The October Image Model Price War: GPT Image 2.5 Splits Into Two Tiers

OpenAI's image lineup changed shape in September, and the effect on cost has taken a few weeks to show up in public comparisons. ChatGPT Images 2.5 shipped on 8 September as two API models rather than one: GPT-Image-2.5 Flare and GPT-Image-2.5 Sunburst.

The pitch from OpenAI is straightforward. Up to 50 percent lower latency, 2K output, better multi-turn editing, and better subject preservation across edits. Flare is the fast default. Sunburst is the slower, precision tier.

The surprise is the price column. Both Flare and Sunburst start at 4 credits per generation, with the 4K tier at 7. The previous GPT Image 2 starts at 9. In other words, the newer OpenAI model is also the cheaper OpenAI model. That flips the default for fresh work to 2.5.

How to split Flare from Sunburst

The practical rule for a production team is simple. Use Flare for drafts, batches, and anything where the first acceptable frame wins. Use Sunburst when an edit has to preserve a face, a label, or a layout exactly, and you can afford to wait.

The case for staying on GPT Image 2 is narrower than it looks. If a campaign has already been approved on the older model, switching mid-campaign changes skin tones, type weight, and lighting in ways a client will notice before you do. The advice from practitioners is to finish the current campaign where it is and start the next one on 2.5.

The rest of the field did not stand still

Tencent released Hy Image 3.5 Preview on 22 September. One model handles both generation and multi-round editing. It goes up to 4096 by 4096, accepts 20 reference images, and is listed at around $0.024 per image at 1K to 2K on Tencent's international API.

The word "Preview" carries real weight here. There are no weights and no technical report. A preview with neither can change behavior from one week to the next. Test it on the API if the 20-reference stack matters to your workflow. Do not build a recurring template on it yet.

Alibaba's Qwen team released Qwen-Image-2.1 on 20 September: 7B open weights, native 2K, RGBA output with a real alpha channel, and up to 10 reference images. The licence, though, is now research-only with no commercial use. That means an agency or a brand cannot ship 2.1 outputs in paid work on the open-weights terms. If you pulled the weights expecting the old terms, read the licence before a client file leaves the machine.

Recraft posted V4.1 Flash on 23 September and called it "the fastest image model on the market." No resolution, no latency figure, no price. One source, no numbers. Treat it as an announcement rather than a spec sheet.

Why the licence terms changed the calculus

The quiet story of September is that the two open alternatives moved in opposite directions on licensing, and both moved in ways that matter more than benchmark rank.

Qwen-Image-2.1 kept the technical promise that made the family popular. Seven billion parameters is small enough to run on a studio workstation, native 2K avoids an upscaling step, and RGBA output with a real alpha channel means a generated subject arrives with transparency already baked in rather than requiring a cutout pass. Ten reference images in a single generation is a genuinely useful ceiling for product work, where a consistent subject across a set of shots is the hard part.

Then the licence closed the commercial door. Research-only, no commercial use. For an agency, that means the open-weights terms cannot cover paid client work. The weights still run, and a designer can still use them to explore, but the output cannot ship on those terms. Practitioners who had built a pipeline around earlier Qwen releases, which were more permissive, are now reading the fine print on a file they already downloaded.

Tencent's Hy Image 3.5 Preview carries the opposite problem. It is available on the API, so commercial use is a matter of paying per image rather than a licence question. But "Preview" means no weights and no technical report, which means the behaviour can shift between one week and the next. Building a recurring template on it, such as a fixed product shot with a fixed lighting setup, is a bet that the model underneath will not move. The safer play is to test it on specific jobs, keep the results, and revisit when the preview turns into a release.

The practical case for keeping two models warm

The reasonable conclusion for a studio is to keep one fast tier and one precision tier configured, and to know which job each handles, rather than to settle on a single winner.

A fast tier, which as of October means GPT Image 2.5 Flare or a comparable model, covers ideation, mood boards, draft frames, and high-volume variation. Speed and cost dominate, and the output is a starting point rather than a deliverable.

A precision tier handles edits where a face, a label, or a layout has to survive untouched. This is where Sunburst's slower path pays for itself, and where a cheaper fast model's drift on skin tone or type weight becomes visible to a client.

The open local tier is a third slot, useful when a workflow has to run without sending assets to an external API. Qwen-Image-2.1's RGBA output fits that slot well, as long as the job is not commercial under the open-weights terms.

Three configured tiers sounds like overhead. In practice it is the opposite. The team that knows which model handles which job stops re-litigating the choice on every brief, and it stops discovering mid-campaign that the model it switched to renders a brand's signature colour slightly differently.

What the leaderboard says, and what it does not

The Artificial Analysis leaderboards, checked on 6 October, put OpenAI in the top three places on both the text-to-image and image-edit boards. That is a clearer lead than the company held in August.

Rank | Model | Text-to-image | Image-edit --- | --- | --- | --- 1 | GPT-Image-2.5 Sunburst | 1425 | 1524 2 | GPT-Image-2.5 Flare | 1398 | 1481 3 | GPT Image 2 | 1383 | 1462 4 | MAI-Image-2.6 (Microsoft) | 1333 | 1428 5 | Nano Banana 2.1 (Google) | 1328 | 1428

The more useful finding for cost is further down. Nano Banana Pro, at 2K, sits at number 16, below Nano Banana 2.1 and at roughly a quarter of Pro's 1K price. Google's own model card shows 2.1 beating the older Pro model on some internal editing evaluations. Independent reviewers note that Pro still produces better images in some hands-on tests.

That gap between rank and reality is the part worth holding onto. Arena rank says little about a model's specialist strengths. FLUX 3 Image, Midjourney, and Ideogram 4.0 at its top setting are either absent from the boards or ranked low, yet each has a niche where it wins. Ideogram 4.0 Quality, for instance, sits at number 20.

What this means for a studio budget

Three things changed in September that affect how a team should plan.

The default model for new work is no longer the premium tier. GPT Image 2.5 Flare at 4 credits does most of what GPT Image 2 did at 9, and does it faster.

Open weights got more useful and more restricted at the same time. Qwen-Image-2.1's 7B weights and RGBA output are genuinely useful for local pipelines, but the research-only licence closes the commercial door that earlier Qwen releases left open. Read licences before building on them, not after.

And the leaderboard is a starting point, not a verdict. A model that ranks 16th on an aggregate board can still beat the number one on the specific job in front of you. Benchmark on your own prompts before switching a production line over.

The month's story is that the cheapest tier of frontier image generation got noticeably cheaper while the licence terms around the open alternatives got noticeably tighter. Both changes push teams to re-examine defaults they set months ago.