When OpenAI released Images 2.6 on November 4, 2026, the API shipped two tiers at once: GPT‑Image‑2.6 Flare (gpt-image-2.6-flare) and GPT‑Image‑2.6 Sunburst (gpt-image-2.6-sunburst). Per the product write-up and model pages, Flare targets fast-iteration generation and editing; Sunburst targets workflows that lean on detailed composition and precise edits. Selection debates often blur “speed claimed in marketing” with “quality a business can accept.” A more durable approach is to freeze prompts and resolution, then evaluate in the order the official docs describe—not to default-route on unverified third-party slogans.

👉 Try the GPT Image 2.6 API on PoYo: per-image pricing and online playground.

Official positioning

According to OpenAI, when quality is low, Flare can generate an image in about six seconds; Flare also improves layout quality for posters, product imagery, and similar professional visuals. Sunburst comes into play when more detail and precision are required: OpenAI states that at quality of medium or high, it can deliver higher-quality generation and keep stronger consistency across edits. Those timing and quality statements come from OpenAI’s announcement. They are not third-party benchmarks and do not establish an absolute ranking across models.

The image prompting guide’s starting advice is engineering-oriented: if an existing GPT Image 2 workflow already meets quality, or you are on GPT Image 2.5 Flare, test gpt-image-2.6-flare first; if GPT Image 2 falls short on complex tasks, or you are on 2.5 Sunburst, test gpt-image-2.6-sunburst first. If Sunburst already meets the bar, compare Flare with the same prompts, references, size, and quality settings before deciding to step down. The migration checklist also stresses saving baseline samples, covering hard edits and precise text, and rolling out by workflow in batches.

Scenario boundaries

Prefer evaluating Flare first when you need short-cycle sketching and concept exploration; fast poster, ecommerce hero, or layout-style visuals; or latency-sensitive pipelines that can ship at low/medium quality. OpenAI explicitly ties Flare to interactive image generation and editing—suitable when multi-turn prompt trials are part of the workflow and you trade for turnaround within an acceptable quality line.

Prefer evaluating Sunburst first when composition is complex, dense text or object relationships are demanding; multi-turn edits must keep character, product, or brand detail consistent; or final assets have low tolerance for texture, small-type, and layout drift. The prompting guide calls Sunburst the higher-requirement base model for workflows that are stricter on detail, layout, or edit accuracy. The announcement also notes edit consistency under medium or high quality as one of Sunburst’s selling points.

Both tiers support quality (auto/low/medium/high/xhigh/max) and complexity (low/medium/high, defaulting to high when omitted). Official guidance: freeze quality when comparing complexity; if layout is correct but small text is hard to read, compare quality tiers at the same complexity. Adjusting complexity has no separate charge, but higher complexity may add latency and unwanted changes. For most edit tasks, OpenAI recommends low complexity, then pairing with the quality tier the workflow needs.

Pricing, endpoints, and integration notes

Per the Flare and Sunburst model pages, list prices verifiable here are identical: text input $5 per million tokens (cached input $1.25), image input $8 per million tokens (cached $2), image output $30 per million tokens. Selection should therefore rest on quality and latency, not on an assumption that “the faster tier is cheaper.” Both endpoints support v1/images/generations and v1/images/edits; model pages mark Responses as unsupported, so do not copy 2.5 Responses tool fields into production. Organizations may need Organization Verification before calling GPT Image.

The prompting guide also reminds readers that the same quality label does not imply identical visual quality or response time across models—measure on your own load. If a region must remain pixel-identical, docs recommend compositing approved edits back onto the original rather than relying on prompt constraints alone. Partners such as Figma Weave are quoted in the announcement for speed plus creative control in reusable workflows; those are customer quotes for scenario reference, not technical specifications.

Bottom line: Flare and Sunburst are a speed–precision split shipped on the same day, not mutually exclusive “high/low marketing names.” Comparing on a fixed eval set and changing one parameter at a time before choosing a default model matches the docs better than arguing which is “stronger”; track failure rates, retries, and accepted per-image cost before scaling.