GPT Image 2

What it is

GPT Image 2 is the anchor of OpenAI’s image line. It launched in April 2026 to the top of the public leaderboards and stayed there through the summer, holding first on the arena.ai reading as late as 29 August 2026 and on the Artificial Analysis image arena on the 6 August reading. On the 15 September 2026 reading of the arena.ai boards (both dated 7 September; arena.ai is the address the old lmarena.ai now redirects to), it sits third on both, and the two models above it are OpenAI’s own. On text-to-image it records an Elo of 1381, behind GPT Image 2.5 Sunburst on 1421 and GPT Image 2.5 Flare on 1399, both marked preliminary; on image edit it records 1461, behind Sunburst on 1520 and Flare on 1491. OpenAI’s line holds the top three seats on both boards at once. The boards re-score continuously, so read those positions as dated readings rather than as today’s. These arenas rank models by pitting their outputs against one another in blind human comparisons and converting the results into an Elo score, the same rating method used in chess, so a top three occupied by a single line reflects a durable preference among the people casting votes rather than a single favourable benchmark.

The 2.5 pair

OpenAI released GPT Image 2.5 Sunburst and GPT Image 2.5 Flare on 8 September 2026, recorded in the company’s API changelog as a pair for image generation and editing. Sunburst is the variant OpenAI points at workflows where editing precision comes first; Flare is the one it describes as built for “fast, high-quality everyday image generation”. Both are billed at “GPT Image 2 token rates”, so the step up in score carries no change in price. A week after release the pair already heads both arena.ai boards on the 15 September reading: Sunburst first on text-to-image at 1421 and on image edit at 1520, Flare second at 1399 and 1491, with the text-to-image scores still marked preliminary while votes accumulate. GPT Image 2 sits directly beneath them on both boards.

Where it is strong

The strengths behind that score fall into three areas. GPT Image 2 is strong on prompt adherence, which means it renders what an instruction actually asks for rather than a loose approximation of it. It is strong on text rendering, the long-standing weak point of image generation, where earlier systems tended to produce garbled or invented lettering inside signs, labels and captions. And it is strong on photorealism, producing pictures that read as photographs rather than illustrations. The model also supports flexible output resolutions up to 4K, high enough for print and large-format work without a separate upscaling pass.

Those three qualities tend to travel together in practice. A model that follows the prompt precisely and renders clean text is far easier to direct, because it takes fewer attempts to reach a usable result, and fewer attempts means a lower cost per finished image. For editorial, marketing and product teams, that dependability is often worth more than a marginal gain in raw fidelity, and it goes a long way towards explaining why the model held first place on the public boards from its launch until OpenAI’s own 2.5 pair arrived above it.

How to access it

For distribution, the model is bundled into ChatGPT subscription plans and billed separately through the API, either per image or per token depending on how a team chooses to integrate it. That arrangement makes GPT Image 2 the path of least resistance for anyone already working inside the OpenAI stack. There is no new supplier to approve, and the same credentials and billing relationship that cover the company’s language models extend to image generation. For a development team weighing its options, that convenience often settles the question before quality is even discussed. Readers comparing systems can start at our image models hub, which sets each one against the others.

The lineage

The lineage is worth setting out, because a good deal of search traffic still arrives under the older names. GPT Image 2 follows GPT Image 1 and GPT Image 1.5, both of which are still cited in older comparison roundups, and the 2.5 pair now extends the line above it. Anyone looking for those earlier releases is almost always after the current generation, and tracing the progression from one release to the next explains how the line came to hold the top three seats it occupies today.

What to watch

The position is easy to describe and hard to unseat. The public arenas have become the closest thing the industry has to an agreed scoreboard, and on the 15 September reading OpenAI holds the top three seats on both at once, with MAI-Image-2.6 and Grok Imagine trading fourth and fifth places beneath them. The same company also leads much of the market in large language models, so a team that adopts GPT Image 2 is often consolidating text and image work with a single provider. What to watch is whether the preliminary Sunburst and Flare scores settle where they started as votes accumulate, and whether any rival breaks into the top three as the next wave of models arrives.