GPT Image 2 vs GPT Image 1.5: are size, ratio and reference-image limits looser?

Bottom line: do not claim looser limits yet

It would be premature to write that GPT Image 2 supports arbitrary dimensions, a wider set of aspect ratios, or more reference images than GPT Image 1.5. The sources here support a narrower statement: whatever other model differences may exist, the cited materials do not quantify how much, if at all, GPT Image 2 changes API freedom for custom size, aspect ratio or reference-image inputs.

In other words, “GPT Image 2 is more flexible” should be treated as a hypothesis until an official comparison or reproducible tests establish it.

What the OpenAI documentation actually establishes

OpenAI’s GPT Image 1.5 model page describes GPT Image 1.5 as an image-generation model with better instruction following and adherence to prompts. A GPT Image 2 model page is also available in the OpenAI API documentation.

The OpenAI Images API reference includes generation and editing endpoints. The Python and TypeScript Images references also mention GPT image models and generated-image fields such as size, background, output format and quality.

That is enough to confirm that image generation, image editing and size-related fields exist in the API documentation. It is not enough to determine the exact supported dimensions for each model, whether either model accepts arbitrary width and height values, which aspect ratios are officially supported, or whether reference-image input rules changed between GPT Image 1.5 and GPT Image 2.

The missing official details include:

• accepted size values, width and height ranges, or maximum and minimum dimensions;
• whether dimensions are arbitrary or limited to fixed presets;
• a model-specific aspect-ratio list;
• reference-image limits, including count, format, file size and multi-image rules.

Three limit claims checked

Why third-party docs are useful but not decisive

Third-party platforms can help show how GPT image models are packaged in real products. Leonardo.AI’s GPT Image-1.5 REST API example includes fields such as width, height, seed and guidances.image_reference; fal’s documentation lists image_size, background and quality; WaveSpeedAI’s edit documentation includes size, quality, input_fidelity and output_format; and a Higgsfield guide lists aspect-ratio choices such as 1:1, 2:3 and 3:2.

Those details may be relevant if you are building on those specific services. But they are not the same as OpenAI’s native API limits. A third-party wrapper can rename parameters, set its own defaults, restrict values, expose only selected UI choices or add platform-specific behavior. For that reason, a width/height field or aspect-ratio menu on one platform should not be treated as the full capability ceiling of the underlying OpenAI model.

What would settle the question

To say with confidence that GPT Image 2 is more flexible than GPT Image 1.5, at least one of the following would be needed:

1. an OpenAI table listing accepted size values, maximums, minimums or fixed presets for both models;
2. official guidance stating whether each model supports arbitrary dimensions or only fixed aspect ratios;
3. official rules for image edit or reference-image inputs, including count, format, file size and multi-image behavior;
4. a changelog, migration guide or model comparison that explicitly states which limits changed;
5. reproducible tests in the same API environment showing which size, ratio and reference-image combinations are accepted or rejected for each model.

Without that evidence, “GPT Image 2 has looser image limits” remains an unverified product claim rather than a documented fact.

Practical advice for product and engineering teams

If you are deciding whether to move from GPT Image 1.5 to GPT Image 2, do not base the decision solely on assumed gains in image-size freedom, aspect-ratio support or reference-image input. Evaluate output quality, prompt following, latency, cost and API behavior as separate questions. For API limits specifically, rely on official documentation and repeatable tests, not on screenshots or wrappers alone.

A practical test plan would vary one factor at a time: requested size, aspect ratio, number of reference images and file format. Record successful responses, rejections and error messages for both models. Only when the official parameter table or test results clearly show a difference should external copy say that GPT Image 2 supports more sizes, more ratios or looser reference-image rules.

Safest public wording

The most defensible public statement is this: based on the verifiable sources cited here, OpenAI documentation shows GPT Image 1.5 and GPT Image 2 model pages, along with image generation and image editing references. However, the cited materials do not provide a complete official comparison of custom size, aspect-ratio or reference-image input limits. Therefore, the available evidence does not confirm a measurable increase in parameter freedom for GPT Image 2 over GPT Image 1.5.