GPT Image 2.6 on varg

OpenAI released GPT Image 2.6 on 2026-11-04: animated GIF output, two API models (Flare and Sunburst) that "generate images 50% faster than before", image markup, SVG export and a complexity control (@OpenAIDevs, @OpenAI). The token price is the same as 2.5: $30 per million image output tokens, $8 per million image input tokens, $5 per million text input tokens (@dotey). On varg it is gpt_image_2_6, live since 2026-11-04.

On the LMArena boards on launch night, Sunburst (medium) was #2 in text-to-image (1409) and #3 in image edit (1476); Flare (medium) was #4 (1384) and #5 (1462) (@arena).

This skill is the model guide. The other four skills in the set:

1. Flare or Sunburst

Flare is the default. Pin Sunburst only when you will edit the same image again and again. Edits take up to 16 reference images (fal: "A maximum of 16 images are allowed").

2. quality and reasoning_effort are two different dials

• quality (low, medium, high, auto; varg default high) sets render detail and price.
• reasoning_effort (low, medium, high) is how long the model plans before drawing; OpenAI calls it complexity. OpenAI: "Lower complexity can suit focused edits, such as changing a background or a sign. Higher complexity can help with dense text, complex layouts, and relationships between objects" (OpenAI image guide). fal: "Higher settings increase detail, latency, and token usage." xhigh and max are in the schema, but fal rejected both (measured 2026-11-05 on varg). We did not measure whether effort changes the varg price.

Both go in provider_options.fal.

3. Sizes and aspect ratios

Sizes from our 2026-11-05 outputs on varg. 16:9 came back as 1088×608, likely because 1024×576 (589,824 px) is under the 655,360 px floor. For any other size, pass provider_options.fal.image_size: {"width": W, "height": H}: both sides multiples of 16, longest edge ≤ 3840, total 655,360-8,294,400 px (fal schema). Bigger images cost more: fal billed 10.7¢ for one 2048×2048 high image (measured 2026-11-05), so price it with POST /v2/estimate before you go above 1024. In ChatGPT a 16:9 GIF request came back square for @joseamijares; in the API, always set the size yourself.

4. What works and what fal rejects

Every PNG and GIF the model returned in our tests carried a C2PA content-credentials block (17 files, checked 2026-11-05). None of the files made by this set's scripts kept it (checked the same day); keep the original if you need the provenance.

5. Prices (measured 2026-11-05 on varg)

Check any request before you run it: POST https://api.varg.ai/v2/estimate with the same body, or estimate_cost on the MCP. Follow the spend rules in the varg-ai skill (a cap, approved models, estimate before every paid call).

6. The calls

MCP (https://mcp.varg.ai/mcp, OAuth, no key):

call_tool({
  "tool": "image",
  "input": {
    "model": "gpt_image_2_6",
    "prompt": "…",
    "aspect_ratio": "3:4",
    "provider_options": {"fal": {"quality": "high", "reasoning_effort": "medium"}}
  },
  "wait": true
})

It returns url, job_id and cost_credits. An edit adds "files": [{"url": "https://s3.varg.ai/…"}] to input. To use a local file, upload it first: create_upload({"media_type": "image/png", "size": <bytes>}), then curl -X PUT -H "Content-Type: application/octet-stream" --data-binary @photo.png "https://api.varg.ai$PROXY_URL" with the proxy_url it returned; the answer has the file's public_url.

REST (needs $VARG_API_KEY, see the varg-ai skill):

# upload a local image -> {"url": "https://s3.varg.ai/...", ...}
curl -s -X POST https://api.varg.ai/v2/files -H "Authorization: Bearer $VARG_API_KEY" \
  -H "Content-Type: image/png" --data-binary @photo.png

# text to image (Flare)
curl -s -X POST https://api.varg.ai/v2/image -H "Authorization: Bearer $VARG_API_KEY" \
  -H "Content-Type: application/json" -d '{
  "model": "gpt_image_2_6",
  "prompt": "…",
  "aspect_ratio": "3:4",
  "provider_options": {"fal": {"quality": "high", "reasoning_effort": "medium"}}
}'

# edit on Sunburst, custom size
curl -s -X POST https://api.varg.ai/v2/image -H "Authorization: Bearer $VARG_API_KEY" \
  -H "Content-Type: application/json" -d '{
  "model": "fal:openai/gpt-image-2.6/sunburst/edit",
  "prompt": "…",
  "files": [{"url": "https://s3.varg.ai/files/…/photo.png"}],
  "provider_options": {"fal": {"quality": "high", "image_size": {"width": 1536, "height": 864}}}
}'

# each image call answers 202 with {"id": "job_…"}; poll until "completed"; the image is output.outputs[0].url
curl -s https://api.varg.ai/v2/jobs/JOB_ID -H "Authorization: Bearer $VARG_API_KEY"

Model spec with every option: https://varg.ai/models/gpt_image_2_6.md. JSON schema (no key): https://api.varg.ai/v2/tools/image?model=gpt_image_2_6.

7. Prompting text-heavy layouts

What the recipes below have in common:

• Put every word that must appear in quotes, and keep each one short ("Bright citrus", not a sentence).
• Name the layout before the content: "3×2 grid", "top 75% / bottom 25%", "sidebar on the left".
• Name the label words you want ("FRONT, 3/4 VIEW, SIDE VIEW…") and say "minimal clean labels".
• End with what to leave out ("no other text", "no logos", "no photorealistic background").
• For dense pages use reasoning_effort: "high"; for a photo restyle medium is enough.
• Read every word in the output before you use it. Fix one wrong word with a markup edit (gpt-image-markup-edit), not a full regeneration.

8. Recipes

Full prompt templates, settings and notes: references/recipes.md. Most community prompts were written for GPT Image 2.5; our examples show each one on 2.6. Examples were made on varg on 2026-11-04/05 with gpt_image_2_6.

9. Failure → fix

Tested

Measured on varg on 2026-11-05: route times (Flare ~10 s, Sunburst ~49 s), the prices in section 5, the rejections in section 4, 16:9 at 1088×608, and every example linked above (made with this skill's recipes; we did not keep the exact filled-in prompts).

Related skills

gpt-image-gif, gpt-image-sticker-pack, gpt-image-sprite-sheet, gpt-image-markup-edit (this set); varg-ai (setup, keys, every model and price); varg-character-design and character-passport (character sheets for video); kling-4-prompting (stills as key frames for video).