Skip to main content

​
TL;DR

If you needPick
Precise size / quality control (incl. 4K), must match OpenAI official exactlygpt-image-2 (Official)
Predictable flat pricing ($0.03/image), faster generation, minimal paramsgpt-image-2-all (Reverse)
Both models are built on OpenAI’s gpt-image-2 underneath. The differences are in channel nature (official direct vs reverse-engineered), pricing model, endpoint surface, and parameter granularity.
What -all suffix means: It distinguishes the reverse-engineered model from the official one. On APIYI, -all suffix = reverse-engineered model; without suffix = official direct connection.

​
Full Comparison Table

Dimensiongpt-image-2-all (Reverse, cost-effective)gpt-image-2 (Official)
Model namegpt-image-2-allgpt-image-2
Channel natureReverse-engineered (parity with ChatGPT web image gen)Official direct (OpenAI Images API)
PricingPer-call: flat $0.03/callToken-metered: matches official; ~85% of list price after APIYI deposit bonuses
Typical cost/image$0.03 (regardless of size / quality)Measured $0.03 – $0.2 (correlates with prompt length, size, quality)
Token groupDefaultDefault
Token typePer-call or Token-priority both workToken-priority only (this model is token-billed; per-call tokens will be rejected)
Recommended endpoint⭐ /v1/chat/completions (chat-style, primary)/v1/images/generations + /v1/images/edits
Alt endpoints/v1/images/generations, /v1/images/edits(only the two official ones)
Upload formatbase64 or https URL (chat endpoint)multipart file (edit endpoint)
Output formatb64_json (includes prefix) or url (R2 CDN)b64_json (raw base64, no prefix)
Reference image countMultiple (chat-mode upper bound is high)Max 5 (image[])
Mask inpainting❌ Not supported✅ Supported (alpha channel required)
Prompt adherenceGoodExcellent
Generation speed~60 seconds~100-120 seconds, complex + 4K can reach 3-5 minutes
Resolution controlPrompt-only, output between 1K-2Ksize parameter, 1K / 2K / 3840×2160 4K
Common output sizes16:9 → 1672×941, 9:16 → 941×1672, 1:1 → 1254×12548 presets + any valid custom size
Quality parameter❌ No quality✅ low / medium / high / auto
Transparent background—❌ Not supported (background: transparent errors)
Chinese prompts✅ Native✅ Native
Text renderingHigh fidelityHigh fidelity (strongest at high tier)
Content restrictionsLooserStricter (OpenAI official policy)
API docsGPT-Image-2-All OverviewGPT-Image-2 Overview
🔑 Create or manage API tokens: https://api.apiyi.com/token
When creating a token in the console, choose a group (Default is fine) and a token type (Per-call / Token-priority). Calling gpt-image-2 (official) requires a “Token-priority” token — per-call tokens will be rejected due to billing-mode mismatch.

​
When to Pick Each

​
Pick gpt-image-2-all (Reverse) when

💰 Predictable cost

Stable $0.03/image with no size/quality tier. Ideal for batch production with hard cost ceilings (infographics, marketing assets, e-commerce thumbnails).

⚡ Speed-first

~60s generation, almost 2× faster than the official version — better for real-time UX.

🗨️ Chat-style workflows

Primary endpoint is /v1/chat/completions — multi-turn iterative editing, text-to-image, and reference editing all share one endpoint. Simplest integration.

🌏 Chinese + marketing text

Native Chinese prompt support, excellent text rendering for signage / posters / infographics — great for Chinese-audience content production.

​
Pick gpt-image-2 (Official) when

🖼️ Precise size control

size accepts any valid resolution (incl. 3840×2160 4K). Required for movie posters, desktop wallpapers, video covers where exact aspect ratio / resolution matters.

🎚️ Quality tiers

quality supports low/medium/high/auto. Use low for drafts to save cost; high for print-grade finals.

🎯 Mask inpainting

Alpha-channel mask supported — precisely modify a region while preserving the rest.

🔌 Same as OpenAI Official

Goes through the official Images API — fields and behavior identical to OpenAI official. Existing OpenAI-SDK-based code / systems migrate with zero changes and stay stable long-term.

​
Key Differences in Detail

​
1. b64_json format gotcha (migration trap!)

# gpt-image-2-all: b64_json already has prefix — drop into <img src>
all_b64 = resp["data"][0]["b64_json"]
# "data:image/png;base64,iVBORw0KGgo..."
img_tag = f'<img src="{all_b64}">'  # ✅ direct use

# gpt-image-2: b64_json is raw base64, no prefix — decode or prepend manually
official_b64 = resp.data[0].b64_json
# "iVBORw0KGgo..."
with open("out.png", "wb") as f:
    f.write(base64.b64decode(official_b64))  # ✅ write file
img_tag = f'<img src="data:image/png;base64,{official_b64}">'  # ✅ browser render
When switching between the two, the b64_json handling code must change, or you’ll get a corrupted data URL or a decode failure.

​
2. Resolution control

gpt-image-2-all (in the prompt): 
"Landscape 16:9 cinematic, old lighthouse at sunset"   → ~1672×941
"Portrait 9:16 phone wallpaper, cyberpunk city"        → ~941×1672
"1024×1024 square logo, minimalist cat line art"        → ~1254×1254
gpt-image-2 (size parameter, strict): 
client.images.generate(
    model="gpt-image-2",
    prompt="...",
    size="2048x1152",   # ✅ output exactly this
    quality="high"
)

​
3. Upload / output format differences

Operationgpt-image-2-allgpt-image-2
Upload referencebase64 data URL or https URL (in chat messages’ image_url)multipart image[] file field
Download outputDefault url (R2 CDN, 24h validity), can switch to b64_json (with prefix)b64_json (raw base64, requires decode)
Multi-image fusionMultiple image_url blocks in chatimage[] array, max 5

​
4. Cost ballpark

Scenariogpt-image-2-allgpt-image-2
1024×1024 draft$0.03~$0.006 (low)
1024×1024 medium quality$0.03~$0.053 (medium)
1024×1024 high quality$0.03~$0.211 (high)
2048×1152 high quality$0.03~$0.20+ (token-metered)
3840×2160 4K high quality❌ Not supportedToken-metered, significantly higher than 1K
Edit / multi-image fusion$0.03Input tokens rise sharply, single call can hit $0.1+
Bottom line: For batch / low-quality workloads, gpt-image-2-all isn’t always cheaper (1K low is actually less expensive on the official tier). For the mid-to-high quality range without 4K, gpt-image-2-all’s flat $0.03 is more stable and budgetable. For 4K or precise parameter control, gpt-image-2 is the only choice.

​
Client Settings

Settinggpt-image-2-allgpt-image-2
Timeout (conservative)300 seconds360 seconds (4K high quality realistically reaches 3-5 minutes)
Retry strategyExponential backoff on 5xx / timeout, max 2 retriesSame
Concurrencychat endpoint is naturally concurrency-friendly1 image per call — issue parallel requests for multiple
Request IDrequest-id response headerx-request-id response header

​
FAQ

Yes. Both run on the Default channel — the same API Key calls both with no extra config.
It can. When the image-generation intent isn’t strong enough, gpt-image-2-all’s chat endpoint may return plain text. Workaround: prepend a fixed prefix like “Generate image:” to the user prompt, or constrain output via a system message.
  • Stick with the OpenAI SDK / must match OpenAI official: pick gpt-image-2 (official). Drop input_fidelity, avoid background: transparent, leave the rest unchanged.
  • Want to cut cost too: pick gpt-image-2-all (reverse). Flat $0.03/image, simplest migration via the chat endpoint.
Yes. A common pattern: primary gpt-image-2-all (predictable cost, faster), fallback gpt-image-2 (switch to it when 4K or precise control is needed). The two models’ response shapes differ — normalize at the business layer.