GPT Image 2.6: Flare vs Sunburst, Pricing and API Guide

On November 4, 2026, OpenAI released Images 2.6 in ChatGPT, Codex and the API. The API version ships as two models, GPT-Image-2.6 Flare for fast iteration and GPT-Image-2.6 Sunburst for detailed work. Both bill at the 2.5 generation's token rates: $8 per million image input tokens and $30 per million image output tokens. The news is in the controls rather than the price. Developers can now ask for an animated GIF, an SVG file, an edit guided by a drawn annotation, and a planning level the API calls complexity.

This is the third GPT Image release of 2026. GPT Image 2 arrived on April 21, the 2.5 pair on September 8, and 2.6 on November 4, dates taken from OpenAI's API changelog rather than from any marketplace listing. The gap between releases has shortened from twenty weeks to eight. Anyone building image generation into a product should plan for that pace, which is why the last section of this guide is about the models leaving on December 1 as much as the models that arrived.

Every figure below comes from OpenAI's announcement, its model and pricing pages, its image generation guide or the Images 2.6 system card, all read on November 5, 2026. Speed claims are OpenAI's and are labelled as such. We have not run our own latency or quality tests on these models. For the previous generation, our GPT Image 2.5 guide covers the Flare and Sunburst split as it first appeared.

The announcement covers three surfaces at once. In ChatGPT and Codex, OpenAI says Images 2.6 is available globally on every subscription tier across desktop, mobile and web. The post does not say which of the two underlying models ChatGPT uses, and the system card says only that users in chat control reasoning effort through the model picker; it does not mention a Flare or Sunburst choice. In the API, the two models appear by name on the generations and edits endpoints.

The headline claim is speed. OpenAI states that Images 2.6 generates images 50% faster than before, without naming the baseline in the announcement text. Flare carries the sharper number: when quality is set to low, OpenAI says it can produce an image in as little as six seconds. Treat both as vendor figures until a third party times them. The announcement also gives a scale figure, 4 billion images created per week across ChatGPT, Codex and the API, which it restates as more than half a billion a day.

The practical change for developers is the feature list. The changelog entry for November 4 names adjustable complexity, SVG generation and animated GIF generation for the two 2.6 models, and adds markup-guided editing for supported GPT Image 2.5 and 2.6 models. That last clause matters if you are still on 2.5: the markup feature reaches you without a model change. The failure case is assuming the same of GIF and SVG, which the guide ties to the 2.6 models only.

The two model pages carry one-line taglines that do most of the work. Flare is "Image generation and editing for fast iteration". Sunburst is "for detailed visual work". The announcement adds the operating detail: Flare's six-second claim applies at low quality, while Sunburst's advantage, higher quality and greater consistency across edits, is stated for medium or high quality. The system card adds that Flare is optimised for faster generation and that both models improve editing consistency, detail preservation and response to sketches and annotations compared with 2.5.

A worked example makes the split concrete. A social team producing twenty candidate visuals for a campaign brief wants to see them in a minute and discard most of them. That is Flare at low quality. The one visual that survives, which now needs the product label corrected and the brand colour matched across three crops, is Sunburst work at medium or high quality, because consistency across edits is the property OpenAI claims for it.

The failure case is choosing by name rather than by stage. Teams that run every request through Sunburst at high quality pay in latency for drafts nobody will keep. Teams that run final edits through Flare at low quality get speed on the one asset where speed was never the constraint. Neither model is cheaper per token, so the choice is about time and consistency, not the rate card.

OpenAI prices image models per token, not per image. For the 2.6 models that means $8 per million image input tokens, $2 per million cached image input tokens, $30 per million image output tokens, $5 per million text input tokens and $1.25 per million cached text input tokens. The pricing page shows identical rows for the 2.5 pair and for GPT Image 2. Only the older models differ: GPT Image 1.5 charges $32 for image output and $10 for text output, and GPT Image 1 Mini sits at $2.50 input and $8 output.

The sentence to carry from OpenAI's guide is a warning, not a rate. Equal token rates do not mean equal cost per image, because token consumption can differ by model and quality setting. The guide tells developers to read the usage object on each response to measure consumption for their own prompts, sizes and quality levels. OpenAI sets no fixed per-image price for the 2.6 models. Its guide's cost calculator estimates output tokens and output cost for one image at a chosen quality and size, excluding input tokens and partial images. We have not added an estimate of our own.

Two billing footnotes apply. Cached input rates for GPT Image 2 and 2.5 apply only to the image generation tool in the Responses API, not to direct Images API requests, and the 2.6 models are not yet on that tool. Streaming partial images costs an additional 100 image output tokens per partial. Our GPT Image 2.6 comparison sets these rates against six other image models, each in the units its vendor publishes.

The new request field is called complexity and it is separate from quality. OpenAI's guide draws the line plainly: quality controls rendering detail and fidelity, complexity controls how much the model plans before it renders. The accepted values are low, medium and high. If the field is omitted, both models use high. Adjusting it carries no separate fee, so the cost of a wrong setting is time and output, not money.

OpenAI's recommendation is specific. For most editing tasks it suggests low complexity for faster edits, paired with whatever image quality the workflow needs. For scenes with many elements or constraints, dense text or relationships between objects, it suggests medium or high. The guide's own example, a paper-cut football scene generated by Sunburst at medium quality, shows the trophy, the uniforms and the placement of a sign changing as complexity rises.

The failure case is written into the documentation. Higher complexity can increase latency and introduce unwanted changes, so OpenAI tells developers to compare results for their task. In practice that means the default is the expensive setting in time, and an edit pipeline that never sets the field is planning every background swap as if it were a full scene. Set it low for edits, raise it when the brief is dense, and keep the two dials in the request template so nobody has to remember.

GIF output is set with output_format on the generations endpoint, and only there. The edit endpoint does not produce GIFs. The guide asks for the subject, the motion and the desired loop in the prompt, and exposes three fields for the animation itself: frame_count from 2 to 64 with a default of 9, frame_duration_ms from 20 to 2,000 milliseconds per frame rounded to 10 and defaulting to 200, and size, which describes each finished frame and must use dimensions that are multiples of 16 unless set to auto.

The constraints are about budget. Frame count, frame dimensions and the number of animations have to fit the model's size limits and a total pixel budget per request; if a request exceeds them the fix is fewer frames, smaller frames or a smaller n. GIF output does not support streaming, partial images, custom compression or the prompt rewriter. OpenAI's own examples are marketing-shaped: a brand mascot, social reactions, animated storyboards before video production, and a typing indicator in a workplace assistant built by a startup called Town.

The system card is the part of the release most coverage will skip, and it is where the caveat lives. OpenAI states that GIF generation introduces risks that are not apparent in still images, including meaning carried across frames, and that GIF-specific rules block some content categories more broadly than for stills. On a targeted adversarial set, its GIF moderation recall was 77.7% overall, with sexual content at 77.8% and self-harm at 77.5% as the weakest categories. OpenAI's own wording is that motion-related violations remain challenging to detect. For a brand pipeline that means a human still reviews every GIF before it ships.

SVG is also an output_format on the generations endpoint, with n set to 1. The guide describes a two-step process: the API creates a reference image, then produces the SVG from it, and the quality setting controls that reference image. The returned document is static vector content with no scripts, external resources or embedded raster images, which is what a web team wants for icons and illustrations. SVG is not available from the edit endpoint, and an SVG cannot be fed back in as a raster input for edits.

The billing is the detail to budget for. OpenAI's cost section says you pay the image model's rate for the reference image and the authoring model's rate for the tokens it uses to write the SVG, and the total depends on how many tokens each model consumes. Two meters, one request. A logo pipeline that assumed SVG cost the same as a PNG will under-forecast, and the only way to size it is to run the real prompts and read usage.

Markup-guided editing is the feature marketers will recognise from ChatGPT. On the edits endpoint, you send an annotation layer as markup, for example a transparent overlay with a red circle around a vase, and describe the change in the prompt. Markup applies to the first input image; additional images act as references. The first image and the annotation can each hold at most 16,777,216 pixels, the annotation file is capped at 50 MB, and markup cannot be combined with a mask or the prompt rewriter. Markup edits use the same token rates as other edits with no extra fee, and the changelog extends them to supported 2.5 models too.

The system card, published November 4, 2026, evaluates both models against their 2.5 counterparts on a shared set of 3,077 adversarial prompts designed to produce policy-violating images. It reports three outcomes: safe generation, unsafe generation blocked by moderation, and unsafe generation not blocked. OpenAI's own caveat comes first: these results are not representative of how often violative requests or outputs arise in production traffic, and they measure the evaluated configuration, not the complete deployed system.

Sunburst improved slightly. Its unsafe-not-blocked rate fell from 1.92% to 1.72%. Flare moved the other way, from 1.49% to 2.14%, which OpenAI states as 46 to 66 of 3,077 requests. The card says neither overall difference is statistically significant. One change is significant: before moderation, the share of Flare outputs graded non-violative fell from 78.36% to 75.04%, with the largest contributions from the "Other" and sexual categories. Fourteen of the twenty net additional unblocked outputs sat in the "Other" category, which covers religious figures, depictions of real people and jailbreak attempts.

For a buyer the reading is straightforward. The deployed safety stack is the same foundation as 2.5 with GIF-specific additions, the Preparedness Framework assessment finds neither model crosses the Bio High or Cyber High thresholds, and OpenAI still applies High-capability biological mitigations as a precaution. The failure case is treating those numbers as a guarantee about your own outputs: they describe an adversarial offline test, and OpenAI says so twice.

OpenAI's deprecations page, read November 5, 2026, lists three image models with a shutdown date of December 1, 2026: gpt-image-1-mini, gpt-image-1.5 together with its dated snapshot, and the chatgpt-image-latest alias. The recommended replacement for all three is gpt-image-2.5-sunburst or gpt-image-2.5-flare. The 2.6 models are not named as replacements, which is a reasonable reading of how new they are rather than a judgement on them. The original gpt-image-1 was shut down earlier, on October 23, 2026, with the same 2.5 replacements.

The migration that will bite is gpt-image-1-mini. It is the only model on the sheet with a lower rate card, $2.50 image input and $8 image output against $8 and $30 for everything newer. A team that chose Mini for cost has no like-for-like successor on OpenAI's price page, and will need to re-measure token consumption on a 2.5 model at low quality before it can say what the change costs. Anyone who still relies on the chatgpt-image-latest alias should also pin a named model now, because the alias disappears on the same day.

A note on sources. Marketplace listings elsewhere show December 1 expiry dates for OpenAI's image-mini routes. We have stated only what OpenAI's own deprecations page lists, and a listing date is not a vendor notice. Our alias and retirement ledger tracks these dates across vendors as they are published.

Strip the release to what changes a marketing workflow and three things remain. GIFs move a brand mascot or a product loop from a motion designer's queue to a prompt, with a moderation recall figure that says a human still approves the output. SVG moves icons and simple illustrations into an editable format, at the cost of two models' tokens per request. Markup turns "change the vase on the left" from a paragraph into a circle, and reaches 2.5 users without a model switch.

The example we would run first is a product-image refresh. Draft twenty hero compositions on Flare at low quality and low complexity, pick two, then move those two to Sunburst at high quality for the edits that have to hold across every crop, using markup to point at the label and the shadow. Read usage after each stage and record tokens per approved asset, because that figure, not the rate card, is the cost of the pipeline. Our GPT Image 2 review covers the generation that shipped in April.

The failure case is the default. A request template that omits complexity, routes everything to one model and never reads the usage object will produce slower drafts than necessary, finals no better than before, and a bill nobody can explain. Teams building creative production on these models, or deciding whether to, can work through that template with our content engine and paid media practices.