OpenAI's GPT Image 2 spent most of August and early September at or near the top of the Artificial Analysis Image Arena — the closest thing the industry has to an independent, blind-vote leaderboard for image models. Then, on September 8th, OpenAI shipped GPT Image 2.5. Within a day, it had taken over both of Arena's image boards, pushing GPT Image 2 down the list.
That's not a one-off. It's the pattern right now: Google, OpenAI, Black Forest Labs, and Midjourney are all shipping updates on a cadence measured in weeks, sometimes days, and the leaderboard reshuffles every time one of them lands. If you picked a product photography tool six months ago because it used "the #1 model," it almost certainly doesn't anymore — not because the tool got worse, but because the ranking is a snapshot, not a fact.
Why chasing the leaderboard is the wrong strategy for product photos
A general-purpose leaderboard measures one thing: which model produces images people prefer in a blind head-to-head vote, averaged across every kind of prompt anyone throws at it — portraits, landscapes, abstract art, product shots, all mixed together. That's a reasonable way to rank models for "best all-purpose image generator." It's a bad way to pick a model for a specific, narrow job like keep this label's text legible while I put the product in a new scene.
Models don't move up and down the leaderboard at the same rate on every sub-skill. A model can lose ground on the overall ranking while staying genuinely excellent at the one thing you actually need — and the reverse happens just as often. Betting your whole pipeline on whichever model tops the general leaderboard this week means re-testing your entire product line every time the rankings shuffle, which is often.
What we do instead
Oranokai doesn't route every generation through one "best" model. Different slots in a product photo need different strengths, so the model changes per job:
- GPT Image 2 (available in Low/Medium/High tiers) is what we point people toward specifically for text on packaging and labels — ingredient lists, brand names, certifications, the fine print that most image models redraw into gibberish the moment a product changes scenes. This is the exact failure mode we wrote about back in August — held-product shots where the label comes back almost-but-not-quite legible.
- Nano Banana Pro and Flux Kontext cover other jobs — general scene composition, background placement, consistency across a multi-image set — where GPT Image 2 isn't necessarily the strongest fit.
None of that depends on whichever model happens to sit at #1 on a given Tuesday. It depends on which model is actually good at the specific thing a given image slot needs to do, tested against that job, not against a generic benchmark average.
What this means if you're evaluating tools
Don't ask a product photography tool "which model do you use?" as if there's one right answer. Ask instead:
- Does it use more than one model, or does it force everything through one pipeline regardless of what the shot needs?
- For your specific failure mode — label text, in this case — has it actually been tested, or is that an assumption?
- What happens when the ranking shuffles again next month? If the answer is "we re-evaluate what to use, not which brand name to chase," that's the right answer.
A tool that's tied its identity to one model's leaderboard position is one update cycle away from being wrong. A tool that matches models to jobs stays right regardless of who's #1 this week.
GPT Image 2's label/packaging tiers are live now on the Generate page (paid plans — the fidelity these tiers are built for costs more per image than the free-tier ceiling covers). Try it on your own product photo →