OpenAI's GPT-Image-2.6 Flare and Sunburst Top Every Image Leaderboard

OpenAI’s Images 2.6 family has taken the top two positions on the Artificial Analysis leaderboards for text-to-image generation and image editing. The API offers two variants: GPT-Image-2.6 Flare, optimized for lower latency, and GPT-Image-2.6 Sunburst, tuned for higher fidelity. Both surpassed the previous 2.5 generation in independent blind comparisons.

Artificial Analysis derives its quality rankings from head-to-head votes, while measuring latency and cost separately. Its results show distinct performance profiles: Flare nearly halves generation time, while Sunburst spends more time producing finer detail, stronger typography, and more consistent edits.

Flare accelerates as Sunburst slows

Flare’s median text-to-image latency fell from about 55 seconds to 30 seconds in the Artificial Analysis tests. Editing time dropped from 58 seconds to 32 seconds. Those measurements used maximum quality, 1024 × 1024 PNG output, and the API’s default high-complexity setting.

Sunburst required about 135 seconds for each text-to-image generation and 120 seconds for an edit, exceeding its predecessor’s latency in paired tests. It ranked first on AA-Image-T2I v2.1, with its largest gains appearing in reasoning, text rendering, lighting, and material realism. OpenAI also reports stronger image consistency across edits at medium and high quality.

Sunburst leads structure-heavy prompts

Sunburst ranked first in seven of the ten text-to-image use cases tracked by Artificial Analysis. Its strongest categories depend on spatial structure, legible typography, or multi-step reasoning:

• Animation and gaming, including concept art, character designs, and comics
• Architecture and real estate, including interior visualizations, exterior scenes, and floor plans
• UI and UX design, including app, web, and in-car interface mockups
• Reasoning prompts involving mathematics, spatial relationships, logic, or mixed concepts
• Text rendering, including long passages, small type, and artistic lettering

Flare ranked first in the consumer and marketing and advertising categories, where it posted its largest gains over 2.5 Flare. Human anatomy remains weaker for both variants, particularly hands, faces, and dynamic poses.

A 2.6 model led six of the seven editing categories. Sunburst ranked first for enhancement and restoration, text or symbol changes, object-level edits, and identity preservation. Flare led reasoning-based edits and composition or reframing. Microsoft’s MAI-Image-2.7-Preview retained first place for scene and style editing.

Pricing holds at $30 per million tokens

Both variants cost $30 per million image-output tokens, matching the 2.5 generation. At maximum quality, Artificial Analysis calculates an approximate cost of $0.21 for a 1024 × 1024 image, or $211 per thousand images. Actual API spending will vary with dimensions, quality settings, output format, and token use.

Artificial Analysis measured the earlier GPT-Image-2.5 Sunburst at about 6.3 times the per-image cost of Nano Banana 2.1. The OpenAI model led by 47 Elo points in text-to-image testing and 48 points in editing. Elo reflects relative preference in head-to-head votes; the point difference is not a percentage improvement.

Four controls expand the API workflow

Images 2.6 adds controls and output formats that support longer planning, lightweight animation, vector delivery, and more precise edit instructions:

• Complexity: Controls how much planning the model performs before rendering. Artificial Analysis used the default high setting.
• Animated GIF output: Creates a looping GIF from a prompt or source image, with a maximum of 16 frames.
• SVG export: Produces vector output that remains sharp when resized.
• Markup-guided editing: Accepts drawn annotations that identify where and how an image should change.

Choose by latency budget

Flare fits product interfaces, rapid prompt iteration, marketing pipelines, and batch workloads where a two-minute response would impede throughput. Its roughly 30-second median latency also makes it the practical option for applications that display results during an active user session.

Sunburst fits final assets that require accurate text, detailed materials, structured layouts, or repeated edits to the same subject. OpenAI’s release notes identify character consistency across edits as a training target, and the independent editing results support that claim.

Teams migrating from 2.5 can test the new model IDs without an increase in image-output token pricing. Before deployment, developers should revalidate prompts, latency limits, output-token consumption, anatomy-sensitive use cases, and any workflow that depends on consistent results across multiple edits.