little as six seconds in the best case, and it improved layout in posters and product images. Sunburst is built for quality and consistency across edits, at medium quality and up. Both cost the same. On the Arena leaderboard, Sunburst ranks above Flare, and one developer has already rolled back from Flare 2.6 to Flare 2.5.

Does Images 2.6 make better images than Images 2.5?

Not clearly. On the Arena leaderboard, where people choose between images without knowing who made them, 2.6 Sunburst is second behind 2.5 Sunburst in generation, and third in editing. The gaps are small and there are still few votes. What's new in 2.6 is speed and new capabilities like GIF and SVG, not a jump in quality.

Can you make a GIF in ChatGPT?

Yes. With Images 2.6 you ask for a looping GIF, or upload an image and ask to animate it. Then you can request changes in the chat, adjust the speed, save and share. GIFs uploaded from elsewhere can't be edited. The animations are short, a few seconds, and in the first tests the quality is still uneven.