OpenAI prices GPT-Image-2.6 at $30 per 1M image tokens with no Batch discount
OpenAI added GPT-Image-2.6 Sunburst and GPT-Image-2.6 Flare to its API on November 4, 2026, in the same release that brought ChatGPT Images 2.6 to ChatGPT and Codex. The two model pages list exactly the per-token rates of the previous generation: $5 per 1M text input tokens, $8 per 1M image input tokens and $30 per 1M image output tokens, with cached input at $1.25 and $2. What changed is how the models can be bought and called. Both 2.6 models are marked as unsupported on the Batch endpoint, which on GPT-Image-2.5 halves image output to $15 per 1M tokens, and both are limited to the Image API's generation and edit endpoints.
The new models bring four capabilities that OpenAI's API changelog lists for both of them: an adjustable complexity setting, markup-guided editing, SVG generation and animated GIF generation. In its announcement, OpenAI describes the pair as delivering "50% faster image generation", says Flare "generates images in as little as six seconds" and positions Sunburst as the model for "higher-quality generation and greater consistency across edits". On the consumer side, ChatGPT Images 2.6 adds animated GIFs that can be edited frame by frame and saved reference images of people, pets and characters that can be reused without uploading them again.
For developers the headline rate card hides two cost questions. The first is the loss of the 50% Batch discount for teams that generate large catalogues offline. The second is the number of output tokens each image consumes, which OpenAI has not published for 2.6 and which early testers on OpenAI's developer forum report to be higher at low and medium quality than on GPT-Image-2.5. Prices in this report are as listed on OpenAI's official model and pricing pages on November 5, 2026.
··········
GPT-IMAGE-2.6 SUNBURST AND FLARE: WHAT OPENAI RELEASED IN THE API.
OpenAI's image generation guide summarizes the split in one line: GPT Image 2.6 "includes Sunburst for detailed visual work and Flare for fast iteration." The model pages repeat that positioning, describing Sunburst as image generation and editing "for detailed visual work" and Flare as image generation and editing "for fast iteration". Both accept text and image input and return images, and both are served through v1/images/generations and v1/images/edits. Every other endpoint, including Responses, Chat Completions and Batch, is listed as unsupported, and each model currently has a single model ID with no dated snapshot for pinning.
The new complexity parameter accepts low, medium and high, and both models default to high when the field is omitted. OpenAI's guide separates it from quality, which still "controls rendering quality", and states that changing complexity carries no separate fee. According to the guide, lower complexity suits focused edits such as replacing a background or a sign, while higher complexity helps with dense text, complex layouts and relationships between objects, at the cost of extra latency and a greater risk of unwanted changes elsewhere in the image.
Markup-guided editing changes the way edits are specified. In the documentation quoted on OpenAI's forum, a markup layer is a transparent image on which the developer draws visible annotations to show what should change, where the older mask approach marked editable regions through transparency. The documentation says markup must not be combined with mask or prompt_rewriter. SVG output, according to the same documentation, requires n=1 and does not support streaming, partial images or output compression, and SVG edits are not supported.
........
Item | GPT-Image-2.6 Sunburst | GPT-Image-2.6 Flare |
OpenAI positioning | Detailed visual work | Fast iteration |
Speed claim | Higher quality, edit consistency | Images in as little as 6 seconds |
Endpoints | Image generations, image edits | Image generations, image edits |
Batch API | Not supported | Not supported |
Complexity setting | low, medium, high (default high) | low, medium, high (default high) |
Markup edits, SVG, GIF | Yes | Yes |
Dated snapshot | None listed | None listed |
........
··········
GPT-IMAGE-2.6 API PRICING COMPARED WITH GPT-IMAGE-2.5.
The per-token rate card did not move. Both 2.6 models list the same text and image prices as GPT-Image-2.5 Sunburst and Flare, which OpenAI launched in September at rates that were themselves unchanged from GPT Image 2. The difference sits in the Batch column. OpenAI's pricing page gives GPT-Image-2.5 a Batch tier at half price, with image output at $15 per 1M tokens, text input at $2.50 and image input at $4. The 2.6 model pages list no Batch price and mark the Batch endpoint as unsupported. On November 5 the main pricing page still listed only gpt-image-2, gpt-image-2.5-sunburst and gpt-image-2.5-flare in its image generation section, so the 2.6 rates below come from the model pages.
........
Model | Text input | Image input | Cached text / image | Image output | Batch image output | Available from |
gpt-image-2.5-sunburst | $5.00 | $8.00 | $1.25 / $2.00 | $30.00 | $15.00 | Sep 8, 2026 |
gpt-image-2.5-flare | $5.00 | $8.00 | $1.25 / $2.00 | $30.00 | $15.00 | Sep 8, 2026 |
gpt-image-2.6-sunburst | $5.00 | $8.00 | $1.25 / $2.00 | $30.00 | Not supported | Nov 4, 2026 |
gpt-image-2.6-flare | $5.00 | $8.00 | $1.25 / $2.00 | $30.00 | Not supported | Nov 4, 2026 |
........
All prices are in USD per 1M tokens. Text output is not billed because the models return images only, and OpenAI has not published a per-image price table for any of the four models, pointing developers to the calculator in its image generation guide instead. The guide's calculator and code samples still reference the 2.5 models, so there is no official per-image figure for 2.6 at any quality or size as of this writing.
··········
WHAT A PRODUCT-IMAGE WORKLOAD COSTS ON GPT-IMAGE-2.5 AND GPT-IMAGE-2.6.
Because image generation is billed by output tokens, the cost of a picture depends on how many tokens the model spends on it. OpenAI has not published those counts for 2.6. The only public data point is a test posted on OpenAI's developer forum, in which the same 280-token prompt produced 344 output tokens on GPT-Image-2.5 at medium quality and 1,136 output tokens on GPT-Image-2.6 at medium quality. The same tester reported that the gap appears at low and medium quality and disappears at high quality and above. It is a single unofficial measurement, and the example below uses it only to show how sensitive the bill is to token counts.
The scenario is an e-commerce tool that generates 2,000 product images a day at medium quality, each from a 280-token text prompt, with no image input and no caching. On GPT-Image-2.5 at standard rates, one image costs 280 × $5/1M plus 344 × $30/1M, or about $0.0117. On GPT-Image-2.6 with the reported 1,136 output tokens, the same image costs 280 × $5/1M plus 1,136 × $30/1M, or about $0.0355. Teams that ran the job through Batch on 2.5 paid half of the standard figure, about $0.0059 per image.
........
Setup (medium quality, 280-token prompt) | Output tokens per image | Cost per image | Cost per 2,000 images | Cost per 30 days |
GPT-Image-2.5, standard | 344 | $0.0117 | $23.44 | $703.20 |
GPT-Image-2.5, Batch | 344 | $0.0059 | $11.72 | $351.60 |
GPT-Image-2.6, standard | 1,136 (forum test) | $0.0355 | $70.96 | $2,128.80 |
........
Under these assumptions, moving a medium-quality workload from GPT-Image-2.5 to GPT-Image-2.6 roughly triples the bill at standard rates and multiplies it by about six for a team that used Batch. At high quality, where the forum test found similar token counts, the two generations should cost about the same per image at standard rates. The change therefore favours interactive, high-quality work such as design tools and editing flows that need Flare's speed or Sunburst's consistency, and it penalizes offline, high-volume catalogue generation at low or medium quality, which loses both the cheaper token profile and the Batch discount. Developers can keep GPT-Image-2.5 for those jobs, since both 2.5 models remain listed, with their Batch rates, on OpenAI's pricing page.
··········
GPT-IMAGE-2.6 PRICING AGAINST GOOGLE NANO BANANA 2.1 AND NANO BANANA 2.
Google's Gemini API pricing page, checked on the same date, lists the closest rivals with per-image token counts, which makes the comparison easier to read. Gemini Nano Banana 2.1 charges $1.50 per 1M input tokens and $30 per 1M image output tokens, the same output rate as OpenAI, and counts 1,120 tokens for a 1K image, 1,680 for 2K and 3,780 for 4K, which Google converts to $0.0336, $0.0504 and $0.113 per image. Batch halves those prices to $0.0168, $0.0252 and $0.0567. Gemini 3.1 Flash Image, sold as Nano Banana 2, charges $0.50 per 1M input tokens and $60 per 1M image output tokens, or $0.067 for a 1K image at standard rates and $0.034 through Batch.
........
Model | Input per 1M | Image output per 1M | Batch | Published per-image price |
OpenAI gpt-image-2.6-sunburst / flare | $5.00 text, $8.00 image | $30.00 | Not supported | Not published |
OpenAI gpt-image-2.5-sunburst / flare | $5.00 text, $8.00 image | $30.00 | $15.00 output | Not published |
Google Nano Banana 2.1 | $1.50 | $30.00 | $15.00 output | $0.0336 (1K), $0.0504 (2K), $0.113 (4K) |
Google Nano Banana 2 (gemini-3.1-flash-image) | $0.50 | $60.00 | $30.00 output | $0.067 (1K), $0.101 (2K), $0.151 (4K) |
........
The comparison is not like-for-like. OpenAI meters quality and complexity rather than fixed resolutions, Google publishes token counts per resolution, and neither company has published a shared quality benchmark between these models. On the forum figure of 1,136 output tokens, a medium-quality GPT-Image-2.6 image would cost close to a 1K Nano Banana 2.1 image at standard rates, while Google keeps a Batch tier that OpenAI has withdrawn for its newest image models.
··········
WHAT CHANGES FOR CHATGPT USERS AND OPENAI API DEVELOPERS.
In ChatGPT, OpenAI's announcement presents Images 2.6 as a rollout to ChatGPT and Codex users, with two visible additions. Animated GIFs can be created from a prompt or from an uploaded image, and users can adjust the speed, edit individual frames and loop the result. Reference images let people save likenesses of themselves, their pets or recurring characters and reuse them in later images. OpenAI has not published which plans get GIFs or reference images first, how long a GIF can be, how many frames it may contain or how many animations each plan allows, so free and paid users should expect plan-specific limits to appear in the app before they appear in documentation. Our ChatGPT image generation guide covers the general workflow, and our report on ChatGPT Images 2.5 and the Flare and Sunburst models explains the September generation that 2.6 replaces in ChatGPT.
For API developers, the switch is a model ID change on the Image API: gpt-image-2.6-sunburst or gpt-image-2.6-flare in place of the 2.5 IDs. Applications that call image generation through the Responses API or Chat Completions cannot use the 2.6 models today, because both endpoints are listed as unsupported. Jobs that rely on Batch should stay on GPT-Image-2.5 until OpenAI adds 2.6 support, and teams that need reproducible output may prefer the dated gpt-image-2.5-sunburst-2026-09-08 and gpt-image-2.5-flare-2026-09-08 snapshots, because the 2.6 models have no pinned version yet.
Before migrating a production workload, the practical step is to measure. Running the same prompt set on both generations at the quality level the product uses, and reading the output token counts returned in each response, gives the real cost ratio for that workload. Setting complexity to low or medium for narrow edits may reduce latency and unwanted changes, although OpenAI has not said whether complexity affects the number of billed tokens. Rate limits for 2.6 also need checking in each organization's limits page: the 2.6 model pages publish no default table, while GPT-Image-2.5 lists 20 images per minute at the Build tier, 150 at Launch and 250 at Grow.
··········
GPT-IMAGE-2.6 SUNBURST AND FLARE: CONFIRMED FACTS AND OPEN QUESTIONS.
The verified facts come from OpenAI's API changelog and model pages. GPT-Image-2.6 Sunburst and Flare launched on November 4, 2026, priced at $5 per 1M text input tokens, $8 per 1M image input tokens and $30 per 1M image output tokens, with cached input at $1.25 and $2. Both support complexity control, markup-guided edits, SVG output and animated GIF output, both run only on the Image API's generation and edit endpoints, and neither supports Batch. The speed figures, 50% faster generation and images in as little as six seconds for Flare, are OpenAI's claims and have not been tied to a published methodology.
The open questions are mostly about cost. OpenAI has not published output token counts or per-image prices for 2.6, the main pricing page did not yet list the models on November 5, and the only public measurement of higher token use at low and medium quality comes from one developer's test. Until OpenAI documents those numbers or adds Batch support, GPT-Image-2.6 is the faster and more capable choice for interactive editing, while GPT-Image-2.5 remains the cheaper option for large offline image jobs.
··········
FOLLOW US FOR MORE.
·····
DATA STUDIOS
·····
datastudios.org
