On September 8, 2026, OpenAI reset the image-generation board. ChatGPT Images 2.5 shipped to every ChatGPT, ChatGPT Work, and Codex tier worldwide, and two new API models — GPT-Image-2.5 Flare and GPT-Image-2.5 Sunburst — landed on the pricing page the same day. The scale here is not hypothetical: OpenAI says people already create more than 3 billion images a week across ChatGPT Images and the GPT-Image models in the API.
But the genuinely interesting move isn't the quality bump. It's that OpenAI split the upgrade in two — and kept the price card identical for both halves. Here's what actually changed, how the two new models differ, and how they stack up against Google's Nano Banana 2 family and ByteDance's Seedream lineup on the three axes that matter: cost, quality, and performance.
The headline upgrades are editing and fidelity. Images 2.5 is better at changing only what you asked for — swap a product, a background, or a line of copy while the subject, composition, and brand treatment stay put. Multi-turn editing got "stamina": earlier edits survive later ones instead of the image quietly degrading every round. Reference-photo fidelity improved too, so faces and products carry through new settings and styles more recognizably. Rounding it out: more natural lighting, richer textures, transparent backgrounds in complex layouts, and more accurate real-world information inside images.
The API grew two new quality tiers — xhigh and max — above the old "high" ceiling, and output now runs to 4K (max edge 3840px, both dimensions multiples of 16, aspect ratios 1:3 to 3:1). In ChatGPT, the launch shipped Sketch (@Sketch — draw a reference directly), templates, comments pinned to images, and shareable prompts.
Safety moved too, in a direction worth knowing about: the system card reports the share of unsafe images reaching users on OpenAI's adversarial set fell to 1.09% for Sunburst and 1.41% for Flare, versus 1.64% for Images 2.0. And in a genuinely surprising provenance twist, 2.5 images now carry an invisible SynthID watermark from Google DeepMind alongside C2PA metadata — OpenAI embedding its biggest rival's watermarking tech.
What didn't change: the token rates. Both new models bill exactly like gpt-image-2 — $8 per million image-input tokens ($2 cached), $30 per million image-output tokens, $5 per million text-input tokens. The upgrade is free per token. What's absent is also telling: no cheaper Batch API pricing for 2.5 at launch (the gpt-image-2 batch rows still exist), and OpenAI published no average price-per-image figure this time.
Honest fine print from OpenAI's own limitation list: complex prompts can take up to two minutes, text rendering is still imperfect on precise placement, characters and brand elements can still drift across separate generations, and repeated edits can still touch details you meant to keep. OpenAI's own advice for pixel-identical regions: composite the approved edit back into the original rather than prompting harder.
OpenAI's split is by workflow, not by model size marketing:
CellCog ran the head-to-head the day of launch — same prompt, all five quality settings, both models, usage reporting on. The results are the cleanest public data we have:
| Measure | Flare | Sunburst |
|---|---|---|
| Image tokens per image | identical at every tier | identical at every tier |
| Price per image | identical | identical |
| Generation time | baseline | ~1.5–2× longer |
| Generation quality | no measured difference | no measured difference |
So the two models produce the same pixels at the same price — Sunburst just takes longer to get there. That may sound backwards until you see where Sunburst earns its keep: in edit-heavy workflows, where CellCog and The Decoder both observed Sunburst consuming more tokens per edit (longer internal reasoning runs), which is precisely the job it was built for.
Meanwhile the quality ladder shifted a rung. CellCog measured tokens per 1024×1024 image directly:
| Quality | Output tokens | Cost per image |
|---|---|---|
| low | 196 | $0.006 |
| medium | 439 | $0.013 |
| high | 1,756 | $0.053 |
| xhigh (new) | 3,122 | $0.094 |
| max (new) | 7,024 | $0.211 |
Read that carefully: 2.5 at "high" spends what GPT-Image-2 spent at "medium," and 2.5 at "max" costs what GPT-Image-2's "high" cost. The new tiers are the top of a shifted ladder, not a surcharge on the old rungs. And there's an opacity wrinkle in ChatGPT itself: the UI gives you no control over which model you're talking to. The Decoder's testing suggests Chat mode mostly routes to the weaker tier while Work mode gets the precision behavior — in the API, at least, you choose explicitly.
Now the fun part. Flare and Sunburst versus Google's Nano Banana 2 Lite and Nano Banana Pro, and ByteDance's Seedream 4.5 and Seedream 5.0 (yes, the user-facing name is Seedream — one of the more common misspellings in AI right now is "seadream").
| Model | Typical cost per image | Pricing model |
|---|---|---|
| GPT-Image-2.5 Flare/Sunburst | $0.006 – $0.211 (1K, tier-dependent) | tokens: $8/M in, $30/M out |
| Nano Banana 2 Lite | $0.034 (1K) | $0.25/M in, $1.50/M out |
| Nano Banana 2 | ~$0.045 (512px) – $0.151 (4K) | resolution-tiered |
| Nano Banana Pro | ~$0.134; ~$0.24 (4K implied) | resolution-tiered |
| Seedream 4.5 | $0.04 | flat per image |
| Seedream 5.0 Pro | $0.045 (≤2.36M px) / $0.09 above | pixel-tiered, first ref image free, extra refs $0.003 |
| Model | Image Edit | Text-to-Image |
|---|---|---|
| GPT-Image-2.5 Sunburst | 1520 (#1) | 1421 (#1) |
| GPT-Image-2.5 Flare | 1491 (#2) | 1399 (#2) |
| GPT-Image-2 | 1461 | 1381 |
| Seedream 5.0 Pro | 1394 | — |
| Nano Banana Pro | 1390 | — |
| Nano Banana 2 | 1387 | — |
| Seedream 4.5 | — | — |
Sunburst and Flare took the top two spots in all three image arenas — including multi-image editing — within a day of launch. The mandatory caveat: both scores are flagged preliminary, built on roughly 3,100 and 2,900 votes respectively against ~78,700 behind GPT-Image-2. Strong early signal, not a settled verdict. (When GPT-Image-2 itself launched, it opened with a +242-point lead over second place; the 2.5 gap over the previous champion, at +30 edit-Elo over GPT-Image-2, is far more modest so far.)
| Model | Latency | Source caveat |
|---|---|---|
| Nano Banana 2 Lite | ~4s (3.37s in one timed test) | fastest of the six by a wide margin |
| Nano Banana 2 | ~10s (9.95s timed) | |
| Nano Banana Pro | ~21s (21.07s timed) | |
| Flare | up to 50% below GPT-Image-2; 2–4× faster per Manus | OpenAI/relative, no absolute seconds published |
| Sunburst | ~1.5–2× Flare | CellCog measurement |
| Seedream 5.0 Pro | 86s @ 1K / 110s @ 2K (independent timing) | one source's first-take average; provider-infrastructure-dependent — some providers report sub-2s |
That Seedream row deserves an asterisk the size of a billboard: measured latency ranges wildly by provider infrastructure, and Google's own screenshot comparisons put Lite at 3.37s against 21.07s for Pro. Latency claims in this market are only as honest as the pipe they were measured through.
The subtle story this release tells: sticker prices no longer tell you what you'll pay. Both OpenAI models bill identical token rates, yet Sunburst routinely costs more per edit in early tests because it reasons longer. Seedream 5.0 Pro prices by pixel area. Nano Banana 2 prices by resolution tier and silently leaks margin through a 15–20% failure rate and ~1.2 retries per image (real cost ≈ $0.067 × 1.2 ≈ $0.08), while Pro's ~3–5% failure rate keeps its effective cost closer to sticker. The models with the flattest pricing cards — Nano Banana 2 Lite at $0.034 and Seedream 4.5 at $0.04 — are suddenly the ones you can predict on a spreadsheet.
And that makes Nano Banana 2 Lite the sleeper of this comparison: within ~19 Elo points of full Nano Banana 2 on generation and ~80 on editing, at half the price and a fifth of the latency. When the second-cheapest model on the board is also the fastest by 3–6×, "cheap and cheerful" stops being an insult.
Routing by workload, from what the data actually supports today:
Two operational notes worth their weight in tokens. First, WaveSpeed's migration advice generalizes: quality-tier names are not fidelity-equivalent across generations — 2.5's "high" is a different product than 2.0's "high" — so re-run your saved trouble prompts (small text, dark scenes, hard light) through both tiers before moving production traffic. Second, the "Draft on Flare, evaluate on Sunburst" pattern is the closest thing this release has to a free lunch: identical pixels at identical prices, with speed as the only trade.
Three open questions will decide how this board looks in a month. Whether Sunburst's +30 preliminary edit-Elo lead holds as votes accumulate past the low thousands. Whether OpenAI ships Batch API pricing for 2.5 — its absence is currently a hidden 50%+ surcharge for anyone with overnight workloads. And whether OpenAI ever documents how ChatGPT routes between the two models, because right now the answer to "which model am I talking to?" in the consumer app is, officially, nobody's saying.
The one-line summary: OpenAI took the quality crown by splitting one upgrade into two price-identical models, Google owns the speed-and-cost floor with a Lite tier that embarrasses models costing twice as much, and ByteDance is quietly selling print-ready 4K at a flat rate while both giants fight over Elo points. There has never been a better week to be paying by the image.
