GPT ഇമേജ് 2

OpenAI's newest image model, succeeding GPT Image 1.5 with better instruction following and text rendering at a lower output-token price. It generates and edits images from text and reference images, and accepts the same size, quality and fidelity controls as its predecessor.

OpenAI launched this model a year after its first ChatGPT Images release, and the framing changed with it: the argument is no longer that images generated by a model can be beautiful, it is that they can be finished. The lab describes the shift as moving image generation from rendering to strategic design - the target is a visual you can put in front of someone without editing it first.

The specific claim is fidelity to what you asked for. OpenAI says the model can conceptualise more sophisticated images and then actually deliver them, following instructions, preserving requested details and rendering the elements that usually break image models: small text, iconography, interface elements, dense compositions and subtle stylistic constraints, at up to 2K resolution through the API. Output shape is unusually flexible for this class of model - aspect ratios as wide as 3:1 and as tall as 1:3, requested in the prompt or picked from presets to regenerate an existing image in new dimensions.

Non-Latin text is where the post is most useful for readers here, because OpenAI states the previous limitation outright: its earlier image models were consistent in English and other Latin-script languages and less precise beyond that, particularly when the text was complex or dense. It reports significant gains in non-Latin rendering, naming Japanese, Korean, Chinese, Hindi and Bengali, and is careful to say the goal is not correct glyphs but language that reads coherently, in outputs where the script is part of the design rather than a label pasted on top. Style fidelity moves in the same direction: better capture of what makes a photograph read as a photograph, including its small flaws, and of cinematic stills, pixel art and manga.

This is also OpenAI's first image model with thinking. When a thinking or pro model is selected, it can search the web for current information, reason through the structure of an image before generating it, check its own output, and produce several distinct images from a single prompt - up to eight in one pass with character and object continuity between them. The lab pairs that with a December 2025 knowledge cutoff, which it says matters most for explainers, educational graphics and visual summaries where being right counts as much as looking right. Editing is first-class: masked inpainting, reference images and transparent backgrounds are all supported.

OpenAI publishes the failures too. It says the model still struggles with anything requiring a complete and coherent physical world model - origami instructions, puzzles like a Rubik's Cube, and details that must appear correctly on hidden, angled or reversed surfaces - and that very dense or repetitive detail, of the fine-grains-of-sand kind, tests its limits. Labels and diagrams may still need checking for accuracy, especially where they depend on precise arrows or part names. API output above 2K is in beta and the lab warns it can be inconsistent.