GPT Image 2.6 vs 2.5: Faster, Not Cheaper, Not Yet Better

On November 4, OpenAI released Images 2.6 in ChatGPT, Codex and its API, with two API models: GPT-Image-2.6 Flare for speed and GPT-Image-2.6 Sunburst for detail. The company says it generates images 50% faster than before, adds GIFs, and keeps characters, products and brand details more consistent across edits. Within a day, developers had run their own tests, and the picture is more mixed than the launch post.

For a brand or an agency, the useful question is narrow: does 2.6 change what a campaign costs, how fast a team iterates, or what a client can approve? Here is what the official documents say, what the first independent tests measured, and what to check before switching.

What OpenAI says changed

The launch post makes five claims. Generation is 50% faster than before. Flare, at low quality, can return an image in as little as six seconds, and OpenAI says it improves layout for posters and product imagery. Sunburst targets higher-quality images and better consistency across edits at medium or high quality. GIFs arrive for animated mascots, reactions and storyboards. In the API, image markup lets a team point to the exact area to change, SVG export is added, and a new complexity setting controls how much planning the model does; OpenAI recommends low complexity for most edits.

Availability is broad: all ChatGPT tiers on desktop, mobile and web, Codex, and the API. OpenAI also says people now create 4 billion images a week across its products.

Price per token: unchanged. Price per image: depends

On OpenAI’s pricing page, 2.6 Flare, 2.6 Sunburst, 2.5 Flare and 2.5 Sunburst carry the same rates: $8 per million image input tokens, $30 per million image output tokens and $5 per million text input tokens. The same page lists 2.5 in a second table at exactly half those rates, which looks like batch pricing; 2.6 does not appear there yet.

Identical token prices do not mean identical image prices. CellCog ran 80 generations on launch night and found that, for the same 1024 × 1024 image, 2.6 used 4.2 times the output tokens of 2.5 at low quality (824 against 196) and 3.3 times at medium (1,449 against 439). From high quality up, the counts matched. Kingy AI saw the same direction: three Sunburst stills cost about $0.045 on 2.5 and about $0.125 on 2.6.

Output tokens for the same image

1024 × 1024 pictogram, by quality setting

Output tokens for the same image

Setting	Tokens
2.5 · low	196
2.6 · low	824
2.5 · medium	439
2.6 · medium	1,449
2.5 · high	1,758
2.6 · high	1,758

For a team that drafts dozens of low-quality options before choosing one, that difference shows up on the invoice. For a team that renders finals at high quality, the per-image cost is about the same.

Faster, but mostly Flare

CellCog’s timings support the speed claim for Flare: across four prompts, 2.6 Flare was roughly twice as fast as 2.5 Flare. Sunburst went the other way. On the same prompts, 2.6 Sunburst was slower than 2.5 Sunburst in every row, by 17% on one image and 27% on another. These are timings from one laptop on one evening, not a fleet benchmark, but they are the clearest numbers published so far.

Time to render one poster, high quality

1024 × 1536, one run per model, one laptop

Time to render one poster, high quality

Model	Time
2.5 Flare	15.1 s
2.6 Flare	8.0 s
2.5 Sunburst	25.3 s
2.6 Sunburst	25.4 s

Better pictures? Not proven yet

On Arena’s text-to-image leaderboard, as reported by i-SCOOP on November 5, 2.6 Sunburst at medium quality scored 1,409 and ranked second, and 2.6 Flare scored 1,384 and ranked fourth. The same report puts 2.5 Sunburst at 1,427 and first in late October. Blind votes, then, do not yet place 2.6 above 2.5 on image quality. Arena said both new models run almost twice as fast.

Hands-on reviews point the same way. Kingy AI found both versions rendered every requested text string correctly, and neither kept a product design fully consistent across its test. Analytics Vidhya called a markup edit on an armchair’s fabric surgical, but judged text accuracy on a bilingual menu not reliable. One developer quoted from Hacker News by eesel AI found Sunburst solid but rolled Flare back to 2.5, judging its results too simple. CellCog could not get SVG export to work reliably: 14 of 15 requests failed on launch night.

GIFs: useful for social, not a video replacement

OpenAI pitches GIFs for animated mascots, social reactions and storyboards before video production. Kingy AI’s first GIF came out at 512 × 512 pixels and 10 frames, with a loop that did not close cleanly. That is enough for a reaction post or a quick animatic. It is not a substitute for a produced spot, where motion, sound and continuity across shots carry the message.

What this changes for a brand campaign

Speed matters most in the early rounds, when a creative team is looking for the idea: more options in the same hour means more decisions in front of the client sooner. It matters less at the end, when every frame has to match the product, the logo and the previous shot. That part still depends on references, review and someone deciding what stays, and no release removes it.

Before switching a pipeline, three checks are worth an hour. Run your own brief at the quality you actually use and compare cost per approved image, not cost per token. Test edits on your real product and logo, since consistency is the claim that matters most and the one least proven so far. And keep 2.5 available until 2.6 shows up in batch pricing and independent leaderboards. For how we weigh any new model before it touches client work, see how to pick a model and tries per usable shot.