GenGrades
EN中文

这篇内容还没有中文版

下面是英文原文。中文版发布后,这里会换成中文。 打开英文页面

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

发布于 更新于

OpenAI提供 API闭源权重

GPT Image 2.6 是什么

先用一两句话说清楚。

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.

主要功能

它能做什么,每项一句话,以及我们有没有实测。

  • Animated GIF output

    已实测

    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

    已实测

    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

    已实测

    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

    部分实测

    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

    部分实测

    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

    已实测

    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.

适合谁

它帮得上谁、解决什么问题、在哪个榜单里。

适合

  • 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

解决的问题

  • 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

擅长什么

我们给它的真实输入、它返回的输出,以及效果好的原因。

  1. 案例 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.

    输入

    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)

    为什么有效

    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 的输出: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.
    我们实测的输出,除水印外未做修改。
  2. 案例 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).

    输入

    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.)

    为什么有效

    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 的输出: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).
    我们实测的输出,除水印外未做修改。
  3. 案例 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).

    输入

    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.

    为什么有效

    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 的输出: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).
    我们实测的输出,除水印外未做修改。
  4. 案例 4

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

    输入

    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.

    为什么有效

    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 的输出:Animated sticker from one sentence (2.6 Flare, November 5, 2026; the eight frames laid out side by side).
    我们实测的输出,除水印外未做修改。
  5. 案例 5

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

    输入

    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.

    为什么有效

    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 的输出:Two-colour shop logo as a real SVG file (2.6 Flare, November 5, 2026; shown rendered).
    我们实测的输出,除水印外未做修改。
  6. 案例 6

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

    输入

    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.)

    为什么有效

    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 的输出:Crowded isometric scene at complexity low and high (2.6 Sunburst, November 5, 2026; low on top, high below).
    我们实测的输出,除水印外未做修改。

短板

做得不好的地方和证据,以及我们还没法定论的地方。

做得不好

做得不好

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

    证据: 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.

    证据: 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.

    证据: 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.

    证据: 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.

尚不确定

尚不确定

  • Speed.

    原因: 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.

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

  • Quality above 1K.

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

价格、限制与格式

每个数字都附来源和核对日期。

项目数值备注与来源
Model IDsgpt-image-2.6-flare, gpt-image-2.6-sunburst

来源:developers.openai.com核对于

Release date2026-11-04

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

来源:openrouter.ai核对于

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

来源:developers.openai.com核对于

Text input price5 USD / 1M tokens

Cached text input 1.25 USD.

来源:developers.openai.com核对于

Image input price8 USD / 1M tokens

Cached image input 2 USD.

来源:developers.openai.com核对于

Image output price30 USD / 1M tokens

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

来源:developers.openai.com核对于

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.

来源:developers.openai.com核对于

Quality levelslow, medium, high, xhigh, max

来源:developers.openai.com核对于

Complexitylow, medium, high (default high)

No separate fee.

来源:developers.openai.com核对于

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

Above 2560×1440 is experimental.

来源:developers.openai.com核对于

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

来源:developers.openai.com核对于

SVGn = 1, generation endpoint only

来源:developers.openai.com核对于

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

来源:developers.openai.com核对于

Azure OpenAINot listed (GPT Image 2.5 is)

来源:learn.microsoft.com核对于

常见问题

先给结论,再讲原因、例子、限制和下一步。

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.

公开测试怎么说

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.

来源

来源

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

更新记录

更新记录

  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.

GenGrades 推荐阅读

  • 指南

    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.

  • 工具评测

    MagicRemover

    MagicRemover 是一个在浏览器里使用的工具,能从照片中擦除人物、物体、文字、logo 和水印,也能去掉视频里的字幕。照片每天免费 2 次,登录后再加 2 次,输出约 100 万像素;超出后按积分付费,最低 $9.90 买 330 积分。MagicRemover 是我们自己的产品,和 GenGrades 出自同一团队。

  • 模型评测

    Claude Fable 5.2

    Claude Fable 5.2 是 Anthropic 对外开放的最强模型,2026 年 11 月 9 日发布,每百万 token 输入 8 美元、输出 40 美元,比 Fable 5.1 便宜 20%。和 Opus 5.5 相比,它的 token 单价是两倍;在 max 推理强度下,它在 Artificial Analysis 指数上以 65 比 59 领先,每个任务却要 26.54 美元,Opus 是 10.04 美元;Anthropic 仍建议先用 Opus。

  • 工具评测

    Hailuo AI

    海螺 AI(Hailuo AI)是 MiniMax 推出的 AI 视频和图片应用,有网页版和安卓版;2026 年 7 月 31 日起以开放权重的 MiniMax H3 为核心,能生成最长 15 秒、带声音的 2K 视频。付费套餐每月 10.50 美元起,含 1,000 积分。已付款项不退,迪士尼等片方针对海螺的版权诉讼也在 2026 年 5 月闯过了驳回起诉这一关。

  • 工具评测

    Devin

    Devin 是 Cognition 推出的 AI 编程智能体(AI 软件工程师):它在自己的云端电脑上干活,做完就提交 PR,并配有 Devin Desktop(原 Windsurf)和命令行工具 Devin CLI。2026 年 11 月,Cognition 推出了更便宜的 SWE-2 Flash 模型和 Devin Review 2.0。套餐每月 0 到 200 美元;最知名的外部测评来自 Answer.AI,20 个任务只成功了 3 个。

  • 模型评测

    Kimi K3.1 Pro

    Kimi K3.1 Pro 是月之暗面(Moonshot AI)的开放权重模型,总参数 2.8 万亿,上下文 100 万 token,2026 年 10 月 23 日发布,11 月 4 日公开权重。API 每百万 token 输入 2 美元、输出 10 美元,是 Claude Opus 5.5 的一半。Vals AI 测得它的 Terminal-Bench 4.0 成绩接近 Opus,但独立综合榜单把它排得更低;它速度慢,还有一份外部审查质疑它的跑分领先。