OpenAI GPT Image 2.5

Go Bananas supports both API models released on 8 September 2026:

Go Bananas model ID	OpenAI API model	Best suited to
openai-gpt-image-2.5-flare	gpt-image-2.5-flare	Fast everyday generation
openai-gpt-image-2.5-sunburst	gpt-image-2.5-sunburst	Precise generation and editing

Both are optional pro-tier models. Existing defaults and tenant allowlists are preserved.

Enable and use

Add the desired model IDs to the tenant's allowed models in Admin and configure its OpenAI key pool. Existing OpenAI provider credentials are reused. Provider account access and billing still apply. The composer then offers both models, and the API and MCP tools accept their IDs for generation, editing, references and character preferences.

{
  "prompt": "A clean banana logo on a transparent background",
  "model_id": "openai-gpt-image-2.5-flare",
  "quality": "xhigh",
  "background": "transparent",
  "output_format": "png",
  "size": "1024x1024"
}

CLI users can select the model explicitly on the provider lane:

gobananas generate "A clean banana logo" --provider --model openai-gpt-image-2.5-flare

Controls

• Quality: auto, low, medium, high, xhigh, max. The last two require GPT Image 2.5.
• Output: PNG, JPEG or WebP. Transparent backgrounds require PNG or WebP. JPEG/WebP support compression from 0 to 100.
• Size: auto or custom dimensions. Both edges must be multiples of 16 and at most 3840 pixels; aspect ratio must be within 1:3 to 3:1; total pixels must be 655,360–8,294,400. OpenAI describes resolutions above 2560×1440 as experimental.
• Reference generation uses the Images edit endpoint. Masked edits are temporarily unavailable for Flare and Sunburst because regional-edit quality verification has not passed. Requests with mask_image_id are rejected before provider dispatch; use GPT Image 2 for masked edits.

Go Bananas retains a 240-second provider timeout for these models. Website Continue Editing submits a background job and follows its execution status, so closing the original request does not end the job. Synchronous API edits still wait for the result. The timeout is an application limit, not an OpenAI latency guarantee; uncertain jobs must be inspected rather than resubmitted.

Cost and subscription behaviour

The two models use the same token rates, but can consume different numbers of tokens per image. check_quota reports their monetary estimate as unknown / variable, with an explanation, instead of applying GPT Image 2's older per-image heuristics. Returned token usage is retained. Provider billing is authoritative.

Selecting Flare or Sunburst uses the tenant's OpenAI API connection. Go Bananas' separate Codex subscription flow is not a guarantee of either API model or of these exact quality controls, so it remains separate. No tenant is automatically switched from subscription generation to a billed API model.

Official references

Checked 9 September 2026:

• OpenAI API changelog
• GPT Image 2.5 Flare
• GPT Image 2.5 Sunburst
• Image generation guide