GPT-Image-2.6 vs GPT-Image-2.5: Complete Comparison and Report on Pricing, Batch Discounts, GIF and SVG Output, and Arena Rankings
GPT-Image-2.6 is the image generation and editing family that OpenAI released in its API on November 4, 2026, in two variants, GPT-Image-2.6 Sunburst for detailed visual work and GPT-Image-2.6 Flare for fast iteration. GPT-Image-2.5 is the family it follows, released on September 8, 2026 with the same Sunburst and Flare split, and it remains on sale with no deprecation notice on OpenAI's deprecations page. The two generations are close relatives, which makes the choice between them a question of price structure, endpoints and new output types rather than of a different model philosophy.
The verdict from OpenAI's own pages is clear on cost. Both families carry identical standard rates of $5 per 1M text input tokens, $8 per 1M image input tokens and $30 per 1M image output tokens, yet only GPT-Image-2.5 is listed for the Batch API, where its image output drops to $15 per 1M tokens, half the standard rate. GPT-Image-2.6 adds an adjustable complexity setting, SVG output and animated GIF output, while markup-guided editing arrived for both generations on the same day. On independent measurements, GPT-Image-2.5 Sunburst at max quality holds first place on both Artificial Analysis image leaderboards (1223 Elo for text-to-image, 1199 for editing), and no GPT-Image-2.6 variant has been listed there yet.
For high-volume work that can wait for asynchronous results, such as catalog renders or overnight asset generation, GPT-Image-2.5 through Batch is the cheaper and better-measured choice. For products that need GIF animation, vector SVG files or control over how much the model changes during an edit, GPT-Image-2.6 is the only option of the two. Prices, endpoints and leaderboard positions in this comparison were verified on OpenAI's official pages and on Artificial Analysis on November 5, 2026.
··········
GPT-IMAGE-2.6 AND GPT-IMAGE-2.5 RELEASE TIMELINE AND MODEL IDENTITY.
OpenAI's API changelog dates the GPT-Image-2.5 release to September 8, 2026, when Sunburst and Flare arrived for image generation and editing through the Image API and the Responses API image generation tool. The same entry introduced the xhigh and max quality settings and stated that both models use the GPT Image 2 token rates. The GPT-Image-2.6 entry is dated November 4, 2026 and lists adjustable complexity, SVG generation and animated GIF generation as the new capabilities, with markup-guided editing added for both GPT Image 2.5 and 2.6 at the same time.
The snapshot structure differs. Each GPT-Image-2.5 variant has an undated alias and a dated snapshot, gpt-image-2.5-sunburst-2026-09-08 and gpt-image-2.5-flare-2026-09-08, so developers can pin a fixed version. The GPT-Image-2.6 model pages list a single identifier per variant, gpt-image-2.6-sunburst and gpt-image-2.6-flare, with no dated snapshot published so far. For teams that need reproducible outputs across months, the presence of a dated snapshot on the older family is a practical difference.
........
Item | GPT-Image-2.6 Sunburst | GPT-Image-2.6 Flare | GPT-Image-2.5 Sunburst | GPT-Image-2.5 Flare |
API release | Nov 4, 2026 | Nov 4, 2026 | Sep 8, 2026 | Sep 8, 2026 |
Model ID | gpt-image-2.6-sunburst | gpt-image-2.6-flare | gpt-image-2.5-sunburst | gpt-image-2.5-flare |
Dated snapshot | None listed | None listed | 2026-09-08 | 2026-09-08 |
OpenAI description | Detailed visual work | Fast iteration | Editing precision | Fast everyday generation |
Inputs / output | Text, image / image | Text, image / image | Text, image / image | Text, image / image |
Deprecation status | None | None | None | None |
........
OpenAI's image prompting guide gives the clearest description of the 2.5 pair: Flare is the small model optimized for speed, with image quality comparable to GPT Image 2, and Sunburst is the base model with higher image quality than GPT Image 2. The 2.6 model pages repeat the same division of labor in different words, describing Sunburst as suited to detailed compositions and image editing workflows and Flare as intended for interactive generation and editing.
··········
GPT-IMAGE-2.6 VS GPT-IMAGE-2.5 PRICING PER 1M TOKENS.
At standard rates the two generations cost exactly the same. OpenAI's pricing page lists all four models at $5.00 per 1M text input tokens, $1.25 for cached text input, $8.00 per 1M image input tokens, $2.00 for cached image input and $30.00 per 1M image output tokens. Text output is not billed because the models return images. The GPT-Image-2.5 model pages add that these token rates match GPT Image 2.
The difference appears in the second price table. GPT-Image-2.5 Sunburst and Flare are listed there at half the standard rates: $2.50 for text input, $0.625 for cached text input, $4.00 for image input, $1.00 for cached image input and $15.00 per 1M image output tokens, the same discount that GPT Image 2 receives. GPT-Image-2.6 appears only in the standard table, and both of its model pages list the Batch endpoint as not supported.
........
Rate (USD per 1M tokens) | GPT-Image-2.6 standard | GPT-Image-2.5 standard | GPT-Image-2.5 Batch |
Text input | $5.00 | $5.00 | $2.50 |
Cached text input | $1.25 | $1.25 | $0.625 |
Image input | $8.00 | $8.00 | $4.00 |
Cached image input | $2.00 | $2.00 | $1.00 |
Image output | $30.00 | $30.00 | $15.00 |
Batch available | No | Yes | Yes |
........
The pricing page publishes no Flex or Priority tier for any GPT Image model, and it does not publish per-image prices for GPT-Image-2.5 or GPT-Image-2.6 by quality and size. Per-image tables exist only for older models such as chatgpt-image-latest, so the cost of a single 2.6 image depends on how many output tokens a request produces, a figure OpenAI has not yet documented for the new family. Our earlier report on the GPT-Image-2.6 launch price list covers the standard rates in more detail.
··········
BATCH API ACCESS AND ENDPOINTS FOR GPT-IMAGE-2.6 AND GPT-IMAGE-2.5.
The endpoint lists on the model pages explain the pricing gap. GPT-Image-2.5 Sunburst and Flare are supported on three endpoints: Batch, image generation and image edit. GPT-Image-2.6 Sunburst and Flare are supported on two, image generation and image edit, with Batch marked as not supported. Both families mark Chat Completions, Realtime, Assistants and the audio endpoints as not supported.
A second difference concerns the Responses API. The GPT-Image-2.5 pages state that the models can be selected directly in the Image API or as the model for the Responses API image generation tool, and the September changelog entry describes the same two routes. The GPT-Image-2.6 changelog entry names only v1/images/generations and v1/images/edits, and the image generation guide says that complexity control, markup-guided edits, SVG output and animated GIF output are available through the Image API. OpenAI's pages fetched for this comparison do not state whether GPT-Image-2.6 can be chosen inside the Responses API image generation tool.
........
Endpoint or route | GPT-Image-2.6 | GPT-Image-2.5 |
v1/images/generations | Supported | Supported |
v1/images/edits | Supported | Supported |
v1/batch | Not supported | Supported, half price |
Responses API image tool | Not stated | Documented |
Chat Completions, Realtime | Not supported | Not supported |
Streaming partial images | 0 to 3 (guide) | 0 to 3 (guide) |
........
For developers who generate images asynchronously, the Batch restriction is the single largest commercial difference between the two families, because it doubles the effective output rate of any workload that GPT-Image-2.5 could have run through Batch.
··········
GPT-IMAGE-2.6 COMPLEXITY CONTROL, SVG AND ANIMATED GIF OUTPUT.
The new capabilities are concentrated in three areas. The first is a complexity setting, sent as the complexity request field with the values low, medium and high, which defaults to high on both GPT-Image-2.6 variants. OpenAI's guide describes it as separate from the quality setting, warns that higher complexity can increase latency and introduce unwanted changes, and advises comparing results for each task. Adjusting complexity has no separate fee, although any change in output tokens would still be billed at the token rates.
The second and third areas are new output types. The changelog lists SVG generation and animated GIF generation for GPT-Image-2.6, and the model pages refer to the image generation guide for details. The public API reference for the Create image method, as fetched on November 5, still lists only png, jpeg and webp as output formats and names only GPT-Image-2.5 and older models, so the parameters for requesting SVG or GIF files had not yet appeared in the reference documentation on that date. The 2.6 model pages also mention that GIF and SVG outputs are part of the family without giving frame counts, duration limits or file size limits.
........
Capability | GPT-Image-2.6 | GPT-Image-2.5 |
Complexity setting | low, medium, high (default high) | Not available |
SVG output | Yes | No |
Animated GIF output | Yes | No |
Markup-guided edits | Yes (Nov 4) | Yes (added Nov 4) |
Mask-based edits | Not specified on 2.6 pages | Yes, inpainting |
Transparent background | Not specified on 2.6 pages | Yes, PNG or WebP |
Fee for the new controls | None for complexity | Not applicable |
........
Markup-guided editing is the one new feature that does not separate the generations, since the November 4 changelog entry adds it to GPT Image 2.5 and 2.6 together. A team that wants to draw on an image to direct an edit can therefore stay on GPT-Image-2.5 and keep the Batch discount.
··········
QUALITY SETTINGS, SIZES AND OUTPUT FORMATS ON GPT-IMAGE-2.5 AND 2.6.
The GPT-Image-2.5 documentation is detailed. Both variants accept the quality values low, medium, high, xhigh, max and auto, with auto as the default, and the xhigh and max levels are exclusive to the 2.5 family among GPT Image models; earlier models top out at high. Custom sizes are allowed as WIDTHxHEIGHT strings when width and height are multiples of 16, the aspect ratio stays between 1:3 and 3:1, no edge exceeds 3,840 pixels and the total pixel count falls between 655,360 and 8,294,400. Resolutions above 2560x1440 are labeled experimental, and the maximum supported resolution is 3840x2160.
The GPT-Image-2.6 pages do not list quality values, sizes, output compression or background options. The image generation guide says only that quality, size, format and compression can be customized on both APIs, without stating how 2.6 differs from 2.5 on those parameters.
........
Parameter | GPT-Image-2.5 (documented) | GPT-Image-2.6 (documented) |
Quality values | low, medium, high, xhigh, max, auto | Not listed on model pages |
Recommended sizes | 1024x1024, 1536x1024, 1024x1536 | Not listed |
Custom size rules | Multiples of 16, 1:3 to 3:1, max 3840x2160 | Not listed |
Raster formats | PNG (default), JPEG, WebP | Not listed |
Images per request (n) | 1 to 10 | Not listed |
Moderation values | auto (default), low | Not listed |
........
For a developer migrating prompts, the 2.5 parameter set is fully specified today, and the 2.6 parameter set has to be tested request by request against the live API.
··········
RATE LIMITS AND ORGANIZATION VERIFICATION FOR OPENAI IMAGE MODELS.
GPT-Image-2.5 Sunburst and Flare share a published rate limit table. The Free tier is not supported, the Build tier allows 250,000 tokens per minute and 20 images per minute, the Launch tier 3,000,000 tokens per minute and 150 images per minute, and the Grow tier 8,000,000 tokens per minute and 250 images per minute. Higher tiers are not shown on the pages.
The GPT-Image-2.6 pages publish no table. They state that limits vary by usage tier and account and direct developers to the limits page of their organization. The image generation guide adds, for all GPT Image models, that API Organization Verification may be required in the developer console before use.
........
Usage tier | GPT-Image-2.5 TPM | GPT-Image-2.5 IPM | GPT-Image-2.6 |
Free | Not supported | Not supported | Not published |
Build | 250,000 | 20 | Not published |
Launch | 3,000,000 | 150 | Not published |
Grow | 8,000,000 | 250 | Not published |
........
At 20 images per minute on the Build tier, an account can produce at most 1,200 GPT-Image-2.5 images per hour through the synchronous Image API, which is one reason high-volume jobs on that tier tend to move to Batch, the route GPT-Image-2.6 does not offer.
··········
[[ADPLUS_300x250_1]]
HOW THE ARTIFICIAL ANALYSIS IMAGE RANKINGS ARE MEASURED.
The independent figures in this comparison come from Artificial Analysis, which runs two blind-preference arenas for image models. In AA-Image-T2I v2.1, people compare text-to-image outputs from two anonymous models for the same prompt; in AA-Image-Editing v2.1, they compare edits of the same source image. The votes produce Elo ratings with 95% confidence intervals, and the organization publishes the number of comparisons behind each score. Prices on the leaderboard are per 1,000 images on the creator's API at 1024x1024 with default settings.
Two limits apply to the reading. Artificial Analysis lists the GPT-Image-2.5 entries only at the max quality setting, so the rankings say nothing about the low, medium or high settings that most production traffic uses. The text-to-image arena on LMArena, the other public image arena, still showed gpt-image-1 as its newest OpenAI entry on a page last updated in October 2025, so it offers no data on either family compared here.
··········
ARTIFICIAL ANALYSIS RANKINGS FOR GPT-IMAGE-2.5 AND THE MISSING 2.6 ENTRIES.
On the text-to-image leaderboard as fetched on November 5, 2026, GPT-Image-2.5 Sunburst (max) ranks first with 1223 Elo (confidence interval 1213 to 1233) from 13,070 comparisons, and GPT-Image-2.5 Flare (max) ranks second with 1207 Elo (1197 to 1217) from 12,418 comparisons. GPT Image 2 at high quality follows in third place with 1189. On the editing leaderboard, Sunburst (max) is first with 1199 Elo (1188 to 1210) from 18,300 samples, and Flare (max) is third with 1183 (1172 to 1194), behind MAI-Image-2.7-Preview at 1196.
........
Leaderboard (Artificial Analysis) | GPT-Image-2.5 Sunburst (max) | GPT-Image-2.5 Flare (max) | GPT-Image-2.6 |
Text-to-image rank | 1 | 2 | Not listed |
Text-to-image Elo (95% CI) | 1223 (1213-1233) | 1207 (1197-1217) | Not listed |
Text-to-image comparisons | 13,070 | 12,418 | Not listed |
Editing rank | 1 | 3 | Not listed |
Editing Elo (95% CI) | 1199 (1188-1210) | 1183 (1172-1194) | Not listed |
Editing samples | 18,300 | 19,199 | Not listed |
Price per 1,000 images | $210.7 | $210.7 | Not listed |
........
Sunburst's 16-point lead over Flare in text-to-image sits at the edge of the two confidence intervals, which overlap only between 1213 and 1217, and the editing gap, also 16 points, rests on intervals that overlap between 1188 and 1194 (1188 to 1210 against 1172 to 1194). Both variants carry the same leaderboard price at max quality, so on Artificial Analysis data the higher-ranked Sunburst costs nothing extra per image. OpenAI itself has published no benchmark comparing GPT-Image-2.6 with GPT-Image-2.5 on the pages that could be fetched, and the ChatGPT Images 2.6 system card on OpenAI's deployment safety site could not be retrieved for this comparison.
··········
COST OF THREE IMAGE WORKLOADS ON GPT-IMAGE-2.5 AND GPT-IMAGE-2.6.
Because OpenAI publishes token rates and no per-image price for either family, the cost example below works from one measured reference point. The Artificial Analysis price of $210.7 per 1,000 GPT-Image-2.5 images at max quality and 1024x1024 implies about 7,023 output tokens per image at $30 per 1M tokens. Each workload adds 300 text input tokens per request for the prompt. The GPT-Image-2.6 column assumes the same token count per image, an assumption that OpenAI's pages neither confirm nor contradict, so it shows the effect of the price structure alone. The arithmetic was run in Python.
........
Workload (max quality, 1024x1024) | GPT-Image-2.5 Batch | GPT-Image-2.5 standard | GPT-Image-2.6 standard (same tokens) |
Cost per image | $0.1061 | $0.2122 | $0.2122 |
Studio, 3,000 images a month | $318.30 | $636.60 | $636.60 |
Catalog, 20,000 images a month | $2,122.00 | $4,244.00 | $4,244.00 |
Images for a $1,000 budget | about 9,425 | about 4,712 | about 4,712 |
........
The catalog workload shows the size of the Batch effect: $2,122 a month separates GPT-Image-2.5 through Batch from either model at standard rates, on identical token counts. Any extra output tokens that GPT-Image-2.6 produces at a given setting, or that the high default complexity might add, would widen that gap, and any savings from the complexity control would narrow it; OpenAI has published no token figures for either case. For interactive products that must return an image within the user session, Batch is not an option for either family, and the two generations then cost the same per token.
··········
SPEED AND LATENCY GUIDANCE FOR SUNBURST AND FLARE.
OpenAI rates speed on the GPT-Image-2.5 pages in relative terms: Sunburst is listed at medium speed with the highest performance rating, and Flare at very fast speed with a higher performance rating. The image generation guide adds that complex prompts may take up to two minutes to process and that JPEG output is faster than PNG. The prompting guide recommends testing Flare when an existing GPT Image 2 workflow already meets its quality requirements, as a way to gain latency, and keeping Sunburst only when its quality advantage is needed.
The GPT-Image-2.6 model pages list both performance and speed as not specified. The guide's only latency statement for the new family concerns complexity, which it says can increase latency at higher settings. Artificial Analysis tracks generation time for GPT-Image-2.5 Sunburst on its model page, but the values are loaded dynamically and were not available in the fetched version, so no measured latency figure for either family appears in this comparison.
··········
CHATGPT IMAGES 2.6 ACCESS, DEPRECATIONS AND MIGRATION PATHS.
Outside the API, the GPT-Image-2.6 launch coincides with ChatGPT Images 2.6, the version of the image feature inside ChatGPT. The ChatGPT pricing documentation fetched on November 5 states that image generation is not available on the Free plan, lists Plus at $20 a month, Pro at $100, $200 or $500 a month and Business at $20 per user per month billed annually or $25 billed monthly, and says image generations consume included limits about 3 to 5 times faster on average than comparable turns without images. It does not name the Images 2.6 or 2.5 models by plan, and OpenAI's release notes page could not be retrieved, so per-plan availability of Images 2.6 is not confirmed here. Our guide to ChatGPT image generation, plan limits and API prices tracks those plan details.
On the API side, OpenAI's deprecations page still points migrating developers to the 2.5 family. gpt-image-1 was shut down on October 23, 2026, and gpt-image-1.5 and chatgpt-image-latest are both scheduled for shutdown on December 1, 2026, each with gpt-image-2.5-sunburst or gpt-image-2.5-flare as the recommended replacement. The chatgpt-image-latest page, which describes itself as the previous image model used in ChatGPT, also recommends GPT-Image-2.5 Sunburst for API use.
........
Model | Status on Nov 5, 2026 | Shutdown date | Replacement named by OpenAI |
gpt-image-1 | Shut down | Oct 23, 2026 | GPT-Image-2.5 Sunburst or Flare |
gpt-image-1.5 | Deprecated | Dec 1, 2026 | GPT-Image-2.5 Sunburst or Flare |
chatgpt-image-latest | Deprecated | Dec 1, 2026 | GPT-Image-2.5 Sunburst or Flare |
gpt-image-2 | Active | None | Not applicable |
GPT-Image-2.5 Sunburst, Flare | Active | None | Not applicable |
GPT-Image-2.6 Sunburst, Flare | Active | None | Not applicable |
........
Developers who must leave gpt-image-1.5 or chatgpt-image-latest before December 1 therefore have an officially named target in GPT-Image-2.5, with Batch support and dated snapshots, and can treat GPT-Image-2.6 as a second step for the requests that need its new outputs. The launch of the previous generation is covered in our report on ChatGPT Images 2.5 and the GPT-Image-2.5 Flare and Sunburst models, and the cross-vendor view is in our GPT-Image-2.6 vs Nano Banana 2.1 comparison.
··········
GAPS IN THE GPT-IMAGE-2.6 DOCUMENTATION AND EVALUATIONS.
Several figures that would settle the comparison are absent from every official and independent page fetched on November 5. OpenAI has published no output token counts per image for GPT-Image-2.6 at any quality or complexity setting, no per-image price table, no rate limits, no dated snapshot and no quality or size parameters on the 2.6 model pages. The API reference had not yet added SVG or GIF formats or the complexity field, and the frame, duration and size limits for animated GIFs are not documented.
On the evaluation side, OpenAI's announcement page and the ChatGPT Images 2.6 system card returned errors when fetched, and the model pages contain no benchmark tables. Artificial Analysis has not yet added GPT-Image-2.6 to either arena, and it ranks GPT-Image-2.5 only at max quality. Any claim that GPT-Image-2.6 produces better images than GPT-Image-2.5 therefore rests, as of this date, on OpenAI's product descriptions alone, while the claim that GPT-Image-2.5 Sunburst leads the public text-to-image and editing arenas rests on thousands of blind comparisons.
··········
GPT-IMAGE-2.6 OR GPT-IMAGE-2.5 BY WORKLOAD.
The facts above point to a split decision that follows the workload more than the model generation. Where the output format is a raster image and the job can run asynchronously, GPT-Image-2.5 costs half as much per token through Batch and carries the best-documented independent rankings. Where the product needs animation, vector files or control over how much an edit changes, GPT-Image-2.6 is the only family of the two that offers it, at the same standard rate as 2.5.
........
Workload | Recommended model | Reason |
Catalog or overnight asset batches | GPT-Image-2.5 Sunburst via Batch | $15 per 1M output tokens vs $30 |
Highest-ranked image quality today | GPT-Image-2.5 Sunburst (max) | First on both AA arenas |
Animated GIFs for social or ads | GPT-Image-2.6 Flare or Sunburst | Only family with GIF output |
Logos, icons, scalable graphics | GPT-Image-2.6 Sunburst | Only family with SVG output |
Edits that must change little | GPT-Image-2.6 with lower complexity | Complexity setting, no extra fee |
Draw-to-edit workflows on a budget | GPT-Image-2.5 with markup edits | Markup added to 2.5, Batch kept |
Migration off gpt-image-1.5 by Dec 1 | GPT-Image-2.5 Sunburst or Flare | Replacement named by OpenAI |
Pinned, reproducible versions | GPT-Image-2.5 dated snapshot | 2026-09-08 snapshot available |
........
A team generating 20,000 catalog images a month at max quality spends about $2,122 on GPT-Image-2.5 through Batch against about $4,244 at standard rates on either family, so it should keep that volume on GPT-Image-2.5 and route only its GIF, SVG and low-complexity edit requests to GPT-Image-2.6 through the synchronous Image API.
··········
FOLLOW US FOR MORE.
·····
DATA STUDIOS
·····
datastudios.org
