Blog · 2026-11-05

GPT Image 2.6 vs Nano Banana 2.1: GIF, SVG and Masks or a Flat Price

Model pages: GPT Image 2.6 Flare API · GPT Image 2.6 Sunburst API · Nano Banana 2.1 API — pricing, specs and the playground

Choose Nano Banana 2.1 when you want one known price per image, and choose GPT Image 2.6 when the job needs animated GIF, SVG vectors, markup or mask edits, exact pixel sizes or more than 10 reference images.

The deciding fact is how each one is billed. Nano Banana 2.1, Google's image model, costs $0.022 at 1K, $0.036 at 2K and $0.06 at 4K per image on apimodels.app, and neither the aspect ratio, the prompt nor up to 10 reference images changes that. OpenAI's GPT Image 2.6 is billed per token at the same rates as OpenAI's official API, so quality, size, format, references and prompt length all move the price. Per-token is not automatically the larger bill: in our 5 November 2026 tests a 1K low image came to about $0.0077, while a 2K medium image, medium being the default quality on apimodels.app, came to $0.0516.

GPT Image 2.6 comes in two tiers with the same parameters and the same token price list: gpt-image-2.6-flare, the default tier for everyday volume, and gpt-image-2.6-sunburst, which OpenAI positions as spending more time per image on fine detail and on edits that stay scoped across many rounds of revision. Everything below applies to both.

GPT Image 2.6 vs Nano Banana 2.1 at a glance

Pricing — GPT Image 2.6 charges 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. Nano Banana 2.1 charges $0.022, $0.036 or $0.06 per image at 1K, 2K or 4K, with 1K as the default.

Resolution and sizes — GPT Image 2.6 takes aspect_ratio with 1K, 2K or 4K, an exact WIDTHxHEIGHT (both multiples of 16, longest edge up to 3840, ratio up to 3:1, 0.66 to 8.29 megapixels), a preset name or auto. Nano Banana 2.1 takes resolution 1K, 2K or 4K and fifteen aspect ratios, from auto, 1:1 and 16:9 to 21:9, plus 1:4, 4:1, 1:8 and 8:1. At 1K and 16:9, GPT Image 2.6 delivered 1360×768 in our tests and Nano Banana 2.1 delivered 1376×768.

Inputs — GPT Image 2.6 accepts up to 16 reference images, a markup image or a mask (not both), and prompts up to 32,000 characters. Nano Banana 2.1 accepts image for one reference or image_urls for up to 10 in total (URL, data URI or base64; JPEG, PNG or WebP) and prompts up to 20,000 characters; an eleventh reference returns a 400 instead of being dropped. Masks and markup images are not documented for Nano Banana 2.1.

Outputs — GPT Image 2.6 returns png, jpeg or webp, transparent backgrounds included, plus animated GIF and SVG, with up to 10 images per call. Nano Banana 2.1 is one model for text-to-image and editing; its output formats and the number of images per call are not documented.

Controls — GPT Image 2.6 has quality (low, medium, high or auto; medium by default on apimodels.app, while the model's own default is high) and reasoning_effort (low, medium or high), which trades latency for detail. Nano Banana 2.1 is steered by resolution, aspect_ratio and the prompt; no quality setting is documented.

Speed — in our tests a 1K text-to-image took about 20 seconds on both. GPT Image 2.6 took about 27 seconds at 1K high, 22 to 58 seconds for a 1K low edit with one reference, and 13 to 25 seconds at 2K and 4K, including high quality. Nano Banana 2.1 timings at 2K and 4K are not documented.

Delivery and limits — both run as asynchronous tasks that return a taskId to poll, and Nano Banana 2.1 also accepts a callback_url. Results are kept for 30 days on both.

Per-token billing against one price per tier

With Nano Banana 2.1 the bill is known before you send anything: pick the resolution and you have the price. That holds for a prompt of a few words or 20,000 characters, for a square or an 8:1 strip, and for a plain text-to-image call or an edit that combines 10 references.

GPT Image 2.6 meters what you ask for. Our measured requests from 5 November 2026 (16:9, one image) show the range: 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. Treat these as examples, not a price list; a longer prompt, more references or a GIF changes the token count.

Two patterns follow. At low quality, GPT Image 2.6 can cost less per image than Nano Banana 2.1: $0.0077 against $0.022 at 1K, and $0.0283 against $0.036 at 2K. But the default quality on apimodels.app is medium, and the 2K medium request came in $0.0156 above Nano Banana 2.1's 2K price. If your code never sets quality, budget from the medium figure, not the low one.

References shift the balance as well. On GPT Image 2.6 every reference image is billed as input tokens: the 1K low edit with one reference measured about $0.008 more than the 1K low text-to-image. On Nano Banana 2.1 references cost nothing extra. We have not published a measured price for a GPT Image 2.6 edit with several references, so test with your own inputs before committing a multi-reference workflow.

To check the real charge, GPT Image 2.6 returns usage.cost on completed png, jpeg and webp requests; its charged field is what was deducted. On both models a generation can fail through a moderation refusal or an upstream error, and failed requests are not charged.

What GPT Image 2.6 makes that Nano Banana 2.1 does not

These are different jobs rather than a quality contest. Nano Banana 2.1 makes raster images from a prompt and references; GPT Image 2.6 adds four abilities that Nano Banana 2.1 does not document.

1. Animated GIF. Set output_format to gif, with frame_count from 2 to 64 (default 9) and frame_duration_ms from 20 to 2000 (default 200). All frames sit on one sheet whose longest edge is at most 3840 px; with aspect_ratio plus 1K, 2K or 4K, apimodels.app shrinks the frames to fit, and with an exact size, the whole sheet must stay within 3840 px on the longest edge and 8,294,400 pixels in total.
2. SVG. Set output_format to svg for a static vector at a flat $0.035. It is text-to-image only, n must be 1, quality has no effect, and the file has no scripts, external resources or embedded raster, so it suits icons, logos and flat illustrations rather than photographic scenes.
3. Markup-image edits. Draw circles, arrows or notes on a transparent PNG and pass it as the markup image; the model edits the first reference accordingly.
4. Masks. White marks the area to change and black the area to keep. A markup image and a mask cannot be combined in one request.

Masks need a caveat. In our tests of the previous generation, masks behaved as hints rather than hard boundaries, and we have not measured 2.6 yet. For a strict local edit, also name the region in the prompt, for example "change only the label on the bottle, keep everything else unchanged".

GPT Image 2.6 also takes an exact WIDTHxHEIGHT, so a layout slot can be filled at its native size; Nano Banana 2.1 derives the pixel size from the aspect ratio and tier.

Wide banners, tall strips and text in the frame

Nano Banana 2.1's ratio list reaches 4:1, 8:1, 1:4 and 1:8, for banners, skyscraper ads and tall strips. GPT Image 2.6 sizes stop at a ratio of 3:1, so an 8:1 header would have to be cropped from a 3:1 or narrower image. Nano Banana 2.1 also rejects a ratio outside its list with a 400 rather than quietly replacing it, which makes a wrong value visible in testing.

On text inside the image, Google lists text rendering among the 2.1 upgrades, and in our first test on 2026-10-07 the sign "NANO BANANA 2.1" came out spelled correctly. GPT Image 2.6 Flare is described as handling complex layouts, and Sunburst as spending more time on fine type and reflective surfaces. We have not run a side-by-side quality comparison of the two, so judge on your own prompts; on apimodels.app that is a one-string change to model.

What a job costs

The GPT Image 2.6 figures below multiply our measured requests from 5 November 2026 (16:9, one image); they are examples, not fixed prices. Nano Banana 2.1 figures are its per-image prices on apimodels.app.

100 product shots at 1K, text-to-image — Nano Banana 2.1: $2.20. GPT Image 2.6: about $0.77 at low quality or about $1.62 at high quality. We have not measured a 1K medium request.

100 single-reference edits at 1K — Nano Banana 2.1: $2.20, since the reference is free. GPT Image 2.6: about $1.57 at low quality, based on the measured 1K low edit with one reference.

100 images at 2K — Nano Banana 2.1: $3.60. GPT Image 2.6: about $2.83 at low quality or about $5.16 at medium, the apimodels.app default.

100 SVG icons, or 4K hero images — GPT Image 2.6: $3.50 for 100 SVG icons at the flat $0.035, whatever the quality; Nano Banana 2.1 has no documented SVG output. At 4K, Nano Banana 2.1 is $0.06 per image, so 100 come to $6.00; we have not published a measured 4K cost for GPT Image 2.6, so read the usage object from a test request first.

Which one to use

Pick GPT Image 2.6 when the output is not a plain raster image: animated GIFs, SVG icons and logos, edits steered by a markup sketch or a mask, transparent backgrounds, or layout slots that need an exact WIDTHxHEIGHT. It also takes up to 16 references and returns up to 10 images per call. Use Flare for volume and Sunburst when someone is going to zoom in.

Pick Nano Banana 2.1 when you budget per image and want the number fixed before the request: $0.022 at 1K whatever the prompt length, with up to 10 reference images included. It also covers banner and strip ratios beyond 3:1, and 4K at a published $0.06.

Do not switch for nothing: if a pipeline runs on Nano Banana 2.1 at 2K and would move to GPT Image 2.6 at its default medium quality without using GIF, SVG, markup or masks, the measured per-image cost goes from $0.036 to $0.0516 with no new capability in use. Neither model is the right tool for a strict, pixel-exact local edit: masks acted as hints on the previous GPT Image generation and are not yet measured on 2.6, and Nano Banana 2.1 documents no mask input, so name the region in the prompt and review the result, or composite the change yourself.

Calling both with one key

Both models use the same image endpoint and the same Bearer key; only the model value and the format-specific fields change.

# GPT Image 2.6 Flare: animated GIF, 9 frames at 200 ms each
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 illustration",
    "output_format": "gif",
    "frame_count": 9,
    "frame_duration_ms": 200
  }'

# GPT Image 2.6 Sunburst: high quality for close-up product detail
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": "Macro shot of a linen shirt cuff with a stitched label reading ATELIER",
    "quality": "high"
  }'

# Nano Banana 2.1: 2K edit combining two references, flat $0.036
curl https://api.apimodels.app/v1/images/generations \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "nano-banana-2-1",
    "prompt": "Place the bottle from the first image on the marble counter from the second image",
    "image_urls": ["https://example.com/bottle.png", "https://example.com/counter.jpg"],
    "aspect_ratio": "16:9",
    "resolution": "2K"
  }'

# Each call returns a taskId. Poll GET /v1/images/generations?task_id=...
# until state is completed, then read the result URLs.

FAQ