Ship week on a review desk is a job split. Drafts are allowed to wander. The face that goes out the door is not. A ChatGPT Images 2.5 thread helps when someone pins the precision tier before the last pass, and leaves the fast default for the pile nobody will publish. Texture arguments can wait. The wrong unlocked tier cannot.
Reviewers who cannot ship used to lose a day to one pretty still that drifted the mouth on pass four. The new habit is to treat every model name as a quality brag. Sunburst gets praised like a premium skin. Flare gets insulted like a free demo. That talk misses the job. Flare is the lane for iterating a face you do not yet owe anyone. Sunburst is the lane for holding the face you already approved. Pinning is how the desk stops mixing those lanes in one composer. The claim worth testing is not “which looks richer.” The claim is whether the final still still belongs to the same person after the notes land.
Drafts Iterate And Finals Hold The Face
A geek commentary desk does not ship a vibe. It ships a still a reader can match to a trailer, a figure photo, or a press kit. If the mouth shape moved while the lighting got more expensive, the file is a different person. A different person on ship week is a kill, not a compliment to the renderer.
Drafts exist so the writer can try a colder key, a tighter crop, a less muddy jacket. Those tries should be cheap and replaceable. Finals exist so the same face survives the last three notes from a stubborn editor. Those notes should not reopen the identity. Mixing the two jobs in one unlocked chat is how a hero still becomes a cousin of the actor. Cousins do not ship.
The face is the object. Flare and Sunburst are two ways to treat that object. One is for hunting. One is for holding. Reviewers who collapse that split into “better versus worse” will keep losing nights to a prettier stranger.
What A Reviewer Scores Before Anyone Ships
Score the still the way a comments section will score it. Does the person still match the reference? Did the last note change only the asked region? Did the mouth, the scar, the hairline stay put while the jacket got cleaner? If any of those fail, the file is wrong, not late. Wrong files do not go in the pack “as backups.” Backups of drifted faces get reused on the next recap and teach the desk that drift is a style.
Time pressure on ship week makes the wrong tier tempting. Flare will give you another look while the Slack thread is still open. That speed is useful only while the still is a draft. The moment an editor says “that face, lock it,” speed stops being the score. Hold becomes the score. A reviewer who keeps the fast default pinned for the lock is asking the model to hunt again. Hunting after lock is how the scar disappears.
Write the score on a card before the last generate. Face match. Asked change only. No new jewelry. No new haircut. If the card would fail, do not click. The click is cheaper than the hour you will spend explaining why the still no longer matches the figure on the shelf. Ship week comments will quote the still, not your Slack thread. If the face already failed the card, the quote will be about a cousin.
Flare Versus Sunburst On The Same Face
Flare is the fast default in the catalog. Sunburst is the precision tier, built for tighter control across successive edits, and it takes longer. Those are job labels, not trophies. A reviewer who says “always Sunburst” is bragging about spend they have not justified. A reviewer who never pins anything is hoping the agent guesses the lock.
The useful comparison is one face, two moments. Early in the thread, you want looks. Late in the thread, you want the same person after “darker jacket, same mouth.” If both moments use the same unlocked default, the late moment often hunts. Hunting is Flare doing what Flare is for. The fail is asking Flare to do the lock.
| Job | Flare | Sunburst |
| Early drafts | Agent can pick it and iterate | Usually slower than the hunt needs |
| Last notes on an approved face | Risk of a new mouth or hairline | Pin it before the last pass |
| Reviewer move | Leave it unlocked while hunting | Pin in the composer, then ask one change |
| Fail signal | A cousin of the actor ships | A long wait with no pin still drifts |
The table is a ship card. It does not rank beauty. If the face already passed and you still need a jacket note, pin the precision tier first. If you are still choosing between three ugly cousins, let the fast default work and throw the stills away. Thrown drafts are cheap. A shipped cousin is a recap you have to retract. Writers who invert the table — pinning precision for the hunt, then unlocking the default for the lock — pay twice. They wait on drafts nobody will keep, then they let the last note wander. The split only works in that order: hunt on Flare, hold on a pinned Sunburst.
Two API Names Split Hunt From Hold
OpenAI shipped that split on September 8, 2026 as GPT-Image-2.5 Flare and GPT-Image-2.5 Sunburst. Flare, in the API note, is “the default choice for most applications,” aimed at creator work, product stills, visual search, and high-volume drafts. Sunburst is “built for premium visual workflows that benefit from tighter control across edits,” the lane OpenAI names for campaign creative and polished product imagery. Those are job labels. They are not a quality brag.
The same day OpenAI said people already create more than three billion images a week across ChatGPT Images and the GPT-Image models. Volume explains why a fast default exists. It does not tell a ship-week reviewer to leave Auto on for the lock. Pin Sunburst when the mouth already passed. Leave Flare for the cousins you will throw away.
Agent Pick And Pin Inside The Composer
The agent can pick Flare or Sunburst from the request. That pick is useful when you have not named a lane yet. “A colder key on this person” is a hunt. The agent can take the fast default and you can keep talking. “Same face, darker jacket, no new mouth” is a lock. Name Sunburst, or pin it, before that sentence. If you only type the lock sentence and leave the composer on Auto, you are betting the agent heard “hold” instead of “hunt.” On ship week, do not bet. Open the composer and pin the precision tier with your own hand.
Images 2.5 puts that choice in the composer. You can leave the pick to the agent, or you can pin a variant yourself. Pinning is the move that turns a quality argument into a job split. The pin is a door you close so the last note cannot wander into a new identity, not a badge on the still.
Let The Agent Pick Early Draft Tiers
Early stills should be disposable. If the first three faces are wrong, that is the hunt working. Do not pin Sunburst to “make drafts look finished.” Finished-looking drafts are how a writer falls in love with a cousin and then spends the night defending a mouth the reference never had.
Ask for looks in plain speech. Attach the reference if the person already exists. Let the agent choose the fast default when you are still rejecting piles. Rejecting is the job. Publishing is later. A commentary desk that treats every generate like a final will pin too early and then refuse to throw work away.
Pin Sunburst When The Face Must Hold
When the editor points at one still and says that face, open the composer and pin the precision tier. Then ask one change. One. “Darker jacket. Same scar.” If you also ask for a new crop, a new sky, and a cleaner logo in the same paragraph, you reopened the hunt with fancier lighting. The pin cannot save a novel.
This is where the Images 2.5 AI Image Agent earns a place on a ship-week machine: the same chat can hunt, then hold, if someone actually uses the pin. Writers who never open the composer will keep arguing about “the model” as if there were one. There are two lanes. The pin is how you pick the hold lane on purpose.
Multi-turn Edits Have To Keep The Face
The reason a lock is even possible is that this model generation is built to follow notes across many turns and to keep a person from a reference recognizable. That is the multi-turn claim a reviewer can actually use. An unlocked fast default will not automatically hold a mouth through five stacked wishes. Consistency is a property you test by asking for one change and looking at the unasked regions. If the jacket moved and the scar moved, the turn failed even when the lighting looks expensive.
Images 2.5 keeps the thread in one conversation, which is the only way that test is honest. A new chat is a new hunt. If the face already passed, do not start over to “get a cleaner seed.” Starting over throws away the only thing the editor approved. Stay. Pin. Ask the jacket note. Look at the mouth.
Follow-ups That Keep The Same Person
A good follow-up names the stay and the move. Stay: the face, the scar, the hairline. Move: the jacket, the key, the crop if the ears still show. Then generate once. Put the new still next to the approved one. If the mouth moved, discard. Do not “fix” the mouth in the same paragraph as a second jacket note. Two fixes are a new hunt.
Bad follow-ups sound like recap copy. “Make it more like the trailer, more heroic, more 2026.” Heroic is how a new jaw arrives. Trailer is how a second costume sneaks in. The review needed this person in this jacket. The trailer already exists. Do not ask the lock pass to rewrite the film.
Stop Stacking Asks After The Face Locks
Ship week dies in stacked notes. Three Slack comments become one paragraph. The paragraph asks for jacket, sky, and logo. The precision tier then has permission to solve all three. Solving all three is how the unasked mouth drifts while everyone watches the logo. Split the notes. Jacket first. Check the face. Sky later, or never.
If the reference was soft, the hold will be soft. Pinning Sunburst does not invent a true scar that left the crop. Get a tighter still of the real face, then pin, then ask. A long wait on a blind crop is still a cousin. The wait is not the proof. The face match is the proof.
Where The Pin Still Needs A Human Check
The pin does not replace a reviewer. Soft references still drift. Stacked notes still wander. A precision pass can hold a face and still invent a logo the brief never had. Look at the still next to the reference before anyone ships. If the mouth moved, discard. If you will not look, the pin was theatre.
Pin The Face Then Decide Who Ships
Use Flare for the hunt. Skip an unlocked lock pass. Pin Sunburst the moment a face has to survive the last note, and leave the quality-brag thread that treats every generate like a final. A desk that will not look at the mouth after the click should not pin anything.
Images 2.5 belongs on a commentary machine that will actually split those jobs. A writer who wants one pretty still and no pin can keep a single unlocked default. A ship-week reviewer cannot. The face is the object. The pin is the split. Texture can wait until the person in the still is still the person in the reference.
Readers who know the actor will not debate which tier sounded more premium. They will ask why the recap invented a cousin. The answer has to be that the cousin never shipped.
