On November 4, 2026, OpenAI’s Changelog announced GPT‑Image‑2.6 Sunburst (gpt-image-2.6-sunburst) and GPT‑Image‑2.6 Flare (gpt-image-2.6-flare), with adjustable complexity, SVG generation, animated GIF generation, and markup-guided editing for supported GPT Image 2.5/2.6 models. The product-layer name Images 2.6 appears in ChatGPT and Codex; developer calls must use API model IDs. This article only summarizes integration facts verifiable on official pages—no invented discounts, undocumented endpoint behavior, or unverified competitor comparisons.

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

Endpoints and model choice

Per the Flare and Sunburst model pages, both tiers support Image API routes for generation and editing: POST /v1/images/generations for text-to-image; POST /v1/images/edits for edits, reference composition, and markup-guided changes. The same model pages mark Responses (v1/responses) as unsupported. By contrast, the image generation guide still uses GPT Image 2.5 examples for image tools in Responses; the 2.6 body also stresses that complexity, markup, SVG, and GIF are provided “via the Image API.” Production 2.6 integrations should therefore default to the Image API and not assume 2.5-era Responses tool configs can simply swap in a 2.6 identifier. If the business strongly depends on multi-turn conversational editing, re-check whether docs have updated Responses support for 2.6 and treat the model-page Endpoints table as authoritative—do not extrapolate from habit.

On selection, Flare emphasizes fast iteration; Sunburst emphasizes detail and edit consistency; token unit prices verified here are identical. Before using GPT Image, organizations may need organization verification in the developer console. For end-user-facing apps, docs recommend passing a stable, hashed safety identifier of at most 128 characters on every generation or edit request; the older user field is deprecated but still compatible, with a non-empty safety identifier taking precedence.

Output formats, background, and 2.6-specific fields

Generation output formats support portable network graphics (default), JPEG, WebP, plus the 2.6 expansions SVG and animated GIF. JPEG and WebP allow a compression ratio; SVG and GIF do not support that compression field, nor streaming, partial images, or the prompt rewriter. Edits do not support SVG or GIF output; SVG is also not an acceptable raster input to the edits interface. For transparent backgrounds, explicitly request transparency and choose a format that supports an alpha channel; JPEG does not. After decoding, inspect the transparency channel—a drawn checkerboard is not the same as true transparent pixels.

GIF additionally supports frame count (2–64, default 9) and per-frame duration in milliseconds (20–2000, default 200). Quality tiers include auto, low, medium, high, xhigh, and max; complexity includes low, medium, and high, defaulting to high when omitted. Markup edits upload a same-size transparent overlay via the edits annotation field and cannot be combined with a mask. Common sizes include square and recommended landscape/portrait resolutions; custom width×height strings that meet side-multiple and aspect-ratio constraints are also supported. Size, quality, and background can all be set to auto for model choice from the prompt.

The image prompting guide also notes: omit the older input-fidelity parameter on image edit requests; GPT Image 2.6 does not support reasoning-summary-style fields. When migrating existing workflows, save baseline prompts and samples first, then compare under matching conditions by corresponding tier (prior Flare → 2.6 Flare; prior Sunburst → 2.6 Sunburst), changing only quality or complexity one at a time.

Pricing (model-page verified values)

As of this article, Flare and Sunburst model-page list prices match: text input $5 per million tokens, text cached input $1.25; image input $8, image cached input $2, image output $30 per million tokens. Actual bills vary with prompt length, reference images, quality tier, and output resolution; batch and other discounts follow the pricing page at the time. Model pages in the capture scope for this article do not list a fixed per-image price—billing is token-based. The complexity parameter has no separate fee. Rate limits vary by usage tier and account; check the organization console rather than copying public figures from other models.

Minimal integration checklist

Before launch, confirm item by item: organization verification completed as required; Flare or Sunburst formal model IDs passed explicitly on generation and edit requests; still, vector, or motion output format chosen for the deliverable; edit tasks preferentially set complexity explicitly to low to avoid default high planning depth; targeted changes use a markup overlay rather than coordinates described only in text; measure latency, content-block rate, and accepted per-image cost on business corpora. On errors, quota and service faults may use backoff retries, but image-generation user errors (including moderation blocks) should change the prompt or input image first—do not retry blindly. Safety evaluation and mitigations are in the official system card. Product availability in ChatGPT and Codex does not automatically equal quota or verification status for a given API project—check them separately.

Bottom line: the 2.6 interface surface centers on Image API generations and edits, with dual model IDs, expanded output formats, and quality, complexity, and markup fields as controllable capability. Strictly separating chat product names from interface identifiers, and treating model-page endpoint and pricing tables as authoritative, completes a docs-level path from evaluation to launch—without baking unsupported Responses configs or unverified unit prices into production dependencies.