GPT Image 2.5: What Is Confirmed, What Is Rumor, and What to Build On Today

Search interest in GPT Image 2.5 has jumped in the past week, and a fair amount of what is being written about it reads as if the model already exists. It does not, at least not under that name. This post separates what can be verified from what is inferred, and then answers the practical question for anyone shipping image features: what should you build on right now?

Short version: build on the GPT Image 2 API today, keep the model ID in config, and treat the 2.5 label as a placeholder until OpenAI publishes one.

What is actually confirmed

Everything in this section can be checked against a primary source as of September 5, 2026.

OpenAI has not announced GPT Image 2.5. The OpenAI developer documentation still lists gpt-image-2 as the current image model. There is no gpt-image-2.5 model page, no entry in the API changelog, and no model ID in the official SDKs. OpenAI has never published a string containing “2.5” for its image line; the name was attached by commentators.

GPT Image 2 is the model you can call. It launched on April 21, 2026, roughly two weeks after an unlabeled checkpoint was spotted in Arena battles in early April. It supports text-to-image and image editing, up to 16 reference images per edit request, and character-level accuracy for non-Latin scripts. We covered the practical side in GPT Image 2 pricing in 2026 and GPT Image 2 rate limits.

Two anonymous checkpoints have appeared in Arena. The first, mona-lisa-1, showed up in blind image battles around August 9. We wrote about it at the time in Mona-Lisa-1: Is OpenAI testing a new image model?, including reports that its outputs carried OpenAI provenance signals. A second checkpoint, luna-lisa-alpha, followed around August 19 to 22. Neither carries official attribution.

A hidden “major upgrade” notice exists in ChatGPT. On September 3, a disabled in-app modal reading “Image creation got a major upgrade” was found in ChatGPT’s front-end build, promising higher-quality results, faster generation, and smarter creative tools. A matching hero image had been uploaded to OpenAI’s servers on September 1. The gate was off for all users. This is shipped launch copy sitting behind a feature flag, which is the strongest signal so far that something is staged.

The GPT-6 Astra launch on September 3 did not mention an image model. If an image update was meant to ride along with that release, it did not.

What is inferred, and how confident you should be

The name. “GPT Image 2.5” is a guess by analogy with gpt-image-1.5, which shipped in December 2025 as an interim release between 1 and 2. OpenAI could just as easily call the next model GPT Image 3, fold it into a ChatGPT-only update, or delay it. Treat the number as a label for “the next OpenAI image model,” nothing more.

The improvements. Testers comparing luna-lisa-alpha against mona-lisa-1 describe more natural skin and fabric texture, better behavior under harsh lighting, and a clear reduction in the grain that affected dark areas in the earlier checkpoint. Some also report finer typographic control at small sizes and more consistent results across iterative edits. These are anecdotes from a small number of accounts on blind battles. There are no reproducible benchmarks, and the grain claim in particular is disputed.

The timing. GPT Image 2 went from Arena sighting to public launch in about 15 days. Applying that gap to the late-August checkpoint points at early September, which lines up with the September 3 modal. OpenAI’s next scheduled event is DevDay on September 29. Both dates are inferences from precedent, not a schedule.

Pricing and API availability. Nothing has been published. GPT Image 1.5 and GPT Image 2 both reached the API on launch day, so an API route is likely if the model is real, but the token rates are unknown.

What this means if you are building an image feature

The mistake to avoid is waiting. Teams that paused image work in March to “wait for GPT Image 2” lost six weeks and then had to integrate under time pressure anyway. The next model, whatever it is called, will almost certainly land on the same Images API shape: a prompt, optional reference images, a size and quality setting, and an image back. Code written against that shape today carries over.

Concretely:

1. Ship on GPT Image 2 now. On WaveSpeedAI the GPT Image 2 API exposes openai/gpt-image-2/text-to-image and openai/gpt-image-2/edit. Text-to-image starts at $0.01 per image at low quality, $0.06 at medium quality for 1K output, and $0.22 at high quality, with 2K and 4K tiers above that. One API key covers these endpoints and the rest of the catalog, so you are not maintaining a separate OpenAI account, rate-limit envelope, and billing relationship for one model.

2. Keep the model ID in configuration, not in code. When a new OpenAI image model is added, it appears as a new model ID on the same platform. Swapping openai/gpt-image-2/text-to-image for its successor should be a config change plus a quality review, not a rewrite.

3. Build your evaluation set now. If the reported improvements are real, they will show up on exactly the prompts that give GPT Image 2 trouble today: small text, dark scenes, skin under hard light, and multi-step edits. Save those prompts and their current outputs. When the new model arrives you will have a same-prompt comparison on day one instead of relying on other people’s Arena screenshots.

4. Do not price a product on unannounced rates. If your unit economics only work at a hoped-for lower price, they do not work yet. GPT Image 2 pricing is public and stable; plan against it.

What we will do when it ships

WaveSpeedAI adds new OpenAI image models as endpoints alongside the existing ones, with the same request format and per-image pricing, and keeps the previous model available so you can migrate on your own schedule. When OpenAI publishes the next image model under any name, we will update this post with the model ID, the pricing, and a same-prompt comparison against GPT Image 2.

Until then, the model to build on is the one you can call: GPT Image 2 on WaveSpeedAI.