OpenAI ships GPT-Image-2.5 as two API models, Flare and Sunburst

GPT-Image-2.5 reaches OpenAI's API as two models, fast Flare and slower, more precise Sunburst, at the same per-token rates GPT Image 2 charged.

Key points

• OpenAI's model pages for Flare and Sunburst list the IDs gpt-image-2.5-flare and gpt-image-2.5-sunburst, each resolving to a snapshot dated 2026-09-08. Inputs are text and images, output is images, on the v1/images/generations and v1/images/edits endpoints, and both pages mark v1/batch as unsupported. The edits endpoint also takes a mask for inpainting, per OpenAI's image guide.
• Identical rates for both, per those pages: $5 per 1M text input tokens ($1.25 cached), $8 per 1M image input tokens ($2 cached) and $30 per 1M image output tokens, so a cache hit takes 75% off input. Those match GPT Image 2's standard rates on OpenAI's pricing page, which lists the older model's batch rates at exactly half.
• Both model pages list six quality settings (low, medium, high, xhigh, max, auto), and xhigh and max are new in this generation. The Decoder puts a 1024x1024 image at roughly $0.006 on low, $0.053 on high and $0.21 on max, the last from about 7,024 output tokens.
• OpenAI's claims, not independent tests: Flare beats GPT Image 2 on image quality while halving latency, and Sunburst adds precision for demanding edit workflows at the cost of slower generation. You reach the models through the API, ChatGPT or Codex, and there is no batch discount yet.
• On Arena's text-to-image board Sunburst sits first at 1421 and Flare second at 1399, ahead of GPT Image 2 at 1381, but both new entries are flagged preliminary with about 3,100 and 2,900 votes versus roughly 78,700 for the older model.
• The same model pages set rate limits from 100K TPM and 5 images per minute at tier 1 to 8M TPM and 250 at tier 5. Outputs carry C2PA metadata plus an invisible SynthID watermark. The Decoder found no word in OpenAI's announcement or docs on how ChatGPT chooses between Flare and Sunburst; API callers set the model themselves.

“Our most capable model for image generation and editing.” — OpenAI, gpt-image-2.5-sunburst model page

Why it matters

Anyone calling gpt-image-2 outside the Batch API can switch by changing the model string with no per-token price change; batch users would give up the older model's half-price rate, since neither new model accepts batch jobs. The routing call is yours: Flare for thumbnails, mockups and anything a client is waiting on, Sunburst for hero assets and multi-round edits where one wrong element forces a redo. Budget per image rather than per token, though: in early tests cited by The Decoder, Sunburst tended to bill more per image than Flare, likely because it spends more tokens on each one, and a single max-quality 1024x1024 image can run about $0.21.

This is an AI-written summary of the reporting credited above and of the OpenAI documentation linked in the text, read and edited by Nicholas before publishing. The facts belong to those sources and the quote to OpenAI; the wording is ours. Read the original.