Agent Upstart

Cool things people are building with agents — and practical help getting the most out of them.

Playbooks · Coding

GPT-Image-2.6: Flare or Sunburst?

Written November 10, 2026

OpenAI's Images 2.6 ships two API models at the same token rates: Flare for speed, Sunburst for detail. Pick by the job, not by the name.

The problem

A builder calling the image API now has two model names where there was one choice to make, and the names do not say which job each one fits. Flare is the fast one, Sunburst is the careful one, and the price per token is the same for both — so the choice is about time against detail, not about cost per token.

What OpenAI says each one is for

OpenAI's November 4 announcement puts it plainly: Flare generates images in as little as six seconds when quality is set to low, and also improves layout quality for posters, product imagery, and other professional visuals. Sunburst brings higher-quality generation and greater consistency across edits when quality is set to medium or high. Images 2.6 as a whole generates images 50% faster than before, by OpenAI's own measure. Both variants are available in the API, alongside GIF creation, SVG export, image markup, and a new complexity parameter that controls how much planning the model does before it renders.

The rule that follows from that is simple. Use Flare when you will throw half the outputs away: drafts, social graphics, poster concepts, product variations, quick edit rounds. Use Sunburst when the output has to hold up on close inspection or stay consistent across a series of edits: brand assets, product shots that must match across a campaign, character sheets, scenes with many constraints.

What independent tests measured

Artificial Analysis ran both variants through its blind text-to-image and editing leaderboards, and the two models took the top two positions on both. Sunburst ranked first in seven of the ten text-to-image use cases tracked, and Flare ranked first in the consumer and marketing categories. The speed gap in those tests is large: Flare's median text-to-image latency was about 30 seconds, with edits at about 32 seconds, against Sunburst at about 135 seconds per image and about 120 seconds per edit.

VariantText-to-image latencyEdit latencyBest for
FlareAbout 30 secondsAbout 32 secondsDrafts, iteration, batch generation
SunburstAbout 135 secondsAbout 120 secondsDetail, typography, identity-preserving edits

Both variants share one token rate card: $5 per million text input tokens, $8 per million image input tokens, and $30 per million image output tokens. At maximum quality, Artificial Analysis calculates about $0.21 for a 1024 by 1024 image, or about $211 per thousand images. The per-image cost is the same for both variants at the same settings; Sunburst costs more in wall-clock time, not in tokens.

How to call them

Set the model explicitly in the API call: gpt-image-2.6-flare or gpt-image-2.6-sunburst. There is no variant selector in ChatGPT itself, so an API call that names neither gets whatever the default is, not the variant you meant. Both variants take the same parameters, including the complexity setting: low complexity for quick edits, medium or high for scenes with many elements or constraints. Both default to high complexity, which is easy to miss and adds latency to a call that did not need it.

Limits

Related playbooks

Sources

Back to: Coding playbooks hub · Also: Hermes Agent 0.23.0: paid MCP spend limits