Two Models, One Price, Half the Wait: Inside OpenAI's ChatGPT Images 2.5

On September 8, 2026, OpenAI pushed a new image generation release into ChatGPT and its API, and the numbers behind it are bigger than the release itself. ChatGPT Images 2.5 arrived for a user base that, by OpenAI’s own count, now creates more than 3 billion images every week across ChatGPT Images and the GPT-Image API models. Whatever OpenAI ships into that surface is, by default, one of the most-used creative tools on the planet — and this update is built around exactly the friction that shows up at that scale: waiting, re-rolling, and edits that break more than they fix.

The headline claims are concrete. Images 2.5 reduces generation latency by up to 50% compared with Images 2.0 (the April release behind the gpt-image-2 API), sharpens detail with more natural lighting and texture, and — the part production teams will care about most — makes editing dramatically more precise, including the ability to change one element of a picture while leaving the rest of the scene genuinely untouched.

The two-model API split: Flare and Sunburst

In the API, the release lands as two distinct model IDs: gpt-image-2.5-flare and gpt-image-2.5-sunburst (snapshots dated 2026-09-08). The split is OpenAI’s answer to a tension every image-heavy product lives with — speed versus control — and it resolves it with an unusual pricing decision: both models bill at identical token rates.

Flare is positioned as “the default choice for most applications.” It delivers higher-quality images than GPT-Image-2 at 50% lower latency, and OpenAI’s usage guidance points it at creator and social content, product experiences, visual search, rapid prototyping, and high-volume generation. If you run anything on gpt-image-2 today, Flare is the migration target.

Sunburst is the precision instrument: “an extra level of precision for detailed creative work with longer generation times.” It is aimed at production-ready campaign creative and polished product imagery, and it is the model meant for the edits endpoint, where reference images and masks must be respected. The premium you pay for Sunburst is latency, not dollars — both models use the same token budget per image at a given quality and size.

Early real-world measurements suggest the latency story is even better than the headline. An HN user whose AI UI design tool has generated roughly 50,000 images on gpt-image-2 reported average latency of around 104 seconds on the old model dropping to 35–40 seconds on 2.5 — a bigger drop than “up to 50%” implies, at least on that workload. Community quality opinions are more mixed, with some users arguing OpenAI edits still lose fine detail next to competitors. The crowd-voted Arena leaderboards (updated September 7) tell a friendlier story: on text-to-image, Sunburst sits at 1421 and Flare at 1399, ahead of gpt-image-2 at 1381, Microsoft’s mai-image-2.6 at 1331, and Nano Banana 2 at 1261 — with Sunburst’s lead widening further on the image-editing board.

Editing that finally respects the rest of the image

The capability OpenAI is pushing hardest is targeted editing. Images 2.5 follows editing instructions more closely, even across multiple rounds of edits, and can make changes to just one piece of a picture — a product, a background, a piece of copy — while preserving the surrounding scene. Reference-led workflows also get more reliable, with greater image fidelity keeping each variation more closely tied to the original source.

In production, this is the difference between revising an asset and rebuilding it. A team that needs a product shot with three different backgrounds no longer regenerates the product three times and prays for consistency; the model is explicitly built to hold the subject steady while the context changes. OpenAI also says the new model is better at understanding complex visual instructions and translating them into coherent results, getting more images right on the first try before edits are even needed.

Four new tools inside ChatGPT

The consumer side of the release adds four features, none of which ship in the API:

• Sketch — draw directly in ChatGPT and use the drawing as a visual guide for generation. Invoke it with @Sketch in the composer or at chatgpt.com/sketch.
• Templates — presets for popular creative formats: posters, merch, flyers, product photos. (Early reports say templates aren’t available in Work mode.)
• Comments on images — click a spot on a generated image, leave a comment, and the next edit targets that area. It’s a lightweight regional-instruction mechanism hiding inside a UI affordance.
• Share the prompt — shared images now carry their prompt, so recipients can remix them with their own photos.

Availability is across all ChatGPT tiers — including free — on desktop, mobile, and web, with per-plan generation caps still applying.

The quiet repricing: same rates, new ladder

Day-one coverage got one number wrong: several outlets reported that API pricing had doubled. It hadn’t. Per-token rates match gpt-image-2 exactly — $5 per million text input tokens, $8 per million image input tokens, $30 per million image output tokens. What changed is the quality ladder, which decides how many tokens each image consumes.

Images 2.5 expands the ladder from three quality levels to five: low, medium, high, xhigh, and max (plus auto). Per OpenAI’s calculator at 1024x1024, low costs about $0.006 (196 output tokens), while max runs 7,024 tokens — about $0.21 per image. The practical effect: 2.5’s high (1,756 tokens) is the old gpt-image-2 medium budget, and 2.5’s max is the old high. Migrating developers who keep quality: "high" and swap the model ID get roughly 4x cheaper images; those who want the old high-quality output budget move to max and pay effectively the old bill. OpenAI’s own caveat deserves quoting: equal token rates don’t mean equal cost per image — token consumption can differ by model and quality setting, and reference images on the edits endpoint bill as image input with no published per-image count.

Safety and provenance

The system card published alongside the release gives the transparency the image arena rarely gets. Sunburst presents unsafe content in 1.09% of cases and blocks 21.9%; Flare presents 1.41% and blocks 19.2% — both ahead of the Images 2.0 baseline (1.64% unsafe presented, 23.1% blocked). Provenance is unchanged: every image carries C2PA metadata plus an invisible watermark, which the system card identifies as Google DeepMind’s SynthID.

What’s missing

The launch has gaps worth noting. There are no published per-tier rate limits (the model pages point to the org limits page). Batch support is unconfirmed — the pricing page’s Batch tab still lists only gpt-image-2, so Batch-dependent workloads should stay put for now. The thinking parameter documented for gpt-image-2 is absent from the 2.5 docs. And no gpt-image-2 deprecation date has been posted; the April snapshot remains live.

The read

Images 2.5 is not a moonshot — it is infrastructure maintenance for the largest image-generation surface on the internet, executed with unusual care about where the pain actually lives. The two-model split at one price is a genuinely interesting pattern: it lets OpenAI A/B its own tradeoff curve with users self-selecting into speed or precision, without anyone having to run a price-sensitivity experiment on them. The quality-ladder relabeling does quietly change what your invoice means, which is exactly the kind of change that surfaces three weeks later in a cloud bill review rather than a launch-day headline. For the 3-billion-images-a-week baseline, a 50% latency cut and edits that leave the rest of the image alone aren’t incremental niceties — they’re the difference between an AI toy and production tooling.