Every GPT Image 2 post, filterable by category, sortable by date, searchable by keyword.
Nine months after GPT Image 1.5, the follow-up shows up on LM Arena under a tape-themed codename. Here is what is different, what is the same, and what you can reason about today versus what is still rumor.
A working theory of why text is hard for diffusion models, what the leaked GPT Image 2 benchmarks actually mean, and how to prompt for text today on GPT Image 1.5 so your pipeline survives the upgrade.
Three frontier image models, three strong opinions. A working guide to which one you call for which job, based on what each is actually good at in April 2026.
Three tiers, a 40x price spread, and one decision that drives most of your cost. Here is the math on when low is fine, when medium is obvious, and when high earns its keep.
Two concrete recipes: a SaaS dashboard screenshot and a beverage can label. Both produce legible, accurate text if you specify the right details.
At 2048px, prompts that worked fine at 1024 start to break. Three patterns that hold up, plus a worked example that actually tests the resolution limit.