gpt-image-2.6-flare Guide: When to Pick the Fast Image Model

What it costs, and why same price is not same cost

The model page and the pricing page list gpt-image-2.6-flare at $5 per million text input tokens ($1.25 cached), $8 per million image input tokens ($2 cached), and $30 per million image output tokens. Sunburst is listed at exactly the same rates, and so are both 2.5 models.

Because the rates match, price alone cannot choose the model for you. Images are billed by tokens, and the same prompt consumes a different number of tokens depending on model, size, and quality. OpenAI's own migration advice tells you to track failures, retries, and cost per accepted image rather than the cost of a single request. Calling Flare "the cheap one" will mislead a budget. The accurate version is that Flare may get you to an acceptable image sooner, and whether it also costs less depends on your acceptance rate.

Tune quality and complexity separately

Version 2.6 splits "how finely to render" from "how much to plan." quality controls rendering detail such as texture or small lettering, with auto, low, medium, high, xhigh, and max. complexity controls how much the model plans before rendering, with low, medium, and high. If you omit it, both Flare and Sunburst use high. OpenAI states that adjusting complexity has no separate fee.

The documentation is specific about when to move it:

• For focused edits, such as changing a background or a sign, low complexity can suit the job. Pair it with whatever quality your workflow needs.
• For scenes with many elements or constraints, such as dense text, complicated layouts, or relationships between objects, try medium or high.
• Higher complexity can increase latency and introduce unwanted changes, so compare results rather than assuming higher is better.
• When comparing complexity levels, hold quality fixed. If the layout is right but small text is hard to read, then compare quality settings at the same complexity.

Since the default is high, anyone who wants speed should lower complexity explicitly and then check whether response time really improves. It is the setting Flare users are most likely to miss.

Guide an edit with markup

Markup is new to the API in 2.6. You draw circles or sketches on a transparent layer to show where a change should happen, then describe the change in the prompt. The image generation guide demonstrates it with Flare: a striped vase in a kitchen photo, circled in red, is recolored to solid cobalt blue ceramic. The request works like this:

1. Send the original photo as the first input image.
2. Create a separate transparent PNG that holds only your circles or sketches. Do not flatten it onto the photo. After orientation, the annotation layer and the first image must have the same width and height.
3. Send that layer in the markup field. Markup applies only to the first input image, and any additional images act as references.
4. In the prompt, say what to change and what to keep, for example: recolor the striped vase inside the red circle to solid cobalt blue ceramic, keep its shape, the dried stems, and the rest of the kitchen, and remove the red circle from the finished photo.

The documented limits: the first input image and the annotation layer can each hold at most 16,777,216 pixels, the annotation file is capped at 50 MB, and you cannot combine markup with mask or prompt_rewriter. Leave out input_fidelity on edits. OpenAI also tells you to compare the rest of the output with the original to check for stray changes, and if part of an image must stay pixel-identical, composite the approved edit back into the original instead of trusting the prompt.

Migrating from GPT Image 2 or 2.5 Flare

The image prompting guide lays out a six-step migration that fits Flare well:

1. Save a baseline: representative prompts, reference images, and current outputs, including difficult edits, exact text, product geometry, and transparent assets.
2. Pick the matching model. Running 2.5 Flare means testing 2.6 Flare first. A GPT Image 2 workflow that already meets your quality bar should also start with Flare.
3. Keep the first comparison identical: same prompt, references, size, quality, and output format, with complexity at its default.
4. Inspect text, layout, identity, and the details that should survive an edit. For a chain of edits, check the final image and each intermediate one.
5. Change one setting at a time, either quality or complexity. Repeat each request and compare how often the output is acceptable and how long it takes.
6. Roll out by workflow. Move a small share of traffic, then watch response time, failures, retries, and cost per accepted image before expanding.

Mind the retirement dates as well. OpenAI's documentation says gpt-image-1 was shut down on October 23, 2026, and gpt-image-1.5 is scheduled to shut down on December 1, 2026. If you still have either in production, check the parameter differences in the official reference rather than copying old request settings straight into 2.6.

When to move from Flare to Sunburst

Switch your test to gpt-image-2.6-sunburst when you see any of these on the same prompts and references:

• Small text, labels, or dense layouts keep failing even after you raise quality.
• People, products, or brand elements drift after several edit rounds.
• The deliverable needs more detail density, such as formal ad creative or a matched product set.

OpenAI's ordering is to confirm that Sunburst meets your requirements, then try Flare with identical inputs and keep Flare if it passes and responds faster. Neither model is the default answer. Your acceptance criteria are. For the previous generation's version of this question, see the GPT Image 2.5 guide, which covers 2.5 Flare and Sunburst.

Common questions

What is the difference between gpt-image-2.6-flare and gpt-image-2.6-sunburst?

OpenAI positions Flare for speed and fast iteration and Sunburst for detail and edit precision. Pricing, parameters, and output formats are the same. The practical difference is the speed and quality trade-off, and how large it is on your task is something you have to measure.

How much does gpt-image-2.6-flare cost?

Text input is $5 per million tokens, image input is $8, and image output is $30, with cached input at $1.25 and $2. What one image costs depends on size, quality, and prompt, so compare the total cost of accepted images.

Can regular ChatGPT users use Images 2.6?

Yes. OpenAI says Images 2.6 is available in ChatGPT and Codex across all subscription tiers on desktop, mobile, and web. Flare and Sunburst are the two models OpenAI offers developers in the API.

Does Flare support SVG and GIF output?

According to the image generation guide, both models can produce SVG and animated GIF through the Image API. Both are available on the generation endpoint only, and the edit endpoint does not support SVG or GIF output. For GIF limits, rely on the official API reference, since the guide and the reference list different frame ranges.