GPT Image 2.5 Tested: Flare, Sunburst, and Astra as the Real Winner
OpenAI's GPT Image 2.5 fixes a couple of Image 2's nagging issues but plays it fairly safe on its own — the real upgrade is GPT-6 Astra riding along as the brain, driving conversational edits and research-backed prompts. I ran it through the standard tests, a four-angle grid matchup against Nano Banana, and a round of arguing with it over some Fury Road stills.
OpenAI dropped GPT Image 2.5 a few days after GPT-6 Astra, and by the end of testing it, this had turned into more of an Astra video than an image-model one. My read going in, and still my read now: it’s the .5 update. A couple of Image 2’s nagging issues are better — that noise pattern is reduced, not gone — but what actually changes things is putting real intelligence and conversational editing on top of the model. That’s most of what’s below.
Flare vs Sunburst: the specs
Two versions: Flare, the fast one, and Sunburst, the heavier model for precise edits. Per OpenAI’s docs, both are available to all ChatGPT and Codex users now, and through the API at the same price as GPT Image 2 — though OpenAI doesn’t say which version ChatGPT and Codex actually run under the hood. My guess is Flare. There are also quality settings from low up to max. Native output in ChatGPT/Codex came in at 1672×941 for me, so plan on an upscale.
The standard tests: suit, wine glass, pelican
Running the usual battery first. The man in the blue business suit on a busy street is fine — good background and crowd detail, no blob faces — but the man himself has that pasted-in, comped-into-the-scene look 2.5 has been getting flagged for. The wine glass filled to the top with a clock reading 5:15 got both details right, if a little flatly rendered — which is the point of the test: a thinking model works from the actual instructions rather than what’s aesthetically typical, so it’ll fill a glass all the way and set a clock to a time nobody asks for. A diffusion model like Midjourney tends to default to what its training data says wine glasses and clocks usually look like — poured halfway, hands at 10:10 for the symmetry. Midjourney wins on beauty; 2.5 wins on doing what it’s told.
| Test | Result |
|---|---|
| Blue suit, busy street | Strong background/crowd detail; the man himself looks comped in |
| Wine glass + clock at 5:15 | Both details correct, a little uninspired |
| Pelican on a bicycle | Stayed photoreal instead of going cartoony; pedal placement slightly off |
| All three combined, plus wine | Starts to fall apart — flamingo-stretched legs, a redundant digital clock, both pedals on one side |
Worth calling out on its own: handing the wine-glass image to Astra and asking it to empty the glass, advance an hour, and leave evidence someone killed the rest of the bottle. The clock moved forward correctly, the bottle showed up empty, and — unprompted — it added a drunk text to an ex and a receipt for a $2,800 life-size bronze goose. We’ve all made a questionable purchase after a glass of wine. It made sense at the time.
Editing with a brain: the 12-years-later edit and Tim Square
Handed a Midjourney-generated, commercial-photography-style image with the prompt “show me what happens 12 years later,” 2.5 added home improvements and, more affectingly, a kid’s bike sitting out front. It didn’t shy away from the toll of raising kids either, which tracks — they get all that energy from somewhere, and it’s the parents.
Testing real-world knowledge next: a Times Square takeover where every billboard is mine, run on Astra’s Ultra tier for the extra research budget. The image itself is a little rough on composition, but Astra went out and found real references — a billboard for Death Walks Into A Bar, a nod to Paperclip Heart (the film behind the production-office breakdown), and branding pulled from the site. Bonus points for the pun: welcome to Tim Square.
The mirror test and the 4-angle grid
Flamethrower Girl is back for the funhouse-mirror test, also run on Astra’s Ultra tier. Mirrored text reads backwards correctly, which is the actual pass condition, but the reflections don’t consistently match her pose or braid placement — right on the surface detail, gaps in the logic underneath.
The four-angle grid held up better: one 21:9 image split into four quadrants, four angles on the same moment, character consistency intact throughout — in my opinion, better than Nano Banana 2 and Nano Banana Pro, and a cousin of the 2×2 grid technique from the cinematic mega-prompt breakdown. It also caught its own mistake — noticed the feet were cropped out of frame and volunteered a corrected version pulling the camera back, unasked.
Tales from the Loop, graded like War of the Worlds
Same grid method, zero direction, on a Tales from the Loop-style image — mundane middle-America life with a giant robot looming in the background. It nailed the tone without being told what the tone was; a school bus door and the robot’s placement are both slightly off if you go looking, but for no direction given, it’s a solid hit.
Then, since I’d just rewatched Spielberg’s War of the Worlds and that color grade is a choice, I asked Astra to look at the film’s grading and apply it to the same image: blown-out sky, crushed blacks, halation on the lights. It knew exactly what I was asking for.
GTA VI stills, Fury Road, and arguing with it
Pointing Astra at Rockstar’s GTA VI screenshot gallery and asking for cinematic versions of the stills went badly at first — genuinely bad output. What worked was Astra’s own idea: research who actually played Jason and Lucia in the game and use them as photo references. The results are a legitimately interesting idea — actors as their characters, in photographs that were never actually taken, since the real capture was, I’m guessing, mocap suits in a green-screen warehouse rather than an actual shoot.
Fury Road stills were the harder case, and the one place 2.5 is both good and annoying at once: the first pass, dropping Flamethrower Girl into that world, came back competent but bland — Photoshop-comp energy, with the grain pattern that’s characteristic of the GPT Image family creeping back in (a denoiser pass mostly beats that out, but it’s an extra step). Saying so — “these don’t feel cinematic” — and asking again got a noticeably better pass. The pattern holds across this whole video: you’ll get somewhere, but you’ll probably have to argue with it first.
Flare vs Sunburst at max quality — and the verdict
For comparison, I ran the same prompt through Sunburst at max quality on fal. There’s a difference, but to my eye it’s subtle enough that a creative upscaler on low would close most of the gap. If you’re on the free tiers in ChatGPT or Codex, you can always pull the prompt it used and run it through an API version to compare.
Overall: 2.5 doesn’t fix every issue GPT Image has had, but the real upgrade isn’t the image model on its own — it’s having Astra on board as the brain running it.
Astra, now the daily driver
Astra showed up too late for my last video about it; a weekend later, everything had changed, and it replaced Fable 5.1 as my daily driver almost immediately. The clearest proof was practical: I took The Bridge, my oldest AI short film, built back in the Veo-2 era, and redid the entire thing with Astra and computer use — finished in a day. Full breakdown on that next week. I’m not drinking the AGI Kool-Aid, but Astra is the first model in a while that’s dropped my jaw and left me asking what I actually do with it.
If you want the pecking order: Midjourney still wins on pure aesthetics, 2.5 wins on doing what it’s told, and on the four-angle grid it beat Nano Banana 2 and Pro for me. Pick by the job. But the honest headline is that the image model isn’t the story here. Astra is.