GPT Image 2.5 vs GPT Image 2
Five months separate the two releases. This page lines up what is measurably different and what is unchanged under the hood.
GPT Image 2 landed on April 21, 2026 and was, by most accounts, a genuine step change: 2K output, multiple aspect ratios, web research for current information, competent handling of non-Latin scripts, and a split between Instant and Thinking modes. It took the top slot on the image-editing arena and pushed in-image text rendering close to the practical ceiling.
Five months later, GPT Image 2.5 does not try to beat that list. It fixes the parts that annoyed people daily. The distinction matters when you are deciding whether to migrate, because the headline benefits are not visible in a single generated image — they show up on turn four of an edit chain, and in the latency column of a queue dashboard.
Line by line
| GPT Image 2.5 | GPT Image 2 | |
|---|---|---|
| Released | September 8, 2026 | April 21, 2026 |
| API models | Flare and Sunburst | gpt-image-2 (single line) |
| Generation latency | Up to 50% lower | Baseline |
| Quality settings | low / medium / high / xhigh / max / auto | low / medium / high / auto |
| Token rates | $5 text in, $8 image in, $30 image out | Identical rates |
| Scoped editing | Preserves untouched regions reliably | Frequently rewrote nearby areas |
| Multi-turn edits | Earlier edits persist through the thread | Quality degraded over long chains |
| Transparent backgrounds | Improved, called out by early adopters | Supported but inconsistent |
| In-chat drawing | Sketch, via @Sketch | Not available |
| Templates and shared prompts | Yes | No |
Should you migrate? By workload
Single-shot generation at scale
UpgradeSame token rates, up to 50% less latency. There is no argument for staying.
Multi-turn retouching
UpgradeScoped edits and edit persistence are the entire release. This is the workload it was built for.
Transparent overlay assets
UpgradeManus flagged transparent-background generation specifically as improved.
A pipeline that already works
Test firstNothing is deprecated. Route new traffic to Flare, compare on your own prompts, cut over when the numbers agree.
What is genuinely unchanged
Pricing. OpenAI's model pages list identical token rates for both generations — $5 per million text input tokens, $8 per million image input tokens, $30 per million image output tokens. An upgrade that improves speed and holds price is rare enough to state plainly, and it removes the usual reason to delay a migration.
The API contract. Same endpoints, same auth, same billing model, same request shapes. What changes is the model string and the availability of two extra quality tiers. For most codebases the migration is a constant and a config review, not a rewrite.
The resolution story. OpenAI published no new ceiling with this release. The line's stated maximum has been 2K, and the previous generation's API sizes topped out at 1024×1536 and 1536×1024. If you need native 4K, this release does not change your answer, and the comparison you actually want is against Google's line rather than against the previous OpenAI model.
The three complaints this release targets
Generation was slow. Autoregressive image generation has been the slow part of this family since the beginning. Cutting latency by up to half does not just save time; it changes whether iteration inside a conversation is practical at all.
Subjects morphed between edits. Under GPT Image 2, an edit regenerated the picture. The face drifted, the crop shifted, the background rearranged itself. GPT Image 2.5 preserves the subject through a change, which is what launch partner Higgsfield AI described in its own evaluation.
Edits rewrote things you did not touch. The scoped-edit work is the direct answer, and it extends to complex frames where subject and background are hard to separate. Combined with edit persistence across a long thread, it is the difference between a generator you re-prompt and an editor you iterate with.
How to run the comparison yourself
Vendor numbers describe an average workload that is unlikely to be yours, so the only comparison that settles anything is one run on your own prompts. Pick five that represent your real mix — not five impressive ones — and hold everything constant except the model string.
Measure three things. Wall-clock latency at your typical size and quality, because that is where the headline claim either shows up for you or does not. Corrections-to-acceptable, meaning how many edit turns it takes to reach something you would ship, which is the metric the scoped-edit work should move. And cost per accepted asset rather than cost per generation, since a model that needs half as many attempts is cheaper at the same per-image rate.
That third number is the one that usually decides it. Per-image pricing is identical across the two generations, so any reduction in wasted attempts drops straight to the bottom line. A team that currently burns four rejected renders per shipped asset and gets that to two has halved its image budget without changing a single price.