Some tools you stop noticing. You open them, type a line, and a poster with clean lettering comes back, or a diagram whose labels read as real words. For many people who make pictures with AI, GPT Image 2 became that kind of tool this year. On 10 November 2026, OpenAI gave it an end date.
The entry on OpenAI’s deprecations page is short. The gpt-image-2 model and its only snapshot, gpt-image-2-2026-04-21, are deprecated and will be removed from the API on 10 February 2027. The recommended replacements are GPT-Image-2.6 Sunburst and GPT-Image-2.6 Flare. That is three months’ notice for a model whose snapshot is dated 21 April 2026. Counted from that date to the shutdown, it will have lasted 295 days.
What exactly ends on 10 February
The notice concerns the API, the interface developers use to build a model into their own apps and services. It names two model IDs and nothing else. It does not mention ChatGPT. The people directly affected are developers who call gpt-image-2 in their code, and everyone who uses a design tool, plugin or automation built on top of it. Once the model is removed, anything still pointing at those IDs has to be moved to another model first.
Until then the model keeps running. Its documentation page now carries the label “Deprecated”, but image generation, image editing and batch jobs are still listed as supported.
Why this model earned a goodbye
When it arrived on 21 April 2026, the launch post on OpenAI’s developer forum called it the company’s most capable image generation model yet and listed what was new: resolutions up to 2K, more aspect ratios, improved multilingual text rendering, and stronger results for structured pictures such as diagrams, infographics, charts, posters and comics. The same post said it had reached first place on every Image Arena leaderboard, with a lead of 242 points in text-to-image.
Read that list the way a designer would. Where does the eye go first? On a poster, to the headline. For years that was exactly where AI images fell apart: letters melted and words turned into ornament. A model that could set a readable line of type inside the picture changed what you could hand over without retouching. That is the detail many people grew attached to, and it is why a short deprecation notice reads like more than housekeeping.
It also had a symbolic role. When OpenAI removed DALL·E 2 and DALL·E 3 from the API on 12 May 2026, GPT Image 2 was one of the models developers were told to move to. Six months later, it has an end date of its own.
The pace behind the decision
The deprecations page shows how quickly this model family turns over. GPT Image 1 was scheduled for shutdown on 23 October 2026. GPT Image 1.5, GPT Image 1 mini and the chatgpt-image-latest alias go on 1 December 2026. GPT Image 2 follows on 10 February 2027. For the December and February retirements, OpenAI recommends the same destination: the two 2.6 models.
What you get in exchange is concrete. According to OpenAI’s image generation guide, Sunburst is for detailed visual work and Flare is for fast iteration. Both come with a complexity setting in three levels, edits guided by markup drawn onto the image, SVG output and animated GIF output. The guide describes these features for the newer models.
My reading, and it is an interpretation that OpenAI does not state: keeping several generations of image models running side by side costs money and attention, and a company that ships a new generation every few months has little reason to maintain the old ones. For OpenAI that is tidy. For you it means a model you depend on can no longer be treated as a fixed tool. It is closer to a season.
What to do before the date
If you build on the API, three things are worth doing while both generations are still available.
First, find every place where gpt-image-2 or the dated snapshot is written into your code, your automations or the settings of a third-party tool. Second, collect the prompts that matter to you, together with the images they produced, and run the same prompts on Sunburst and Flare. A prompt tuned on one model does not produce the same picture on another, and it is better to see the difference in November than on 11 February. Third, check the cost. The token prices OpenAI lists for Sunburst, $8 per million image input tokens and $30 per million image output tokens, are the same figures the launch post gave for GPT Image 2. Whether a single picture costs the same depends on how many tokens each model uses, so measure it on your own images.
What stays, and what you cannot get back
The pictures you made with GPT Image 2 and saved stay exactly as they are. A retired model does not take its output with it. What disappears is the possibility of going back and asking it for one more. If there is a style, a character or a layout that this model gave you in a way you liked, the next three months are the time to make what you still want from it, and to learn how to ask its successors for the same thing.
It will have lasted 295 days. In that time, readable text inside an AI image stopped being a surprise, and that is a fair sign that a tool mattered. Goodbye, GPT Image 2.
