Skip to main content

​
Overview

gpt-image-2 is OpenAI’s flagship image model (GA April 22, 2026). It’s the strongest text-in-image renderer in production today — accurate typography across Latin, Japanese, Korean, Hindi, Bengali, and 25+ other scripts — with reasoning-augmented composition and photoreal output. Sync API, blob-hosted output, runtime quality dropdown. One SKU exposes all three of OpenAI’s quality tiers (low / medium / high) — the picker default is medium and you opt into higher fidelity per request. No -pro or -mini derivatives; the model ID matches OpenAI’s exact naming.

​
Specs

FieldValue
Model IDgpt-image-2
CreatorOpenAI
Best forText-in-image, photoreal, multilingual typography, logos
Sizes1024x1024, 1024x1536, 1536x1024, 2048x2048
Quality tierslow, medium (default), high
Pricing modePer image, per quality tier
Default latency~30s medium 1024² (low ~19s, high ~60s, n=3 high ~3min)
OutputBlob-hosted URL (Vercel CDN, no expiring OpenAI URL)

​
Pricing

Per image at 1024². List = OpenAI provider cost × 1.35 markup, rounded up to a sensible cent boundary so margin floors at ~35% across every tier.
QualityProvider costKyma listMargin
low$0.010$0.01440%
medium (default)$0.060$0.08135%
high$0.220$0.29735%
Multi-image requests (n: 3) scale linearly: high × 3 = 0.891.Holdsbooktheexacttieramountbeforethecallsoa‘quality=high‘requestreserves0.891. Holds book the exact tier amount before the call so a `quality=high` request reserves 0.297 up front, no refund-and-rebill drift.

​
Compared to other image models on Kyma

Strengthgpt-image-2flux-2-proideogram-v3recraft-v4-pro
Text in image (English)★★★★★★★★★★★★★★★
Multilingual text★★★★★★★★★★★★
Photoreal humans★★★★★★★★★★★★★★★★★
Composition reasoning★★★★★★★★★★★★★★★★
Print quality (4MP)★★★★★★★★★★★★★★★
Multi-reference blend—10 sources——
Native SVG output———recraft-v4-vector
Stars are positioning, not benchmark percentiles — pick by the row that matches your strongest constraint.

​
Use this when

  • You need text inside the image to be legible and accurate (logos, posters, packaging, UI mockups, screenshots, ads).
  • The text is non-English (Japanese, Korean, Chinese, Hindi, Bengali, Arabic, etc.).
  • The composition needs reasoning (“a chart showing X”, “a diagram of Y”, “a recipe card with Z ingredients”).
  • You’d otherwise pay for a designer to set type properly.

​
Pick something else when

  • You need photoreal hero shots with multi-reference blending → flux-2-pro takes up to 10 source images.
  • You need editable vector files (SVG with paths and layers) → recraft-v4-vector.
  • Volume matters more than fidelity — sub-cent budget tier → minimax-image-01 at $0.005/image.
  • You need print-ready 4MP without paying gpt-image-2 high tier prices → recraft-v4-pro at $0.338/image.

​
Example

curl -X POST https://kymaapi.com/v1/images/generations \
  -H "Authorization: Bearer $KYMA_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-image-2",
    "prompt": "A vintage tea brand poster reading 'KYMA TEA — since 2026' in elegant serif typography, soft cream background, watercolor botanical illustration",
    "size": "1024x1024",
    "quality": "high",
    "n": 1
  }'
The endpoint is async — POST returns 202 with a job_id; poll GET /v1/jobs/{id} until status is succeeded. The completed job’s output.url is a Vercel blob URL hosted on Kyma’s CDN, not an expiring OpenAI URL.

​
Quality tier rule of thumb

# low — drafts, fast iteration, sub-cent draft tier alternative
'"quality": "low"'   # $0.014, ~19s, OK for layout sketches and ideation

# medium — production default
'"quality": "medium"' # $0.081, ~30s, the picker default

# high — hero shots, marketing, anything user-facing
'"quality": "high"'  # $0.297, ~60s, near-perfect text-in-image

​
Tier classification

tier: "quality" (top tier in the unified picker taxonomy alongside flux-2-pro, recraft-v4-pro, ideogram-v3). The default medium quality lands at $0.081 — comparable to flux-kontext-pro / recraft-v4 in the Fast tier — but the SKU’s positioning is the high-end option, with quality: "high" available per request when needed.

​
See also