GenGrades

GPT Image 2.6 review: Sunburst, Flare, prices and vs 2.5

Published Updated

OpenAIAPI availableClosed weights

What GPT Image 2.6 is

In one or two sentences, before anything else.

GPT Image 2.6 is OpenAI's image generation and editing model, released November 4, 2026 as gpt-image-2.6-flare for fast everyday images and gpt-image-2.6-sunburst for detailed work and precise edits. It adds animated GIF and SVG output, markup-guided edits and a complexity setting, and replaces gpt-image-2, which shuts down on February 10, 2027.

Main features

What it can do, one sentence each, and whether we actually tested it.

  • Animated GIF output

    Tested by us

    Returns a looping GIF of 2 to 16 frames (default 9) at 20 to 2,000 ms per frame, from both the generation and the edit endpoint.

    One sentence gave us an 8-frame sticker with the same cat in every frame; one frame sent back as a reference gave three more expressions.

  • SVG output

    Tested by us

    Returns a vector SVG for flat graphics such as logos and icons. Requires n = 1 and works only on the generation endpoint; OpenAI makes a raster reference image first and converts it.

    Our two-colour bakery logo came back with the name spelled correctly, stored as live text rather than outlines.

  • Markup-guided edits

    Tested by us

    Takes an annotation layer with circles or arrows to show where to edit. The input image and the layer can each be up to 16,777,216 pixels; markup cannot be combined with a mask.

    Our plate landed inside the drawn circle and the red mark was left out of the result.

  • Complexity setting

    Partly tested

    Sets how much the model plans before drawing: low, medium or high, high by default. It is separate from quality and has no fee of its own; OpenAI says higher complexity can add latency and unwanted changes.

    One crowded-scene test only: high matched the food to every stall sign, low was looser.

  • Edits that keep a person or product

    Partly tested

    Holds faces, clothing, products and layout across rounds of edits; OpenAI positions Sunburst for work where editing precision matters most.

    Two rounds on one character kept the face, glasses, coat and props.

  • Text in images, including Chinese

    Tested by us

    Renders multi-line poster text, including Chinese characters and punctuation.

    All three Chinese lines of our poster were correct.

  • Five quality levels

    Quality can be low, medium, high, xhigh or max; gpt-image-2 offers only low, medium and high.

  • Two tiers, Flare and Sunburst

    Flare is built for fast iteration and Sunburst for detailed work; both have the same features and the same token prices.

    We have not compared the two tiers on identical prompts.

Who it is for

The people it helps and the problems it solves, and where it ranks.

Built for

  • Developers adding image generation or editing to an app or workflow
  • Marketing and e-commerce teams making posters, product shots and ad variations
  • Sticker, mascot and social-content makers who need short looping animations
  • Small businesses that need a simple logo or icon as an editable vector file
  • Teams still on gpt-image-2 who must move before February 10, 2027

Problems it solves

  • Making a short looping animation without a video model
  • Getting a logo or icon as SVG instead of a bitmap that blurs when enlarged
  • Changing one exact area of an image without describing its position in words
  • Keeping a character or product the same across several edits
  • Laying out crowded scenes with exact counts and labels in the right order

What it is good at

Real inputs we gave it and the outputs it returned, with why each one worked.

  1. Case 1

    Retro event poster with three lines of Chinese text. Run by the GenGrades team through APIMODELS on November 5, 2026; GPT Image 2.5 Flare on the left, 2.6 Flare on the right.

    Input

    一张竖版活动海报,暖橙色和奶油色配色,复古印刷质感。正中是一只冒着热气的咖啡杯插画。顶部大标题写「周末咖啡市集」,下面一行小字写「11月15日—16日 · 老街广场」,底部写「上午10点 — 晚上8点 · 免费入场」。所有文字清晰、排版整齐,不要出现其他文字。 (3:4, 1K, medium)

    Why it works

    Every character in all three lines was right on 2.6 Flare, down to the dash and the middle dot. 2.6 split the title over two lines and added striped awnings and market stalls without being asked. It took 76 seconds, against 27 for 2.5.

    GPT Image 2.6 output for: Retro event poster with three lines of Chinese text. Run by the GenGrades team through APIMODELS on November 5, 2026; GPT Image 2.5 Flare on the left, 2.6 Flare on the right.
    Output from our run, unedited apart from the watermark.
  2. Case 2

    Change only the background of a photo of a person (2.6 Sunburst, November 5, 2026; original and GPT Image 2.5 Sunburst shown alongside).

    Input

    Change only the background to a snowy mountain village at dusk with warm lights in the windows. Keep her face, hair, glasses, yellow raincoat, pose and coffee cup exactly the same. (Input: a photo of a young woman with short curly red hair, round glasses and a yellow raincoat holding a paper coffee cup; 1K, medium.)

    Why it works

    Face, curls, glasses, raincoat buttons, tote bag and cup all stayed the same. 2.6 also added a chalkboard reading Coffee, Tea, Dessert that the prompt did not ask for. A second round asking her to wave changed only the raised hand.

    GPT Image 2.6 output for: Change only the background of a photo of a person (2.6 Sunburst, November 5, 2026; original and GPT Image 2.5 Sunburst shown alongside).
    Output from our run, unedited apart from the watermark.
  3. Case 3

    Put an object exactly where a red circle is drawn (2.6 Flare with a markup image, against GPT Image 2.5 Flare working from words; November 5, 2026).

    Input

    Request, summarized: add a small white plate with a croissant inside the circled area of the table. Markup: a transparent PNG the size of the photo with a red circle and arrow. 2.5 got the position in words: on the table to the right of the mug.

    Why it works

    2.6 put a sensibly sized plate inside the circle and left the red mark out of the output; the mug, plant and book were untouched. 2.5 understood "to the right" but drew an outsized croissant that runs off the frame.

    GPT Image 2.6 output for: Put an object exactly where a red circle is drawn (2.6 Flare with a markup image, against GPT Image 2.5 Flare working from words; November 5, 2026).
    Output from our run, unedited apart from the watermark.
  4. Case 4

    Animated sticker from one sentence (2.6 Flare, November 5, 2026; the eight frames laid out side by side).

    Input

    Request, summarized: a chubby orange cat sticker with a white outline that waves a paw, blinks, has a pink heart pop up and returns to its starting pose; output_format gif, looping.

    Why it works

    An eight-frame loop came back in about 25 seconds. Stripes, face and outline are identical in every frame; only the paw, the eyes and the heart move.

    GPT Image 2.6 output for: Animated sticker from one sentence (2.6 Flare, November 5, 2026; the eight frames laid out side by side).
    Output from our run, unedited apart from the watermark.
  5. Case 5

    Two-colour shop logo as a real SVG file (2.6 Flare, November 5, 2026; shown rendered).

    Input

    Request, summarized: a flat logo for a neighbourhood bakery called Morning Loaf, a rounded loaf with half a rising sun behind it, only warm orange and dark brown, the name underneath; output_format svg.

    Why it works

    A genuine SVG came back in about 19 seconds, strictly in two colours with the name spelled right. The name was stored as live text, so another computer may show a different font until it is converted to outlines.

    GPT Image 2.6 output for: Two-colour shop logo as a real SVG file (2.6 Flare, November 5, 2026; shown rendered).
    Output from our run, unedited apart from the watermark.
  6. Case 6

    Crowded isometric scene at complexity low and high (2.6 Sunburst, November 5, 2026; low on top, high below).

    Input

    Isometric illustration of a small night market street with exactly 6 food stalls in a row. Each stall has one clear sign, in this order from left to right: "Noodles", "Dumplings", "Tea", "Skewers", "Ice Cream", "Bao". Paper lanterns overhead, a few people queuing at each stall, warm evening light. (16:9, 1K; run twice, complexity low and then high, sent as reasoning_effort on APIMODELS.)

    Why it works

    Both runs drew exactly six stalls with the signs in order. High matched the food to each sign, with buns in steamers and skewers on the grill; low was more cartoonish and left a garbled, watermark-like mark in the lower-right corner.

    GPT Image 2.6 output for: Crowded isometric scene at complexity low and high (2.6 Sunburst, November 5, 2026; low on top, high below).
    Output from our run, unedited apart from the watermark.

Where it falls short

What it does badly, with the evidence, and what we could not settle.

Weak

Weak

  • Low-quality images cost about four times as much as on gpt-image-2.

    Evidence: At the same token rate, OpenAI's calculator prices a 1024×1024 low image at 0.02469 USD on GPT Image 2.6 against 0.00588 USD on gpt-image-2. Teams that make bulk low-quality drafts should budget for this before moving.

  • GIFs are short loops, not video.

    Evidence: Output is capped at 16 frames; our loops were a few seconds of small movement. Camera moves or a storyline need a video model.

  • SVG is for flat graphics only.

    Evidence: SVG needs n = 1, is not available on the edit endpoint and is traced from a raster reference; our logo came back with live text instead of outlines.

  • Adds things you did not ask for.

    Evidence: In our tests 2.6 added market stalls to a poster, a Coffee, Tea, Dessert chalkboard to a background edit, and a few unreadable letters on a book in a product photo. Write "no other text" when text matters.

Uncertain

Uncertain

  • Speed.

    Why unsure: Most of our jobs through APIMODELS took 25 to 30 seconds and one poster took 76; these were single runs on one day, so they neither confirm nor rule out speed claims.

  • Consistency over long chains of edits.

    Why unsure: OpenAI positions Sunburst for repeated editing; the samples on this page cover two rounds of edits on one character.

  • Quality above 1K.

    Why unsure: All our samples are 1K. OpenAI calls resolutions above 2560×1440 experimental.

Price, limits and formats

The numbers, each with its source and the date we checked it.

SpecValueNote and source
Model IDsgpt-image-2.6-flare, gpt-image-2.6-sunburst

Source: developers.openai.comchecked

Release date2026-11-04

As listed on the OpenRouter and Vercel AI Gateway model pages; the OpenAI model page gives no date.

Source: openrouter.aichecked

API endpointsv1/images/generations, v1/images/edits

Source: developers.openai.comchecked

Text input price5 USD / 1M tokens

Cached text input 1.25 USD.

Source: developers.openai.comchecked

Image input price8 USD / 1M tokens

Cached image input 2 USD.

Source: developers.openai.comchecked

Image output price30 USD / 1M tokens

Same rates as gpt-image-2; not listed in the Batch API price table.

Source: developers.openai.comchecked

1024×1024 image, low / medium / high0.025 / 0.043 / 0.053 USD

xhigh 0.094, max 0.211. OpenAI calculator, output tokens only; one setting covers Flare and Sunburst.

Source: developers.openai.comchecked

Quality levelslow, medium, high, xhigh, max

Source: developers.openai.comchecked

Complexitylow, medium, high (default high)

No separate fee.

Source: developers.openai.comchecked

Image sizeSides multiples of 16; ratio 1:3 to 3:1; 655,360 to 8,294,400 pixels

Above 2560×1440 is experimental.

Source: developers.openai.comchecked

Animated GIF2 to 16 frames (default 9), 20 to 2,000 ms per frame

Source: developers.openai.comchecked

SVGn = 1, generation endpoint only

Source: developers.openai.comchecked

Replacesgpt-image-2, which shuts down on 2027-02-10

Source: developers.openai.comchecked

Azure OpenAINot listed (GPT Image 2.5 is)

Source: learn.microsoft.comchecked

Common questions

Answer first, then why, an example, the limits and what to do next.

How much does GPT Image 2.6 cost?

OpenAI bills GPT Image 2.6 by tokens: 5 USD per 1M text input tokens, 8 USD per 1M image input tokens and 30 USD per 1M image output tokens, the same for Flare and Sunburst. By OpenAI's calculator a 1024×1024 image costs about 0.025 USD at low, 0.043 at medium, 0.053 at high, 0.094 at xhigh and 0.211 at max, before input tokens. Third-party providers set their own prices.

What is the difference between GPT Image 2.6 Flare and Sunburst?

Flare is OpenAI's tier for fast, everyday generation and Sunburst is the tier for detailed work where editing precision matters most. Both have the same features, GIF, SVG, markup edits and complexity, and the same token prices. We have not compared the two on identical prompts.

Can GPT Image 2.6 make GIFs?

GPT Image 2.6 returns animated GIFs of 2 to 16 frames when output_format is set to gif, on both the generation and edit endpoints. In our test one sentence produced an eight-frame looping sticker in about 25 seconds with the character identical in every frame. Keep to short, small movements; camera moves and stories need a video model.

Is GPT Image 2.6 better than GPT Image 2?

GPT Image 2.6 adds GIF and SVG output, markup-guided edits and a complexity setting that gpt-image-2 lacks, and OpenAI names it as gpt-image-2's replacement before the February 10, 2027 shutdown. Per image at OpenAI's rates, 2.6 costs more at low quality (0.025 against 0.006 USD at 1024×1024) and less at medium and high (0.043 against 0.053, and 0.053 against 0.211).

Can GPT Image 2.6 output SVG?

GPT Image 2.6 can return an SVG from the generation endpoint when n is 1; OpenAI creates a raster reference image first and converts it. In our test a two-colour bakery logo came back with the name spelled correctly but stored as live text, so convert the text to outlines before using the file as a final logo. It does not turn photos into vectors.

Do I need OpenAI organization verification to use GPT Image 2.6?

OpenAI's image guide says you may need to complete API Organization Verification before using GPT Image models through OpenAI's own API. Third-party providers that bill in their own credit do not require an OpenAI account.

What the public tests show

GPT Image 2.6 is the November 2026 generation of OpenAI's image model, sold through the API as gpt-image-2.6-flare and gpt-image-2.6-sunburst. This page combines OpenAI's documentation, checked on November 11, 2026, with sample images the GenGrades team generated on November 5, 2026.

How the samples were made

The samples were generated through APIMODELS, an API platform made by the team behind GenGrades, at 1K and medium quality unless noted, and are shown as they came back. Several put GPT Image 2.5 next to 2.6 on the same prompt. Each is a single run, not a scored test suite, which is why this page has no score yet. The full write-up of the same tests is on the APIMODELS blog.

What is new

GPT Image 2.6 keeps the Images API endpoints of earlier GPT Image models and adds four things: animated GIF output, SVG output, edits guided by a markup layer, and a complexity setting for how much the model plans before drawing. Quality goes up to xhigh and max, two levels gpt-image-2 does not have. In our tests the new abilities did what OpenAI describes; everyday posters and product shots looked close to GPT Image 2.5.

Price per image against gpt-image-2

OpenAI charges both models the same token rates, but they spend tokens differently. OpenAI's calculator gives these per-image costs for a 1024×1024 image, output tokens only:

Qualitygpt-image-2GPT Image 2.6
low0.006 USD0.025 USD
medium0.053 USD0.043 USD
high0.211 USD0.053 USD
xhighnot offered0.094 USD
maxnot offered0.211 USD

Low-quality drafts cost about four times as much on 2.6, while high-quality images cost about a quarter. GPT Image 2.6 is not in OpenAI's Batch API price table, and Azure OpenAI lists GPT Image 2.5 but not 2.6.

Moving from gpt-image-2

OpenAI removes gpt-image-2 and the gpt-image-2-2026-04-21 snapshot from the API on February 10, 2027 and names the two 2.6 tiers as replacements. Calls go to the same generation and edit endpoints. Two settings deserve a look before switching: complexity defaults to high, which OpenAI says can add latency, and the per-image cost at low quality rises, as the table shows. Rerun a few of your own prompts at the quality you use today before moving a large batch.

Who should not use it

Skip GPT Image 2.6 if you mostly make bulk low-quality drafts and price matters most, if you need real video rather than a few seconds of looping movement, or if you want a photo converted to vectors. Teams that must buy through Azure cannot get it there yet.

Sources

Sources

  1. OpenAI: GPT-Image-2.6 Flare model page (developers.openai.com), accessed
  2. OpenAI: Image generation guide (developers.openai.com), accessed
  3. OpenAI: API pricing (developers.openai.com), accessed
  4. OpenAI: Deprecations (developers.openai.com), accessed
  5. OpenRouter: GPT Image 2.6 Flare (openrouter.ai), accessed
  6. Vercel AI Gateway: GPT Image 2.6 Flare (vercel.com), accessed
  7. Microsoft Learn: Foundry model region availability (learn.microsoft.com), accessed
  8. APIMODELS blog: GPT Image 2.6 vs 2.5, tested (apimodels.app), accessed

Changelog

Changelog

  1. First version: features, limits and prices from OpenAI's documentation checked on November 11, 2026, with six sample outputs the GenGrades team generated through APIMODELS on November 5, 2026.

Recommended from GenGrades

  • Guide

    GPT Image 2 API Price per Image: 23 Providers Compared

    APIMODELS, our own product, publishes a 30-day success rate on every model page, 99.9% for gpt-image-2; of the other 22 GPT Image 2 API providers compared, only OpenRouter and CometAPI publish any model-level reliability figure. Prices were read on November 11, 2026, and the lowest medium 1K price is APIMODELS at 0.025 USD.

  • Model review

    Claude Fable 5.2

    Claude Fable 5.2 is Anthropic's most capable public model, released November 9, 2026 at 8 and 40 USD per million input and output tokens, 20% below Fable 5.1. Against Opus 5.5 it costs double per token and, at max effort, scores 65 to 59 on Artificial Analysis's index for 26.54 USD per task against 10.04; Anthropic still recommends Opus first.

  • Tool review

    Hailuo AI

    Hailuo AI is MiniMax's AI video and image app for web and Android, built since July 31, 2026 around MiniMax H3, an open-weight model that makes up to 15 seconds of 2K video with sound. Plans start at US$10.50 a month for 1,000 credits. Payments are not refunded, and a studio copyright suit over Hailuo survived dismissal in May 2026.

  • Tool review

    Devin

    Devin is Cognition's autonomous coding agent: it works on its own cloud computer and opens pull requests, and ships with Devin Desktop (formerly Windsurf) and a CLI. In November 2026 Cognition added the cheaper SWE-2 Flash model and Devin Review 2.0. Plans cost US$0 to US$200 a month; the best-known outside test, Answer.AI's, saw 3 successes in 20 tasks.

  • Model review

    Kimi K3.1 Pro

    Kimi K3.1 Pro is Moonshot AI's open-weight 2.8-trillion-parameter model with a 1M-token context, released October 23, 2026, weights out November 4. Its API costs 2 and 10 USD per million tokens, half Claude Opus 5.5's. Vals AI puts it near Opus on Terminal-Bench 4.0, but independent aggregates rank it lower, it is slow, and an audit questions its benchmark lead.

  • Model review

    Tencent Hy4

    Tencent Hy4 is Tencent's flagship language model, released November 4, 2026, and its first Hy model with vision and computer use. Until December 31 Tencent bills half its 6-yuan input and 18-yuan output list price per million tokens; OpenRouter charges 0.417 and 1.25 USD. Mercor ranks it third on APEX-Agents, but its weights are unreleased and coding is slow.