GPT Image 1.5: review an upgrade without losing your baseline

The release

OpenAI released gpt-image-1.5 and chatgpt-image-latest on December 16, 2025.

Keep the old evaluation set

The fastest way to lose a useful comparison is to improve every prompt while changing the model. First run an unchanged set of representative requests, then experiment with prompt revisions separately. Save both sets of results.

Include easy and difficult examples. Product silhouettes, packaging text, reference-image identity, and composition constraints can expose different weaknesses. A single impressive output does not establish a reliable default.

Distinguish a chosen version from a moving alias

An application should record the exact identifier it requests. If you use an alias intended to follow a newer release, plan recurring acceptance checks around that choice. If you choose a specific model version, track its retirement notices instead.

Neither choice removes the need to inspect results. The appropriate selection depends on how much change your production workflow can absorb without a new review.

Make the rollout reversible

Keep prompt templates, request settings, and application configuration versioned together. Start with a subset of jobs and inspect cost per accepted image before increasing traffic. Check the current retirement schedule before adopting this model today; this article explains the historical release, not a current default recommendation.