Blog · 2026-11-05

GPT Image 2.6 vs Qwen Image 3.0: GIF, SVG and 4K or a Flat Price

Model pages: GPT Image 2.6 Flare API · GPT Image 2.6 Sunburst API · Qwen Image 3.0 API · Qwen Image 3.0 Pro API — pricing, specs and the playground

GPT Image 2.6 is the pick when a job needs 4K, animated GIF, SVG, mask or markup edits, or more than three reference images; Qwen Image 3.0 is the pick when you want a flat $0.035 per image at 1K or 2K for text-heavy, multilingual layouts.

The deciding fact is how each one bills. GPT Image 2.6 charges per token, so cost follows quality, size, format, references and prompt length; in our 5 November 2026 tests one image came to between $0.0077 (1K low) and $0.0516 (2K medium). Qwen Image 3.0 has a price list you can quote before the first call: $0.035 at either resolution, plus $0.004 per reference image.

Each comes in two tiers on apimodels.app. OpenAI's Flare (gpt-image-2.6-flare) is the default tier and Sunburst (gpt-image-2.6-sunburst) spends more time per image on detail, both on one token price list. Alibaba's standard tier (qwen3-image) is built for volume and Pro (qwen3-image-pro) for fidelity, at $0.037 for 1K and $0.075 for 2K.

GPT Image 2.6 vs Qwen Image 3.0 at a glance

Pricing — GPT Image 2.6 bills image output at $30 per 1M tokens, image input at $8 (cached $2), text input at $5 (cached $1.25) and text output at $10, each request rounded up to $0.0001; SVG is a flat $0.035 per image. Qwen Image 3.0 is $0.035 per image at 1K or 2K, Pro is $0.037 at 1K and $0.075 at 2K, and each reference image adds $0.004 on either tier.

Resolution and sizes — GPT Image 2.6 takes aspect_ratio with 1K, 2K or 4K, an exact WIDTHxHEIGHT (multiples of 16, longest edge up to 3840, ratio up to 3:1), a preset name or auto; in our tests a 16:9 request delivered 1360×768 at 1K and 2720×1536 at 2K. Qwen Image 3.0 outputs 1K or 2K in eight aspect ratios (1:1, 3:2, 2:3, 4:3, 3:4, 16:9, 9:16, 21:9), and no tier goes above 2K.

Inputs and edits — GPT Image 2.6 accepts up to 16 reference images, prompts up to 32,000 characters, and either a markup image or a mask (white marks the area to change, black the area to keep), not both. Qwen Image 3.0 accepts up to 3 reference images in image_urls, each up to 10MB in jpeg, png or webp, and prompts up to 4.5k tokens; it edits when references are present and has no mask parameter.

Output types — GPT Image 2.6 returns png, jpeg or webp with transparent backgrounds available, animated GIF of 2 to 64 frames, and SVG vectors, with up to 10 images per call. Qwen Image 3.0 returns PNG; how many images one request can return is not documented.

Controls — GPT Image 2.6 has quality (low, medium, high or auto, medium by default on apimodels.app) and reasoning_effort (low, medium or high), which trades latency for detail. Qwen Image 3.0 documents neither; tier and resolution are the levers.

Speed — In our 5 November 2026 tests GPT Image 2.6 returned 2K and 4K images in 13 to 25 seconds. On our own measurements Qwen Image 3.0 standard has a median of 73s and Pro a median of 121s.

Limits — GPT Image 2.6 SVG is text-to-image only, one image per request, and all GIF frames share one sheet of at most 3840 px on the longest edge. Qwen Image 3.0 stops at 2K and has no pixel-level mask control.

A price list against a meter

Qwen Image 3.0 is the easier one to budget. A 4.5k-token prompt costs the same as a one-line prompt, 2K costs the same as 1K on the standard tier, and the only extra is $0.004 per reference. A 2K Pro edit with one reference is $0.079, and fusing three 1K references on the standard tier is $0.047.

GPT Image 2.6 meters the work instead. Quality, size, GIF, every reference image and the prompt all change the token count, so there is no fixed per-image price. Our measured requests (16:9, one image) came to about $0.0077 for 1K low, $0.0162 for 1K high, $0.0157 for a 1K low edit with one reference, $0.0283 for 2K low and $0.0516 for 2K medium. They are examples of the range, not a rate card.

Set against the flat price, the 1K low request cost less than $0.035 and the 2K medium request cost more. Since apimodels.app defaults quality to medium and we have no measured price for 1K medium, test your own default before sizing a budget on it.

apimodels.app holds an upper bound when a GPT Image 2.6 request is accepted, releases the rest afterwards, and returns a usage object on completed png, jpeg and webp requests whose charged field shows the amount deducted, so you can log real per-image cost.

Formats and sizes only GPT Image 2.6 has

If the deliverable is a 4K print file, an exact pixel size such as a 3:1 banner, an animated GIF or an SVG logo, Qwen Image 3.0 is out and price never enters the decision.

GIF output is set with output_format gif, frame_count (2 to 64, default 9) and frame_duration_ms (20 to 2000, default 200). All frames sit on one sheet of at most 3840 px on the longest edge. With aspect_ratio plus 1K, 2K or 4K, apimodels.app shrinks the frames so the sheet fits; with an exact size, keep it small yourself, and the whole sheet must stay within 3840 px on the longest edge and 8,294,400 pixels in total.

SVG is text-to-image only, one image per request, at a flat $0.035 whatever the quality. It is a static vector with no scripts, external resources or embedded raster, suited to icons, logos and flat illustrations rather than photographic scenes.

Local edits are the other gap. GPT Image 2.6 takes a markup image or a mask; Qwen Image 3.0 has neither, so you pass the original through image_urls and describe the change in the prompt. On the previous GPT generation masks acted as hints rather than hard boundaries, and 2.6 is not yet measured, so name the region in the prompt with either model.

Text-heavy layouts and languages

Qwen Image 3.0 is built around in-image text: dense layout from prompts up to 4.5k tokens, 10px small text, 12 languages and 20+ native fonts, aimed at newspapers, storyboards, menus, exam sheets, infographics and UI mockups. Pro keeps those abilities and adds micro-expressions, pores and hair strands near photographic realism.

GPT Image 2.6 is described as handling complex layouts, but no language count, font count or minimum text size is documented for it. For a menu in three languages or a poster full of small type, Qwen Image 3.0 is the model whose documentation addresses the job directly. Legible is not the same as character-accurate, though: spelling, prices, formulas and dates still need a human pass before they ship.

Waiting time and polling windows

Both run as async tasks: create one, receive a taskId, poll. In our tests every GPT Image 2.6 request finished in under a minute: about 20 seconds for 1K low, 13 to 25 seconds at 2K and 4K, and 22 to 58 seconds for a 1K low edit with one reference.

Qwen Image 3.0 standard has a median of 73s, with 90% finishing inside 96s and the fastest seen at 36s, so a 3-minute polling window is plenty. Pro has a median of 121s and a long tail: 10% exceed 324s and the slowest seen is 359s. Allow at least 8 minutes for Pro, because cutting the window at the median is the usual cause of a task that looks stuck but is still running. Qwen Image 3.0 also accepts a callBackUrl if you prefer a webhook.

When a person is waiting for the image on screen, this gap weighs more than the price difference; for overnight batches it hardly matters.

What a job costs

Qwen Image 3.0 figures are fixed prices. GPT Image 2.6 figures multiply our measured 5 November 2026 requests (16:9, one image), so treat them as examples. Moderation refusals and upstream errors happen on both models; failed requests are not charged, so these totals count completed images only.

100 social images at 1K

GPT Image 2.6 at 1K low comes to about $0.77, or about $1.62 at 1K high. Qwen Image 3.0 standard is $3.50 and Pro is $3.70.

100 banners at 2K

GPT Image 2.6 at 2K low comes to about $2.83, or about $5.16 at 2K medium. Qwen Image 3.0 standard is $3.50, the same as at 1K, and Pro is $7.50.

100 one-reference edits at 1K

GPT Image 2.6 at 1K low comes to about $1.57. Qwen Image 3.0 standard at $0.039 per edit is $3.90, and Pro at $0.041 per edit is $4.10.

10 SVG icons

GPT Image 2.6 charges $0.35 at the flat SVG price, whatever quality you pass. Qwen Image 3.0 cannot return SVG; 10 flat PNG illustrations on the standard tier also cost $0.35, but they are raster files.

Which one to use

Choose GPT Image 2.6 when any of these applies:

1. The output must be 4K, an exact pixel size, an animated GIF or an SVG.
2. You want to point at an edit with a markup image or a mask, or you need more than three references.
3. A person waits on the result, and the sub-minute turnaround we measured matters.

Use Flare for everyday volume and Sunburst when someone will zoom in; the per-token prices are identical, so the tier changes time per image, not the rate.

Choose Qwen Image 3.0 when the budget must be fixed per image, 2K is enough, and the work is text-dense or multilingual: posters, menus, UI mockups, infographics. Run volume on qwen3-image and send hero shots or portraits to qwen3-image-pro with an 8-minute polling window. The tiers share parameters, so iterating on one and rendering the final frame on the other is a one-field change.

Neither fits a strict pixel-level local edit: Qwen Image 3.0 has no mask, and mask strictness on GPT Image 2.6 is unmeasured. And if you already run GPT Image 2.5 on apimodels.app for its fixed per-image prices, stay there unless you need GIF, SVG, markup edits or exact sizes; 2.5 remains available.

Calling both with one key

One apimodels.app key calls both families on the same endpoint; switching is a change to the model field. Each call below creates a task.

# GPT Image 2.6 Flare: animated GIF, 9 frames of 200 ms
curl https://api.apimodels.app/v1/images/generations \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "gpt-image-2.6-flare", "prompt": "A paper boat bobbing on small waves, flat colors", "aspect_ratio": "1:1", "output_format": "gif", "frame_count": 9, "frame_duration_ms": 200}'
# GPT Image 2.6 Sunburst: high quality close-up
curl https://api.apimodels.app/v1/images/generations \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "gpt-image-2.6-sunburst", "prompt": "Close-up of a leather wallet, visible stitching", "aspect_ratio": "16:9", "quality": "high"}'
# Qwen Image 3.0: text-to-image (no image_urls)
curl https://api.apimodels.app/v1/images/generations \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "qwen3-image", "prompt": "A cafe menu poster in English and Chinese with small printed prices", "aspect_ratio": "3:4"}'
# Qwen Image 3.0 Pro: edit, because image_urls is present
curl https://api.apimodels.app/v1/images/generations \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "qwen3-image-pro", "prompt": "Keep the person, change the background to a sunlit studio", "image_urls": ["https://example.com/portrait.png"], "aspect_ratio": "1:1"}'
# Each call returns a taskId. Poll it with
# GET https://api.apimodels.app/v1/images/generations?task_id=TASK_ID until state is completed or failed.

FAQ