OpenAI Images 2.6 in the API: Flare vs Sunburst, complexity control, and GIF output

OpenAI released Images 2.6 on November 4, 2026, and the developer-relevant part is easy to miss under the GIF demos: the API now exposes two image models, gpt-image-2.6-flare and gpt-image-2.6-sunburst, plus a new complexity request field that controls how much planning the model does before rendering. If you generate or edit images in a pipeline, the practical change is a speed-versus-precision dial you can set per call, animated GIF and SVG output from the Image API, and token-based pricing that is already listed. Most of the performance numbers are still OpenAI's own claims, so treat them as starting hypotheses, not measurements.

What actually shipped

From OpenAI's announcement and the API documentation:

• Two API models. gpt-image-2.6-flare is positioned for fast iteration; OpenAI says that with quality set to low it can generate an image in as little as six seconds. gpt-image-2.6-sunburst is positioned for detailed work and edit consistency at medium or high quality. The announcement says Images 2.6 generates 50% faster than before, without naming the baseline.
• A new complexity parameter. Per the image generation guide, complexity takes low, medium, or high, defaults to high when omitted, and controls how much the model plans before rendering. Lower complexity suits focused edits like changing a background or a sign; higher complexity helps with dense text, complex layouts, and object relationships. OpenAI warns that higher complexity can increase latency and introduce unwanted changes. Adjusting complexity has no separate fee. Complexity is distinct from quality, which still controls rendering fidelity.
• Markup-guided edits in the API. Image markup, previously a ChatGPT feature, is now available through the API so callers can indicate exactly where a change should apply.
• SVG output and animated GIF output through the Image API, both listed in the guide's GPT Image 2.6 section.
• Availability: Images 2.6 is live in ChatGPT and Codex on all subscription tiers, and both 2.6 models are in the API now.

One documentation gap worth knowing: the image generation guide has a GPT Image 2.6 section, but the code examples throughout the guide still use the 2.5 model names (gpt-image-2.5-flare and gpt-image-2.5-sunburst). If you copy an example, swap the model string.

Pricing, from the pricing page

The OpenAI pricing page already lists both 2.6 models, and the numbers are identical for Flare and Sunburst:

• Image input: $8.00 per million tokens.
• Image output: $30.00 per million tokens.
• Cached input: $2.00 per million tokens.
• Text input: $5.00 per million tokens; cached text input $1.25; no text output price is listed for these models.

Billing is token-based, so the cost of a generated image depends on the size, quality, and complexity you request. The announcement's "as little as six seconds" figure applies to Flare at low quality; Sunburst at medium or high quality will cost more latency and, at higher token counts, more money. If you are budgeting a pipeline, run your own prompts and read the usage field rather than assuming a per-image flat rate.

For comparison, the pricing page lists the same rates for the 2.5 Flare and Sunburst rows, so on list price there is no premium for the 2.6 generation. gpt-image-1.5 shows a slightly higher image output price ($32.00 per million) in the same table.

The claims I would not take at face value

Vendor launch posts optimize for excitement. Three claims here need your own testing:

1. "50% faster than before." No baseline model, settings, or measurement method is named in the announcement.

2. "As little as six seconds" for Flare. That is a best case at low quality, not a median, and it says nothing about edit rounds where consistency matters more.

3. Consistency claims. Sunburst "brings higher-quality image generation and even greater image consistency across edits" per the announcement. Consistency across edit rounds is exactly the property that breaks image pipelines, so test it on your assets: same character or product, five edit rounds, count the drift.

The announcement also cites usage scale (4 billion images per week across ChatGPT, Codex, and the API) and quotes launch partners OpenArt and Figma Weave. Those are marketing context, not engineering signal.

What I would check

These are my untested suggestions, not findings:

• A latency ladder on your own prompts: Flare at low/medium/high quality crossed with low/medium/high complexity, then Sunburst the same way. The announcement recommends low complexity for most editing tasks; verify where quality breaks on your content.
• GIF output constraints: frame count, resolution, duration limits, and file size. The guide confirms animated GIF output through the Image API but the announcement gives no limits; check the API response metadata on real calls.
• Whether markup-guided edits reduce retry loops in your edit pipeline compared with text-only edit prompts. That is where the cost saving would actually come from.
• The safety_identifier field. The guide notes the Images API user parameter is deprecated in favor of safety_identifier (a privacy-preserving hash, 128 characters or fewer). If your code still sends user, plan the migration; existing requests remain supported for now.

What I would not conclude

• That Flare replaces Sunburst or vice versa. They are positioned as different points on the speed-precision curve and are priced identically, so the choice is per call, not per project.
• That six-second generation holds at production quality settings. It is explicitly a low-quality figure.
• That GIF output replaces a video model. The demo prompts in the announcement describe six-frame sprite sheets turned into loops, which is a different thing from general video generation.
• That the 2.5 models are deprecated. The guide still documents them and its examples still use them; no deprecation date appears in the materials I read.

Frequently asked questions

What are the model IDs for Images 2.6 in the API?

gpt-image-2.6-flare and gpt-image-2.6-sunburst, per the OpenAI pricing page. Flare targets fast iteration, Sunburst targets detailed work and edit consistency.

What is the new complexity parameter?

A request field (low, medium, high; default high) that controls how much planning the model does before rendering, per the image generation guide. Lower complexity suits focused edits, higher complexity helps dense layouts. It carries no separate fee and is independent of the quality setting.

How much does GPT Image 2.6 cost?

Token-based: image input $8.00, image output $30.00, cached input $2.00 per million tokens, text input $5.00 per million, per the pricing page. The same rates are listed for the 2.5 Flare and Sunburst models. Per-image cost depends on your settings; there is no flat per-image price in the materials I read.

Can the API generate animated GIFs and SVGs now?

Yes. The image generation guide lists animated GIF output and SVG output through the Image API for the GPT Image 2.6 models. Limits on frames, duration, or size are not stated in the announcement.

Is Images 2.6 available outside the API?

Yes, per OpenAI: in ChatGPT and Codex across all subscription tiers on desktop, mobile, and web, starting November 4, 2026.

What happened to the user parameter?

The guide says it is deprecated in favor of safety_identifier, a stable privacy-preserving hash of a username or email, capped at 128 characters. Existing user requests remain supported, and a nonempty safety_identifier takes precedence.

Should I switch production workloads from 2.5 to 2.6?

Not on the strength of the announcement alone. The list prices match, but the speed and consistency numbers are vendor claims without published methodology. Run your own prompts through both generations and compare latency, cost per accepted image, and edit consistency before cutting over.

Sources

• Introducing Images 2.6 | OpenAI - read November 6, 2026. Announcement claims: 50% faster generation, Flare at as little as six seconds on low quality, markup in the API, SVG export, GIFs, availability in ChatGPT and Codex, partner quotes, and the 4 billion images per week figure.
• OpenAI API pricing - read November 6, 2026. Model IDs and per-million-token prices for gpt-image-2.6-flare and gpt-image-2.6-sunburst, plus the 2.5 and 1.5 comparison rows.
• Image generation guide | OpenAI Developers - read November 6, 2026. Complexity field values and default, no separate fee, markup-guided edits, SVG and animated GIF output via the Image API, safety_identifier guidance, and the note that examples still use 2.5 model names.
• The Code newsletter, November 5, 2026 issue - where I first saw the release; its summary matches the announcement.

Not read: the Images 2.6 system card (linked from the announcement), the image generation cost calculator, and any independent benchmark or hands-on test. All speed and quality figures above are OpenAI's claims unless attributed to the pricing or guide pages.