OpenAI released Images 2.6 on November 4, 2026 across ChatGPT, Codex, and the API. The direction: faster, more precise.

This article breaks down the features and API changes, and how developers should actually use Images 2.6.

The numbers

The core figures from OpenAI's announcement:

Metric Value
Generation speed 50% faster than previous
Weekly generation volume 4 billion (ChatGPT + Codex + API combined)
Daily average 500 million+
GPT-Image-2.6 Flare (low quality) 6 seconds minimum
GPT-Image-2.6 Sunburst (medium/high) Higher quality, stronger editing consistency

4 billion images per week signals that image generation has moved well beyond experimentation. 500 million+ images daily means image generation is now part of production workflows, not a research exercise.

The API offers two models:

GPT-Image-2.6 Flare:
  - 6-second generation at low quality
  - Improved layout for posters, product imagery, and
    professional visuals
  - Suitable for rapid iteration
  - Lower API pricing

GPT-Image-2.6 Sunburst:
  - Medium/high quality setting
  - Higher detail and precision
  - Stronger character/product/brand consistency across edits
  - Suitable for complex scenes with multiple elements
  - Higher API pricing, but quality difference is significant

The difference between Flare and Sunburst isn't just a quality setting. Sunburst maintains character, product, and brand details better across editing rounds. The difference becomes clear in workflows that modify images multiple times or change specific parts of previous outputs.

Developers using the API should design workflows that account for the pricing difference. A hybrid approach — Flare for rapid iteration, Sunburst for final output — is efficient.

GIF sprite sheets: animation from a single prompt

The most visible new feature in Images 2.6 is GIF generation. A single prompt produces a seamless looping animation sprite sheet.

Looking at OpenAI's example prompts reveals the structure:

"Create a single animation sprite sheet for a seamless
 looping GIF. A super-cute chibi adult office worker sits
 at a desk watching a computer monitor with a clearly
 plunging stock-market line chart. The sequence: hopeful
 expression, notices the chart falling, eyes widen, face
 crumples, large cartoon tears stream down both cheeks,
 then quietly wipes tears while still staring at the
 falling chart. Oversized round head, tiny body, big
 expressive eyes, soft rounded features, messy fluffy
 hair, white shirt and loosened tie. Exactly 6 sequential
 frames arranged in a precise 3-column by 2-row grid,
 read left-to-right then top-to-bottom. Every panel has
 identical dimensions. Fixed front three-quarter camera,
 same character, same desk, monitor, outfit, and cozy
 miniature office background. Clear incremental facial
 and tear motion; sixth frame transitions smoothly toward
 the first. Polished kawaii chibi 3D toy-like style..."

Breaking down the prompt structure:

GIF prompt structure:
  1. Task specification: "single animation sprite sheet"
  2. Sequence description: "The sequence: A → B → C → D → E → F"
  3. Fixed elements: character, background, camera angle
  4. Grid structure: "Exactly N frames, M-column by N-row grid"
  5. Style: "kawaii chibi 3D toy-like style"
  6. Constraints: "No captions, no words, no logos..."

The critical detail is explicit grid specification — "Exactly 6 sequential frames arranged in a precise 3-column by 2-row grid." GIF sprite sheets require clear frame counts and layout order. Dimensions must be identical across all panels, and the final frame must transition smoothly back to the first.

OpenAI's second example prompt reverses the sequence — "notices the chart rising, eyes widen with delight, jumps up cheering" — same character, same background, different sequence. This shows that changing only the sequence description produces different animations from the same character design.

Picsart's Davit Avoyan evaluated the results as "more accurate outputs, better character consistency, faster generation." Figma Weave's Itay Schiff called it "speed and creative control together." Town, a startup, explored character animation through the API.

Four new control points in the API

Images 2.6 adds four new control elements to the API:

1. Image markup (already available in ChatGPT)
   → Now in API. Specify exact change locations

2. SVG export
   → Ready for websites, apps, and print immediately

3. Complexity parameter
   → Controls how much planning the model performs
   → Lower: suitable for quick edits
   → Higher: scenes with multiple elements or constraints

4. Image quality parameter
   → Controls visual detail and fidelity
   → low / medium / high

Image markup arriving in the API means instructions go beyond "edit this" to "change this specific element's color to X." Previously, you could regenerate the entire image or specify regions. Now you can point to specific elements via markup.

SVG export makes designs immediately usable across websites, apps, and print. Converting raster images to vector SVG eliminates scaling issues — particularly useful for brand identity workflows.

The complexity parameter is an interesting design choice. Most editing tasks recommend low complexity; scenes with many elements or constraints recommend medium or high. This lets developers control "how much the model thinks."

Recommended usage patterns:
  - Most editing: low complexity + required quality
  - Complex scenes / multiple elements: medium or high complexity
  - Rapid iteration: Flare + low quality
  - Final output: Sunburst + medium/high quality

Real-world usage scenarios

Three scenarios prove useful for teams adopting Images 2.6:

Scenario 1: Brand mascot animation

Generate a 6-frame sprite sheet from a single prompt, then export as SVG for web embedding. Sunburst handles character consistency better, so generate the final version with Sunburst and iterate with Flare. Accounting for API pricing, a strategy of 8 Flare iterations followed by 2 Sunburst generations is efficient.

Scenario 2: Product image editing

Use image markup to specify exact parts of a product (color, logo, texture) for modification. Test multiple versions quickly with low complexity, then generate the final version with Sunburst. Since markup instructions may be rejected by safety filters, implement fallback logic.

Scenario 3: Social media campaigns

As Davit Avoyan evaluated, combining "fast generation + accurate results + character consistency" reduces the time from single idea to full social campaign. OpenArt's Coco Mao noted "GIFs are the native language of social media," and Images 2.6 "makes every creator fluent in it."

How do the safeguards work?

OpenAI published safety evaluations and safeguards for Images 2.6 separately in the system card. Image generation safety primarily focuses on illegal content generation, impersonation of specific individuals, and adult content.

Developers using the API should monitor not only output quality but also how safety filters operate. Specifically, when using image markup, implement fallback logic for cases where markup instructions are rejected by safety filters.

Safety filter behavior is an important practical consideration. When image markup specifies exact change locations, the safety system may interpret certain markup instructions as potentially violating content policies — even when the underlying intent is legitimate. Teams should design their workflows to handle safety filter rejections gracefully, with fallback mechanisms that either adjust the markup or fall back to full-image regeneration. This is particularly relevant for product imagery workflows where specific color or branding changes might trigger safety filters.

The system card also covers content policy around realistic impersonation, which becomes more relevant as image generation quality improves. Developers generating images of real people — employees, public figures, or customers — should implement review processes to verify that generated images don't inadvertently create misleading or unauthorized representations.

Bottom line: consistency beats speed

The question Images 2.6 raises isn't "how fast can you generate images?" It's "how well can you maintain character or product details across multiple editing rounds?"

50% faster generation is a metric. 4 billion images per week signals we've reached production scale. GIF sprite sheets, image markup, SVG export, and the complexity parameter — these four control points give developers more precise control over image workflows.

The real differentiator isn't speed — it's consistency. The ability to maintain character, product, and brand details across multiple editing rounds has more impact on actual projects. A Flare/Sunburst hybrid strategy that accounts for API pricing is the core of an efficient workflow.

One thing worth considering: the meaning of "4 billion images per week." 500 million+ daily means image generation is no longer experimental — it's production. This suggests API design should focus on "editing and iteration" rather than pure "generation." Images 2.6's four control points — GIF, markup, SVG, complexity — all center on iteration and editing. Town's exploration of character animation through the API fits the same pattern.

Read further

Sources

Learn this in practice

If you want to control image generation through validated workflows with iteration and consistency rather than intuition, the process of building CLAUDE.md · Skills · Hooks · Subagents · MCP in Claude Code connects directly to this article.