Sponsored by
When a newer AI model drops, is it just… better?
Yesterday, I made a point that the labs are heads-down on knowledge work and coding related improvements.
Media models though? Quiet.
Q4 2025 was the peak (Sora, the image model wars, all of it) and since then, not a lot.
So when Images 2.5 dropped last week, there was a bit of chatter, but it got buried under everything else going on.
And the last time I did a proper image face-off was over 2 months ago. So for a change of pace, I thought I'd sit down, run the new model, and tell you if it's any good.
But first, some catch up:
Partner with us
Tools that caught my attention
Tencent's "agentic creative studio with memory," where you hand it a brief and its agent builds a full set of on-brand assets on a single editable canvas. It remembers your style and brand rules across sessions, so it stops asking you to re-explain your taste every time.
The upscaler creators use to turn a soft AI image into a crisp, high-res one, now grown into a broader image, video and audio platform with over a million subscribers. Think of it as the finishing step after you generate something you like.
A brand-and-vector focused image generator with strong control over text and layout, on V4 now. Built for designers and marketers who want on-brand, editable output rather than one-off art.
The setup
Images 2.5, on the API, comes as two models.
Flare is the fast one, and it's the default (basically what you get inside ChatGPT now), built for speed and volume.
Sunburst is the slower, better one, tuned for careful edits and detailed work where you want it to change one thing and leave everything else alone.
Both cost the same per image, Sunburst just takes longer. Since I wanted to see the best 2.5 can do, I ran everything on Sunburst.
The matchup: GPT Image 2 (the old one), GPT Image 2.5 Sunburst (the new one), ran side by side on OpenRouter.
Test 1: can it put me in a poster and keep my face
The whole promise of 2.5 is that it holds onto a face while it reinvents everything around it. So I fed both models a photo of me and asked for the full treatment.
Prompt: Turn this photo of me into a hand-painted 1970s Bollywood movie poster. Put my face on the lead hero, dramatic pose, bold film title in Devanagari at the top, side characters, a retro action scene, folded and slightly faded printed-paper texture, saturated colours.
My take: Not even close. The old one gave me a modern illustration that could be any stock hero.
2.5 gave me a folded paper, sun-fade, a whole second villain character it invented on its own, that hand-painted grain you remember from single-screen theatres, and it still looks like me.
Test 2: does it remember 1998
I wanted to see which one could build a memory from scratch, a very specific one.
Prompt: A middle-class Indian living room on a Sunday morning in 1998. A boxy TV playing Doordarshan, a steel tumbler of chai gone cold on a wooden stool, incense smoke drifting, soft light through a window grille falling on a terrazzo floor. Warm, lived-in, a little faded.
My take:
Look at what 2.5 added without being asked: a 1998 wall calendar, a stack of VHS tapes labelled Sholay, Border, Taal, the Doordarshan logo on the screen, a newspaper reading "India tests nuclear capability", chappals by the door.
The old model just drew a darker, emptier room that's technically correct. That's the whole upgrade in one frame.
Test 3: give me something shareable (and about AI)
Wanted one that's pure fun and, since this is an AI newsletter, about AI.
So: the classic Indian dairy hoarding, the hand-painted kind with the chubby mascot and a pun, except the pun is about AI taking your job.
Prompt: A hand-painted topical billboard cartoon in the classic Indian dairy-advertisement style. A cheerful chubby cartoon girl mascot in a polka-dot dress waving, bright blue background, a witty punny headline in bold red about AI taking over office jobs, an original made-up dairy brand logo in the corner, aged weathered hoarding texture.
My take: Honestly a tie.
The old model has a better line with better lettering. 2.5 went darker, funnier, and a mascot clutching an AI tablet, though a couple of its letters wobble if you look closely.
Two months of newsletters and this is the frame I most want to print and stick on a wall.
(Quick note: I told it to invent its own brand, so please don't go recreating these with the real Amul girl, that's someone's trademark.)
Test 4: the one where the new model loses
Tiny vintage matchbox labels are the kind of small-type, fine-line print work 2.5 is supposed to be better at. So this should have been an easy win for the new model.
Prompt: A sheet of four vintage Indian matchbox labels for made-up safety-match brands, 1950s lithograph style, bold flat colours, simple icons (elephant, lion, rooster, lotus), small brand text, aged foxed paper.
My take: The old model won this.
GPT Image 2 gave me four crisp, bright, perfectly readable labels that look like they came off a real 1950s press.
2.5 over-did the aging, muddied the colours, and smeared some of the fine detail into that heavy distressed look it seems to love.
For flat graphic print work with clean lines, newer just didn't mean better. Good reminder that these upgrades come with a personality of their own.
My take
Two things I'm walking away with:
One, 2.5 is a jump, but of a specific kind.
Everywhere the win was about texture, light, and making something look like a real object in the world. That reference-and-realism stuff is definitely better.
Two, that same personality is why it lost the matchbox.
On flat, graphic print work, all that extra texture and aging becomes noise, and the older model's simpler hand looks nicer.
So "is 2.5 better" has an annoying answer: depends what you're making.
Zooming out, this was a nice reminder of how lopsided things have gotten
A media model this good launched last week and barely made a dent on Twitter, because the labs and all of us are so deep in coding / knowledge work right now.
It was fun to spend a day on this. You should try these on your own stuff, the poster one especially. Reply and show me, I'll feature the best ones.
Until next time,
Vaibhav 🤝🏻
If you read till here, you might find this interesting
#AD 1
AI agents built to get work done
Skydive agents take on work for you, using the same tools your team already uses.
Talk to your agent on the web, Slack, email, iMessage, or right from your terminal. Wherever you pick up the conversation, your agent keeps the context.
Put agents to work across customer support, sales, marketing, engineering, ops, and more.
Give your agent a job. They’ll take it from there.
#AD 2
Scale LinkedIn like you scale cold email
You can finally scale your LinkedIn outreach without bothering your team. Aimfox lets you add dedicated LinkedIn profiles that start conversations and route interested replies to your existing sales team.
