GPT Image 2.6 Guide: GIF, SVG, Pricing & What's New
A sourced guide to GPT Image 2.6: how to output GIFs and SVGs, why official token rates are not the same as cost per image, and what changed since 2.5.
The quick answer. GPT Image 2.6, released on 4 November 2026, adds native GIF output, SVG output, a complexity setting, and — in ChatGPT — saved named reference images. The API models are gpt-image-2.6-flare (speed) and gpt-image-2.6-sunburst (detail and edit consistency). Per-token prices match GPT Image 2.5 exactly. Per-image prices do not: at low and medium quality, the official calculator bills about 4.2× and 3.3× the output tokens of 2.5 for the same 1024×1024 image.
VSModels has not run its own generations for this guide. The facts below come from official documentation re-read on 10 November 2026 (UTC+8) and from dated third-party tests. Many of those tests are a single take or a handful of prompts, and several come from platforms that sell this model or a rival. Treat them as evidence, not a verdict. For the generation-to-generation comparison, see GPT Image 2.6 vs 2.5: What Changed and Should You Switch?.
What GPT Image 2.6 is
OpenAI's product name is Images 2.6 (also written ChatGPT Images 2.6). The community name, and the one this page uses, is GPT Image 2.6. The launch post is dated 4 November 2026 and places the release in ChatGPT, Codex, and the API, on every ChatGPT plan (Introducing Images 2.6).
In the API there are two models, with the same token rates:
gpt-image-2.6-flare— the fast tier. OpenAI says that atquality: "low"it can finish in as little as six seconds (model page; launch post).gpt-image-2.6-sunburst— the detail tier, aimed at richer compositions and more consistent edits at medium or high quality (model page).
The immediate predecessors are GPT Image 2.5 Sunburst and Flare (September 2026), and before that gpt-image-2 (April 2026). Both 2.6 models support only the Image API: POST /v1/images/generations and POST /v1/images/edits. The model pages list Batch and the Responses API as not supported — a change from 2.5, which still has a half-price Batch row on the pricing page.
ChatGPT does not show which model or quality setting produced a given image, which is why ChatGPT-path speed tests are hard to compare with API timings. That observation comes from users on Hacker News and aiformortals; it is a community report, not an official statement.
What's actually new — and what isn't
New in 2.6, with official documentation behind it:
- Animated GIF output on both
/v1/images/generationsand/v1/images/edits. You can guide the animation with still reference images on the edits endpoint. Older briefs that said GIFs were generations-only are out of date; the image generation guide is clear on this. - SVG output, generations only,
nmust be 1. An SVG cannot be sent back as an edit input, and the edits endpoint cannot emit SVG. complexity:low,medium, orhigh, defaulthigh, no separate fee. This is planning effort, not the same thing asquality, which still controls rendering fidelity (low,medium,high,xhigh,max, orauto).- Saved reference images in ChatGPT, so you can reuse a person, pet, or character without uploading the file every time.
Not new, or easy to over-claim:
- Markup-guided edits — circling or sketching on a transparent overlay — are documented for supported GPT Image 2.5 and 2.6 models, not only 2.6 (changelog; the guide's markup example).
- There is no API parameter that "locks" a character. Consistency is a prompting-plus-references workflow. The Image API accepts up to 16 input images on edits (edit reference).
- OpenAI's "generates images 50% faster than before" is a vendor claim with no published baseline (launch post). Third-party API tests find a large speed-up for Flare and, at default
complexity, a slowdown for Sunburst. Details are in the speed section below and in the vs 2.5 comparison.
How to generate a GIF
Set output_format to "gif" on either generations or edits. Describe the subject, the motion, and the loop in the prompt. To start from stills, send those stills to the edits endpoint.
Documented controls, from the image generation guide as of 10 November 2026:
frame_count: integers from 2 to 16, default 9. Earlier write-ups that cited a 64-frame ceiling are stale. CellCog, a studio that sells its own image agent, reported that asking for 32 frames returned HTTP 400, with the API message "Use an integer from 2 to 16, or omit frame_count to use 9."frame_duration_ms: 20 to 2,000 milliseconds per frame, rounded to 10 ms, default 200.size:auto, or aWIDTHxHEIGHTstring whose edges are multiples of 16.sizedescribes each finished frame. GIF output does not support streaming, partial images, or customoutput_compression.
What third parties have seen, in small tests:
- Simple, single-character, fixed-camera loops are the use case that works. CellCog published a 12-frame mascot wave from one text prompt on Flare (11 seconds, $0.056 on their bill — their measurement, not an official rate card). They said the character held glasses and hoodie across all twelve frames (CellCog).
- Complex action and real-photograph animation are weaker. PhotoWorkout reported that, through their ChatGPT-login setup, a cartoon character was drawn frame by frame, but a request to animate a real photo instead pasted AI-drawn leaves over the untouched original. Users on Hacker News described mottling and flicker. These are short, informal tests, not a benchmark.
How to generate an SVG
Set output_format to "svg" on POST /v1/images/generations only. n must be 1. There is no streaming, no partial images, and no output_compression. The API first creates a raster reference image — quality applies to that reference — and then writes SVG from it. You decode data[0].b64_json as base64 and save the UTF-8 XML with a .svg extension. The document is static: no scripts, no external resources, no embedded rasters (guide).
Two limits matter in production:
- The edits endpoint will not output SVG, and an SVG is not a supported raster input for later edits. You cannot generate a logo as SVG and then mark it up or inpaint it through the same Image API path.
- Early reliability was poor on anything with lettering. CellCog's SVG test — again, a commercially interested studio, one published run — reported 9 successful files out of 48 requests. Prompts with lettering went 3 for 41; text-free prompts went 6 for 7. Failures were HTTP 502 after 11 to 91 seconds. Recraft V4.1, in the same write-up, returned 9 of 9 at a flat $0.08 per SVG. A separate OpenAI forum thread describes 500/502 errors on poster prompts while simple icons succeeded. Some users say an undocumented
svg.modelparameter is required; that is not in the official guide, so treat it as unresolved.
When a file does come back, CellCog found that GPT Image 2.6 kept lettering as editable text and produced files 3 to 17 times smaller than Recraft's, with richer scene and mascot drawings. That is a useful property if the request succeeds. It is not a reason to call native SVG production-ready on launch-week data.
Reference images and character consistency
There is no "lock character" field in the API. What exists is:
- In ChatGPT, saved named reference images. OpenAI's community announcement puts it as "Save once. Create again and again" (community post). TechRadar described saving a cat photo as "Chester" and later prompting with
@Chester. - In the API, up to 16 images on the edits endpoint, plus the prompt. That 16-image cap is the same order of limit 2.5 already had.
Independent and platform tests have not shown a dramatic consistency jump over 2.5. Flixly — a platform that sells these models — ran four prompts and put 2.6 Sunburst at or near the top, slightly ahead of 2.5 on character consistency in a related write-up, with Nano Banana 2.1 and Pro matching on instruction-following and finishing faster. That is four prompts, first take, on a seller's stack. For how Nano Banana 2.1 and Pro already compare with each other, see Nano Banana 2.1 vs Pro, Nano Banana 2.1 vs 2, and Nano Banana 2 vs Pro.
Markup-guided edits
Markup is an Image API edits feature for supported 2.5 and 2.6 models. You keep the original as the first input image, attach a separate transparent PNG of the same width and height after orientation, and send that overlay as markup. The prompt describes the change and what must stay put. Do not flatten the annotations onto the photo, and do not combine markup with mask. Each of the first image and the overlay may be at most 16,777,216 pixels; the overlay also has a 50 MB file cap (guide).
Because markup already works on 2.5, it is not a reason by itself to move a working 2.5 edit pipeline.
Pricing: same token rates, not the same cost per image
Re-read on 10 November 2026 (UTC+8) from the pricing page and both model pages, standard processing, per 1 million tokens, identical for Flare and Sunburst and identical to 2.5:
- Text input: $5 (cached $1.25)
- Image input: $8 (cached $2)
- Image output: $30
The 2.6 models do not appear in the Batch price table. GPT Image 2.5 Flare and Sunburst still do, at half those rates. Cached-input rates for GPT Image 2 and 2.5 are footnoted as applying only to the Responses API image tool, not to direct Images API calls. 2.6 is not on Responses at all, so treat cached rates on the 2.6 model pages as published list prices, not as a confirmed Images API discount.
The sentence that matters is in the official guide: "Equal token rates don't mean equal cost per image." Token use can differ by model and quality. The on-page calculator, re-read the same day, defaults to GPT Image 2.6, quality low, 1024×1024, and shows 823 output tokens / $0.02469. That figure is image output only; it excludes prompt tokens, input-image tokens on edits, and the extra 100 output tokens billed for each streamed partial image.
The same official calculator, using OpenAI's published formula, gives these 1024×1024 image-output estimates at $30 per million tokens:
- Low — 2.5: 196 tokens, $0.00588. 2.6: 823 tokens, $0.02469. About 4.2×.
- Medium — 2.5: 439 tokens, $0.01317. 2.6: 1,449 tokens, $0.04347. About 3.3×.
- High — both: 1,756 tokens, $0.05268.
- Xhigh — both: 3,122 tokens, $0.09366.
- Max — both: 7,024 tokens, $0.21072.
The 4.2× and 3.3× ratios are not a third-party inference. They are the multipliers the official calculator applies to the 2.5 estimate for 2.6 at low and medium. High and above are unmultiplied.
Measured API usage is very close, not always identical. CellCog ran the same 1024×1024 prompt across qualities (they say every low cell repeated exactly) and recorded 824 tokens on 2.6 low, versus 196 on 2.5, and 1,758 tokens at high, versus the calculator's 1,756. OmniaKey and users on Hacker News reported 1,449 tokens at medium versus 439 on 2.5, which matches the calculator. Do not use older comparison pages that still price 2.6 low at 2.5's $0.006; those estimates are stale.
SVG billing is separate and variable: you pay the image model's rate for the reference raster and the authoring model's tokens for the SVG itself (guide). There is no official flat "price per SVG." CellCog's successful files ranged from about $0.03 to $0.23 on their invoice; that is their bill, on launch-week traffic, not a rate you should budget as official.
Quality, speed, and conflicting scoreboards
Quality boards should not be mixed. As of 10 November 2026, Artificial Analysis ranks models at max quality. GPT Image 2.6 Sunburst leads text-to-image at 1259 Elo, ahead of 2.6 Flare at 1228 and 2.5 Sunburst at 1227; it also leads the editing board at 1209, with 2.6 Flare at 1206 and 2.5 Sunburst at 1199. AA's listed price of $210.72 per 1,000 images is exactly the official calculator's max 1024×1024 output cost.
Arena's text-to-image board, in a snapshot from 6 November 2026, ran 2.6 at Medium and marked the 2.6 rows preliminary (about 2,972 votes for 2.6 Sunburst). In that snapshot, 2.5 Sunburst led at 1429, with 2.6 Sunburst (Medium) at 1409. That is the opposite direction from AA, on a different quality setting, with a much smaller 2.6 sample. DesignArena snapshots in the same week disagreed with one another and are not usable as a single ranking. None of these boards prove that "2.6 is better than 2.5" in general.
Text is already strong on both generations. Fontly ran nine typographic tests and reported 75 images for $1.79, including Thai, Arabic, Devanagari, stacked Vietnamese marks, and a 120-word poster at low quality, with no misspelled word from either Flare or Sunburst. They are clear that most conditions had only one or two images per model. PhotoWorkout, BulkImagen, and PixelDojo likewise saw posters, menus, and a 29-string settings sheet come back correctly — and PixelDojo and BulkImagen sell access to these models. Dense comic-style grids are a thinner, mixed data point and should not be generalised from a few published frames.
Unrequested extras are the more consistent complaint: slogans, props, or addresses the prompt did not ask for. Kayaweb — a self-funded API test of 61 images for $2.05 — reported that 6 of 10 Sunburst French posters added text that was never requested. BulkImagen and PixelDojo describe the same habit. DesignArena commentary in secondary write-ups claimed Flare added fewer extra slogans; that conflict is unresolved, and the quality settings are not aligned.
Speed depends on which model and which path you use:
- Flare, on the API, is clearly faster in the published tests. CellCog recorded 2.6 Flare at 18.0 seconds versus 45.1 for 2.5 Flare at max, and about 2× at every quality they timed (80 generations, one machine, one evening). Kayaweb timed Flare high at 10 seconds versus 16.8 for 2.5, averaging three runs and including network latency.
- Sunburst, at default
complexity=high, was slower than 2.5 Sunburst in those same tests: CellCog reported 17% and 27% slower on two prompts; Kayaweb timed high at 37.5 seconds versus 24.5, and said droppingcomplexitytolowbrought latency back in line with 2.5 without changing the token bill. - ChatGPT-path tests did not show the advertised speed-up. PhotoWorkout's 2.6 Sunburst averaged 30.1 seconds against 31.9 for 2.5, about 6%, through a ChatGPT login rather than a direct API call.
OpenAI's own guide also warns that higher complexity can increase latency and introduce unwanted changes.
How it sits next to other image models
This page is a 2.6 guide, not a full competitive matrix. A short orientation, with the limits attached:
- Versus 2.5 is the comparison with enough evidence to write carefully: new formats, a real low/medium cost increase, Flare speed, Sunburst default latency, and conflicting boards. That write-up is GPT Image 2.6 vs 2.5.
- Versus Nano Banana 2.1 and Pro, 2.6's distinctive features are GIF, SVG, and markup; Nano Banana 2.1's are 4K, a 14-image reference budget, search grounding, and Batch. Direct same-prompt sets are still small and mostly from seller platforms such as Flixly. Start with the existing cluster: 2.1 vs Pro, 2.1 vs 2, 2 vs Pro, and Nano Banana vs Midjourney.
- Other frequently named rivals — Seedream 5.0 Pro, Midjourney V8.2, Recraft V4.1 (especially for SVG), FLUX 3, Qwen-Image 3.0 — appear in one-off or commercially affiliated tests. They are not scored here.
FAQ
Is GPT Image 2.6 the same price as 2.5? Per token, yes. Per image, only at high, xhigh, and max. At low and medium, official output-token estimates are about 4.2× and 3.3× higher. 2.6 also has no Batch discount.
Can I use GPT Image 2.6 in the Responses API?
Not according to the current Flare and Sunburst model pages, which list Responses and Batch as not supported. Use /v1/images/generations and /v1/images/edits.
What's the difference between Flare and Sunburst? Same price, different jobs. Flare is the fast everyday model; Sunburst is the detail and edit-consistency model. Third-party API timings show Flare is much faster than 2.5 Flare, while Sunburst at default high complexity is often slower than 2.5 Sunburst.
Can 2.6 lock a character from a reference photo?
Not with a dedicated API flag. In ChatGPT you can save a named reference and recall it (TechRadar's @Chester example). In the API you send up to 16 reference images on edits and describe what must stay consistent.
Does markup-guided editing require 2.6? No. Official docs say supported GPT Image 2.5 and 2.6 models.
Are the GIFs and SVGs production-ready? GIFs are usable for short, simple, single-character loops; they struggle with complex motion and real-photo animation in the published tests. SVG success rates in the largest published trial were low, especially with text, and some requests returned 502. Both findings are early, small, and partly from a commercially affiliated tester.
Is 2.6 better than 2.5? Not as a blanket statement. See the vs 2.5 comparison. AA at max quality ranks 2.6 first; Arena at Medium, in an early snapshot, ranked 2.5 Sunburst ahead. Cost and speed depend on quality and on whether you call Flare or Sunburst.
Sources
Official, re-read 10 November 2026 (UTC+8) unless noted:
- Introducing Images 2.6 (4 November 2026)
- Image generation guide, including the on-page token calculator
- Pricing
- GPT-Image-2.6 Flare and Sunburst model pages
- Images edits API
- API changelog
- ChatGPT Images 2.6 system card
- Community announcement: Images upgraded
Third-party tests and boards (sample sizes are small unless the source is a public leaderboard; several authors sell the model or a competitor):
- Artificial Analysis text-to-image and editing (re-read 10 November 2026)
- Arena text-to-image (snapshot 6 November 2026; 2.6 rows were preliminary)
- CellCog, GPT Image 2.6 vs 2.5 and CellCog, Recraft V4.1 vs GPT Image 2.6 SVG
- Kayaweb, Fontly, PhotoWorkout
- OmniaKey, PixelDojo, BulkImagen, Flixly
- Hacker News discussion, OpenAI forum on SVG 500/502, TechRadar
Gaps
- No VSModels-run generations. Nothing here is our own benchmark, and no per-image price in this guide is a measured bill from our account.
- The official calculator's 823-token low 1024 figure was confirmed on the live guide on 10 November 2026. CellCog's API usage for the same cell was 824 tokens. We did not call the API to reconcile the one-token gap.
- Arena scores are from a 6 November 2026 snapshot and were not re-opened today; 2.6 was still marked preliminary there. DesignArena views from that week were inconsistent and are not used as a ranking.
- SVG's possible
svg.modelrequirement is a user report, not documentation. - ChatGPT's hidden model and quality picker means consumer-app timings should not be treated as API benchmarks.
- Most head-to-head image tests cited here used one take, or only a few prompts, often on a platform that sells the endpoint.