Moving from GPT Image 2 to GPT Image 2.5: what to test before switching

A GPT Image 2 to GPT Image 2.5 migration starts with a choice between Flare and Sunburst, followed by an evaluation of the images your workflow actually needs. Keep the GPT Image 2 configuration as a baseline, use explicit output settings, and compare quality, latency, and cost per accepted result before changing production defaults.

This checklist is based on OpenAI’s image generation documentation and prompting guidance, checked September 9, 2026. It is an evaluation procedure, not a report of paid API benchmarks.

1. Choose a candidate for your workload

OpenAI positions Flare as a smaller, faster model with quality comparable to GPT Image 2. Sunburst emphasizes higher quality and precise editing.

The Flare vs Sunburst guide explains the selection tradeoffs. A model’s positioning is a hypothesis to test against your application, not a measured result for your traffic.

2. Preserve the current baseline

Before editing a default model string, save representative prompts, source images, model ID, output dimensions, quality, format, and any post-processing settings. Include examples your current integration handles well and examples that regularly need retries.

Define acceptance before looking at the new results. For a product cutout, that might mean a readable label, unchanged proportions, clean edges, and a usable alpha channel. For a poster, exact text and layout may matter more than additional texture.

Use the same input set across the candidates. If practical, hide model names while reviewing outputs so that a preference for the newer release does not become the evaluation criterion.

3. Change the model explicitly, then audit settings

The direct Images API continues to use images.generate() for generation and images.edit() for reference edits. Choose one of these exact IDs:

# Existing baseline
model = "gpt-image-2"

# Evaluate separately, with the same prompts and output settings
model = "gpt-image-2.5-flare"
# or
model = "gpt-image-2.5-sunburst"

Keep separate configuration entries for the baseline and each candidate rather than overwriting the only working configuration. The API walkthrough includes documentation-checked generation and editing examples.

Check these settings during migration:

• Quality: 2.5 supports auto, low, medium, high, xhigh, and max. Compare an explicit shared setting first; evaluate the new higher settings separately.
• Dimensions: custom edges must be multiples of 16, no edge may exceed 3,840 pixels, the aspect ratio must stay within 3:1, and total pixels must be 655,360–8,294,400. OpenAI marks resolutions above 2560x1440 as experimental.
• Transparency: request background="transparent" with PNG or WebP, and preserve the returned alpha channel through subsequent processing.
• Compression: apply output_compression only to JPEG or WebP.
• Legacy parameters: do not copy fields from an older model’s reference without checking their current support. The examples here do not require input_fidelity.

For Responses integrations, change the model inside the image_generation tool definition. The outer language model is a separate selection. For gateways, verify the provider’s route and parameter support before applying OpenAI examples unchanged.

4. Compare accepted-image cost, not token rates alone

Flare and Sunburst have the same published token rates. That does not establish equal token use, equal latency, or equal cost for an image your team can publish.

Track two different values:

Request cost = charges for that request's billable usage

Cost per accepted image = total evaluation image charges
                          / number of accepted output images

Count rejected outputs and retries in the numerator. If ten generations cost a hypothetical $1 and only five meet the requirements, the generation cost per accepted image is $0.20. This is an arithmetic example, not a GPT Image 2.5 benchmark or price quote. Human review and post-processing costs would be additional.

OpenAI now provides a GPT Image 2.5 output-token estimator. Use the correct model family and explicit settings; its image-output estimate excludes other usage categories. For completed requests, use returned usage and the applicable provider rates. Responses also adds the outer language model’s usage. See the pricing breakdown for those distinctions.

5. Move traffic only after the output checks pass

Start with a limited production category whose acceptance rules are clear. Keep the working GPT Image 2 configuration available while comparing request failures, latency, acceptance, and billing. Expand after the new configuration meets the requirements you set.

Inspect the final artifact that users receive, including resizing and compression. A correct transparent PNG can lose its alpha channel later in the pipeline; a legible label can become unreadable after downscaling. The evaluation should cover the delivered image, not only the raw API response.

A successful migration may use Flare for drafts and Sunburst for selected edits, or retain GPT Image 2 for workloads that already perform well. The useful outcome is a configuration supported by your results, with costs and failure handling you understand.