OpenAI gpt-image-2

Looking for the September 2026 release? See GPT Image 2.5 Flare and Sunburst. This page covers the earlier GPT Image 2 model.

OpenAI's state-of-the-art image generation model — gpt-image-2-2026-04-21 snapshot, integrated into Go Bananas! as a tenant-selectable provider.

Choose your provider — every Go Bananas! tenant can mix Gemini Flash, Gemini Pro, and OpenAI gpt-image-2 per request. (Generated with gpt-image-2 itself.)

openai-gpt-image-2 runs alongside the Gemini family in the model registry. Tenants enable it by adding one to five OpenAI API keys in workspace settings and including it in their allowedModels list, then opt in per request via model_id.

Quick decision tree — gpt-image-2 wins for text-heavy work and quality control; Pro wins for grounding and 4K; Flash for everything else.

When to Pick gpt-image-2

Strength	Why
Text rendering	OpenAI calls this its "killer feature" — small text, multilingual characters, dense compositions, infographics
Prompt adherence	Stronger instruction-following than older OpenAI models; reliable for structured layouts
Quality control	First-class low / medium / high / auto quality levels with explicit pricing tiers
Output formats	Native PNG, JPEG, and WebP with optional output_compression
Aspect-ratio range	Broader supported set including 21:9 cinematic, 4:5 / 5:4 social formats
Edits with refs	Up to 16 reference images via /v1/images/edits for consistent multi-image edits

Constraints to Know

Constraint	Value
Maximum edge	≤3,840px
Edge multiple	Both edges must be multiples of 16px
Aspect ratio cap	Long-to-short ≤3:1
Pixel range	655,360–8,294,400 total pixels
Max reference images	16
Default request timeout	240s (set in src/models/registry.ts)
Transparent background	Not supported — background: "transparent" rejected
Google Search grounding	Not supported (Gemini Pro only)
Native "thinking" mode	Not exposed by Go Bananas! at this layer

Per-Request Parameters

gpt-image-2 exposes a richer parameter surface than the Gemini providers. Pass these alongside model_id: "openai-gpt-image-2":

Parameter	Type	Description
model_id	string	"openai-gpt-image-2" — required to select this provider
prompt	string	Text prompt (max 32,000 chars)
negative_prompt	string	Appended as Avoid: … to the OpenAI prompt
system_instruction	string	Prepended to the prompt
aspect_ratio	enum	One of: square, portrait, landscape, 1:1, 2:3, 3:2, 3:4, 4:3, 4:5, 5:4, 9:16, 16:9, 21:9
size	string	"auto" or explicit WIDTHxHEIGHT (validated against the constraints above)
n	int	1–4 images through Go Bananas (mapped to OpenAI n)
output_format	enum	png (default), jpeg, webp
output_compression	int	0–100 — JPEG/WebP only; rejected for PNG
quality	enum	low, medium, high, auto (default)
background	enum	auto (default) or opaque — transparent rejected
moderation	enum	auto (default) or low for less-restrictive filtering
reference_images	array	Up to 16 input images (sent to /v1/images/edits)
mask_image_id	int	Optional stored mask image for edit_image / edit_uploaded_image with OpenAI

Format selection

PNG is best for line art, screenshots, and anything you'll edit downstream. JPEG with output_compression: 80 cuts storage cost ~70% for photographic output. WebP is the best compromise for web delivery.

Recommended Resolutions

Use case	Size	Notes
Square hero / avatar	1024x1024	Fastest, cheapest
Landscape product shot	1536x1024	Standard 3:2
Portrait poster	1024x1536	Standard 2:3
2K square print	2048x2048	Higher detail
4K landscape wallpaper	3840x2160	Maximum edge
4K portrait poster	2160x3840	Maximum edge
Cinematic banner	3840x1648	21:9

The provider rounds and validates against maxEdge, edgeMultiple, maxAspectRatio, minPixels, and maxPixels from the registry. Out-of-range sizes are rejected with a clear error rather than silently truncated.

Editing with Reference Images

gpt-image-2 routes edits through OpenAI's /v1/images/edits endpoint. When you pass reference_images, Go Bananas! automatically:

1. Bundles each reference as a multipart image[] field with the correct extension
2. Uploads in canonical order (text first, references after — matching Gemini's pattern for cross-provider consistency)
3. Preserves input fidelity (high-fidelity processing is automatic for gpt-image-2; not user-configurable)

{
  "prompt": "Place this product on a marble bathroom counter, soft morning light",
  "model_id": "openai-gpt-image-2",
  "quality": "high",
  "output_format": "webp",
  "output_compression": 85,
  "reference_images": [42, 43]
}

Mask-Guided Edits

For localized OpenAI edits, pass mask_image_id to edit_image or edit_uploaded_image. The mask must already exist in the Go Bananas image library; upload or generate it first, then pass its image ID with model_id: "openai-gpt-image-2".

{
  "image_id": 6367,
  "mask_image_id": 6401,
  "prompt": "Replace only the masked sign with a brushed steel logo plaque",
  "model_id": "openai-gpt-image-2",
  "quality": "high"
}

Masking is OpenAI-only in Go Bananas. Gemini edit tools will reject mask_image_id instead of silently ignoring it. OpenAI also requires the mask to match the source image size and format and include an alpha channel; Go Bananas validates the size/format before sending the request.

Quality vs. Cost

OpenAI publishes token pricing for gpt-image-2; Go Bananas uses practical 1024×1024 heuristics for pre-flight estimates:

Quality	Approx. cost per image	Use when
low	~$0.006	Drafts, thumbnails, A/B mock-ups
medium	~$0.053	Most production work
high	~$0.211	Hero assets, print, final marketing
auto	Variable	Let the model pick — best for mixed workloads

Reference-based edits cost more because the input image counts against your input-token budget (full-fidelity processing). Cached identical inputs are billed at a lower rate.

Output cost dominates

Unlike Gemini's flat per-call pricing, gpt-image-2 charges by output tokens which scale with quality and resolution. Budgeting high quality at 4K can be ~30× the cost of low quality at 1K. Use check_quota and get_account_summary to monitor.

Reliability

OpenAIImageProvider (src/services/image-provider/openai-provider.ts) wraps every call with the same production patterns as Gemini:

• Key pool: Each tenant can store up to five OpenAI keys. Consecutive requests rotate their starting key to spread load across the pool.
• Automatic failover: Authentication, authorization, rate-limit, insufficient-quota, and billing-hard-limit errors move immediately to the next configured key instead of retrying the rejected key.
• Per-key circuit breaker: A rejected key is quarantined for 60 seconds, while healthy keys continue serving traffic. Invalid image requests do not poison key health.
• Safe retry policy: Read-only connection checks may retry transient provider/network failures. Image generation and edit POSTs are never replayed after an ambiguous timeout or 5xx, avoiding duplicate billed generations. Key-specific rate and billing responses fail over immediately.
• Timeout budget: 240s per provider request (set in registry, vs. 60s Flash and 120s Pro) because gpt-image-2 takes longer at high quality and large sizes
• Aspect-ratio enforcement: Output is post-cropped if it deviates >5% from the requested ratio, using the same WASM Photon cropper as Gemini

Pooling reduces avoidable failures but cannot guarantee unlimited throughput. Keys under the same OpenAI project or account may share organization-level rate or billing limits.

When NOT to Use gpt-image-2

Need	Use instead
Google Search grounding for live data	gemini-pro-image
Native thinking mode for complex compositions	gemini-pro-image
Gemini-specific high-fidelity object workflows	gemini-pro-image
Resolution tiers (1K/2K/4K presets)	gemini-pro-image
Lowest latency at large batch sizes	gemini-flash-lite-image (Nano Banana 2 Lite)
Transparent backgrounds	Crop out post-generation; gpt-image-2 doesn't expose this

Setting Up gpt-image-2 for a Tenant

1. Add the keys: Open the tenant admin view → Manage OpenAI Key Pool → enter one to five keys. The whole pool is stored as a versioned AES-GCM-encrypted credential in tenant_provider_credentials; existing single-key rows remain compatible.
2. Add to allowlist: Update the tenant's allowedModels array to include "openai-gpt-image-2".
3. Test: Call generate_image with model_id: "openai-gpt-image-2". The frontend ModelSelector will start surfacing it once the allowlist permits.
4. Monitor: check_quota validates Go Bananas storage and request-rate limits, but it cannot preflight OpenAI account billing or project-level rate limits. Provider-key failover reacts when OpenAI rejects an actual request.

Related Reading

• Pro Model Best Practices — comparable deep-dive for Gemini Pro
• Image Generation — full parameter reference for the unified API
• Multi-tenancy — how the model registry and per-tenant credentials fit together

Sources

• OpenAI gpt-image-2 model reference
• OpenAI image generation guide
• Introducing ChatGPT Images 2.0 (announcement)