OpenAI Images 2.6: Flare, Sunburst, GIFs and API Pricing

OpenAI released Images 2.6 on November 4, 2026, bringing faster generation, GIF creation and additional editing controls to its image tools. The launch announcement lists global availability in ChatGPT and Codex across subscription tiers on web, desktop and mobile, alongside two API models: GPT-Image-2.6 Flare and Sunburst.

For teams producing product visuals, illustrations or campaign assets, the interesting question is how quickly an image can survive review and reach the place it will be used. Generation speed helps, but so do predictable revisions and an output format that does not require rebuilding the result by hand.

Flare versus Sunburst

OpenAI says Images 2.6 generates images 50% faster than before. Its most specific speed claim is for Flare at low quality, which can produce an image in as little as six seconds. Sunburst is positioned for greater detail and consistency across edits at medium or high quality.

These are OpenAI's claims, not timings measured for this article. A six-second low-quality example is not a latency promise for a detailed composition or a sequence of revisions.

A sensible evaluation would compare both models on the same brief, then include the corrections needed to make each result usable. If one option produces a draft sooner but needs an extra revision to repair the product label, its initial speed advantage may disappear. That is particularly relevant when someone is waiting to approve the asset during a live discussion.

Complexity and quality control different things

The release separates the amount of planning from the fidelity of the rendered image. That gives developers another decision to make besides choosing Flare or Sunburst.

The image-generation guide accepts low, medium and high complexity, with high as the default when omitted. Quality controls rendering separately. OpenAI warns that higher complexity can add latency and unwanted changes; complexity has no separate fee.

For a narrow revision, I would start by testing low complexity while keeping the quality required by the final asset. Changing a small part of an approved composition should not automatically ask the model to reconsider the whole scene. A more crowded brief, with several objects that need specific relationships, deserves a separate comparison.

Markup-guided edits use a separate transparent annotation layer. The guide supports them on both GPT Image 2.5 and 2.6, so this control is not exclusive to the new models.

In a review workflow, that suggests a useful separation: show where the change belongs, describe what should change and explicitly record what has already been approved. The annotation makes the request easier to locate. It does not remove the need to compare the result with the original.

GIF and SVG output have specific API limits

GIFs make short animated assets possible, while SVG output gives developers vector artwork they can adjust after generation. The API details matter before building either into an existing editor.

Both use /v1/images/generations; neither is an output format for the image-edit endpoint. Neither supports streaming or partial images. GIFs expose frame-count and frame-duration controls. SVG requires n=1 and returns static vectors without scripts, external resources or embedded raster images.

That makes a single universal “generate or edit” button a questionable abstraction. An application should make the requested result clear before dispatching the job. A still-image revision, a new animation and vector generation have different output requirements even when the user describes them in similar language.

For SVGs, inspect the actual paths, text and colors in the tool where the asset will be used. A convincing preview and a maintainable vector document are separate acceptance criteria. For GIFs, watch the complete loop rather than approving its opening frame.

GPT Image 2.6 API pricing

The official model pages for Flare and Sunburst list the same standard token rates:

• Text input: $5 per million tokens.
• Image input: $8 per million tokens.
• Image output: $30 per million tokens.

Those rates do not establish one fixed price per image. Equal rates only produce equal bills when billable usage is also equal.

SVG billing includes the intermediate reference image plus the authoring model's tokens used to produce the vector document.

For a realistic budget, keep the entire revision history with each approved asset. Count rejected outputs and corrections, not just the final request. Then compare cost per accepted result alongside turnaround time. A lower-cost draft can become an expensive workflow if it repeatedly requires manual repair.

What still needs review

OpenAI's Images 2.6 system card highlights a specific problem with animation: harmful meaning can arise across frames even when a still image does not show it. The company says motion-related violations remain difficult to detect and its safeguards may miss some.

That is a reason to review generated motion as motion, particularly in a product that accepts prompts from other people. The system card's adversarial evaluations also should not be read as estimates of how often unsafe outputs appear in ordinary production use.

No image-generation requests were run for this article. Before adopting the release, I would use a small set of owned assets and repeat the same edits on both models. Check product identity, exact text, the survival of approved details and whether the exported file works in its intended destination.

Images 2.6 is worth evaluating because more of the asset-production process can happen in one workflow. The practical gain will depend on whether those steps reduce the work between an attractive first result and something a team is willing to publish.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help AI teams turn model releases and product capabilities into clear technical stories that developers can understand and use.

Book a 30-minute call.