GPT Image 2 vs GPT Image 1.5: How Much Better Is Object Placement?

What can actually be verified

GPT Image 1.5 has an official OpenAI API model page. OpenAI describes it as its latest image generation model, with better instruction following and adherence to prompts. GPT Image 1 also has an official model page; OpenAI describes it as the previous image generation model and says it accepts text and image inputs and produces image outputs.

OpenAI’s image generation documentation also separates two workflows: Generations, which create images from scratch from a text prompt, and Edits, which modify existing images. That distinction matters. Generating a busy scene from text is not the same task as adding or moving an object inside an existing image.

What is missing from the cited OpenAI API changelog, Models and All models excerpts is a quantitative comparison for GPT Image 2 versus GPT Image 1.5 on complex composition, spatial accuracy or multi-object placement. So the strict answer is: GPT Image 1.5 is documented, and its prompt-following improvement is documented; a numeric GPT Image 2 placement gain is not.

GPT Image 2 claims are not consistent across public sources

The difficulty is not that nobody is talking about GPT Image 2. It is that public sources do not agree.

Mew Design takes a cautious line, saying GPT Image 2 appears to be in testing but is not officially public as a named OpenAI model. getimg.ai similarly says OpenAI has not announced GPT Image 2 and that the public lineup stops at gpt-image-1.5. By contrast, another third-party article claims OpenAI launched ChatGPT Images 2.0, also called gpt-image-2, on April 21, 2026, and cites an Arena score of 1,512.

Third-party reviews have a similar problem. RenovateQR says it tested GPT Image 2 with a standardized set of prompts covering product photography, UI mockups, multilingual signage, portraits and branded content. But the excerpt does not provide a reproducible placement-specific success rate, left-right or front-back accuracy, or a full GPT Image 2 versus GPT Image 1.5 comparison table.

The safest wording is therefore: some third-party sources claim GPT Image 2 exists, is being tested or has launched, but the available material does not provide an official, reproducible, task-specific answer to how much better it is at complex composition and object placement.

Why 1,512 vs 1,241 is not a placement-accuracy result

One third-party page lists GPT Image 2 at 1,512 and GPT Image 1.5 at 1,241, a gap of 271 points. That can support only a narrow statement: one third-party source reports a higher overall score for GPT Image 2.

It should not be rewritten as a 271-point improvement in object placement, or as a percentage improvement in spatial accuracy, for three reasons:

• The excerpt does not say the score specifically measures complex composition, relative position, occlusion or multi-object placement.
• The score is not a pass/fail placement accuracy rate. Without the test set, scoring method, sample size and error categories, it cannot be converted into an accuracy percentage.
• The OpenAI API sources provided do not include an official same-test spatial or object-placement benchmark for GPT Image 2 versus GPT Image 1.5.

For accurate product copy, reporting or internal decision-making, the most you can say from that source is that a third-party page lists GPT Image 2 above GPT Image 1.5. You should not say OpenAI has proved GPT Image 2 is X% more accurate at placing objects.

What a useful object-placement benchmark would need

To answer the question properly, a benchmark would need to be stricter than a gallery of good-looking examples. At minimum, it should include:

1. The same prompts for both models — with model names, versions or snapshots recorded clearly.
2. Spatial-relation prompts — left and right, above and below, in front of and behind, overlapping, edge placement, centred placement, hand-held objects and multi-object relationships.
3. A pass/fail rubric — checking object identity, object count, relative position, occlusion, scale, text placement and any required brand elements.
4. Blind or multi-rater review — to reduce cherry-picking and avoid letting reviewers know which model produced which image.
5. Separate scoring for Generations and Edits — because OpenAI’s own documentation treats creating an image from scratch and modifying an existing image as different workflows.

Without that kind of setup, a single screenshot, a social media comparison or a broad model score cannot answer how much placement accuracy improved.

Practical advice for design, product and marketing teams

Until there is an official or reproducible third-party benchmark, do not treat better GPT Image 2 placement as a quantified fact. A more defensible workflow is:

• Use GPT Image 1.5 as the citeable baseline, because it has an OpenAI API model page and is described as having better instruction following and prompt adherence.
• Build a fixed prompt set around your own high-risk cases: product positions, left-right relationships, foreground-background relationships, UI element order, text placement and brand asset preservation.
• Score placement before aesthetics. A beautiful image that puts the hero product in the wrong place is still a failure for many commercial uses.
• Track Generations and Edits separately, since OpenAI’s documentation distinguishes those workflows.
• Avoid promising clients or stakeholders a specific improvement percentage unless your own test set supports it.

Bottom line

The strongest source-backed conclusion is limited: GPT Image 1.5 has an official OpenAI API model page and is described as improving prompt adherence, while GPT Image 1 is documented as the previous image generation model. What cannot be reliably quantified from the available sources is how much GPT Image 2 improves complex composition or object placement over GPT Image 1.5.

That does not prove GPT Image 2 has made no progress. It means the public evidence reviewed here is not enough to support a number. Third-party sources conflict on GPT Image 2’s status, and a broad 1,512 vs 1,241 score should not be treated as a spatial-accuracy or object-placement benchmark.