AI agents · OpenClaw · self-hosting · automation

Quick Answer

GPT-Image-2.5 Flare vs Sunburst: Which API Model to Use

Published:

The short answer

GPT-Image-2.5 FlareGPT-Image-2.5 Sunburst
Model IDgpt-image-2.5-flaregpt-image-2.5-sunburst
RoleFast default (“small model”)Premium quality (“base model”)
Speed vs GPT-Image-2Up to 50% lower latency (Manus: 2–4x faster)Longer generation times
Quality vs GPT-Image-2Comparable or betterHigher
PricingSame per-token rate as GPT-Image-2Identical to Flare
Quality tierslow · medium · high · xhigh · max · autoSame six
Max resolution3840×2160, arbitrary sizes3840×2160, arbitrary sizes
Transparent backgroundsNativeNative
Reference imagesUp to 16Up to 16
Unsafe-generation rate (OpenAI internal)1.41%1.09%
Best forSocial content, prototyping, visual search, high-volume generationCampaign creative, product imagery, multi-turn precision edits

Default to Flare. It is faster, costs the same per token, and beats the previous generation on quality. Switch to Sunburst for the final pass on assets where edit precision across many turns matters more than seconds of latency.

What shipped on September 8, 2026

OpenAI released ChatGPT Images 2.5 on September 8, 2026, and with it two API models rather than one. In OpenAI’s words, Flare “brings the same improvements in quality, editing, and speed to the API and is the default choice for most applications, delivering higher-quality images than GPT-Image-2 at 50% lower latency,” while Sunburst “is built for premium visual workflows that benefit from tighter control across edits.” The previous model, GPT-Image-2, launched April 21, 2026 and held #1 on the Artificial Analysis text-to-image arena (1,339 Elo) through the summer; 2.5 has not yet accumulated enough votes for a rating.

Both models share the improvements OpenAI lists for Images 2.5:

  • Fidelity — sharper detail, more natural lighting, richer textures, better preservation of a subject’s identity when restyling from a reference photo.
  • Precision editing — change a single element (product, background, copy) while keeping subject, composition and brand treatment intact.
  • Multi-turn consistency — later edits build on earlier ones without degrading quality.
  • Layout and style intelligence — better at complex briefs, real-world information in images, transparent backgrounds and holding a requested visual direction as instructions get specific.

The six quality tiers

GPT-Image-2 had four quality settings; 2.5 has six, adding xhigh and max above high:

TierWhat you getUse it for
lowFastest, fewest output tokensThumbnails, previews, rapid iteration
mediumBalancedMost production use
highMore detail, more tokensDense text, diagrams, detailed product shots
xhigh (new)Extended detailPrint assets, fine textures
max (new)Maximum detail and tokensWhen quality is the only priority
auto (default)Model choosesGeneral use

Cost scales with tier because billing is per output token. At 1024×1024 the range runs from under a cent at low to roughly $0.21 at max; the step from high to xhigh adds about 78%, and xhigh to max about 125%. At 3840×2160, medium is about $0.026 and max about $0.40. OpenAI is explicit that a higher tier “does not guarantee a better result for every prompt” — test at medium/high and only climb when a specific quality requirement is unmet.

Where Flare and Sunburst actually differ

Because pricing and feature surface are identical, the decision is about latency versus control:

  • Flare is the “small model.” Its win is speed: OpenAI claims 50% lower latency than GPT-Image-2; early partner Manus measured 2–4x. For feeds, previews, visual search and anything users wait on interactively, Flare is the obvious choice, and its quality is already above the previous flagship.
  • Sunburst is the “base model.” Its win is tighter control across edits — the third, fourth and fifth turn of a product-shot refinement stay on-brief. Generation takes longer, and at the same token price that mostly means you spend time, not money. It also posts the lower unsafe-generation rate (1.09% vs 1.41%, both below the 1.64% baseline).

A sensible pipeline is Flare for exploration and volume, Sunburst for the deliverable — the same asset, the same prompt history, one model swap at the end.

Migration from GPT-Image-2

OpenAI’s migration guidance in the image-prompting docs is two-branched:

  1. GPT-Image-2 already meets your quality bar? Start with Flare and check whether the latency drop improves the workflow.
  2. GPT-Image-2 falls short on a complex case? Start with Sunburst to establish whether its higher quality closes the gap, then see if Flare is good enough for the rest.

Things that are new regardless of which you pick: transparent backgrounds are fully supported (no longer preview) with cleaner alpha on hair, glass and object edges; arbitrary resolutions up to 3840×2160; up to 16 reference images; streaming partial images; multi-turn editing through the Responses API; and C2PA metadata plus invisible watermarking on every output.

Watch-outs at launch: the OpenAI price calculator did not yet support 2.5, and per-image token counts for the new xhigh/max tiers were undocumented, so budget from measured usage rather than GPT-Image-2 estimates.

How the pair compares with rivals

Against Google’s Nano Banana Pro (Gemini 3 Pro Image) and Nano Banana 2 (Gemini 3.1 Flash Image), the GPT-Image-2.5 pair is the only option with native transparency, arbitrary sizes and streaming; Google counters with true 4096×4096 output at about $0.24 per image and stronger web-grounded factual text. Midjourney V8.2 has no API at all. Open-weight FLUX.2 is the fallback for teams that must self-host.

Sources