GPT-image-2.5 / 2 Official vs Reverse

Compare the official gpt-image-2.5-flare / gpt-image-2.5-sunburst / gpt-image-2 vs the reverse-engineered sister models gpt-image-2.5-all, gpt-image-2-all, gpt-image-2.5-vip and gpt-image-2-vip: channel nature, pricing, endpoints, upload/output formats, speed vs quality positioning, prompt adherence — pick the right one.

TL;DR

If you need	Pick
quality knob / mask inpainting / arbitrary custom sizes (beyond the 30 presets) / strict OpenAI-API field parity	Official, token-metered: gpt-image-2.5-flare (speed-first) / gpt-image-2.5-sunburst (editing-precision-first) / gpt-image-2 (previous generation)
Predictable flat $0.03/image + fast output (speed is the advantage)	gpt-image-2-all / gpt-image-2.5-all (Reverse, ChatGPT-web line, ~90s; 2.5-all is backed by Images 2.5)
Predictable flat $0.03/image + locked sizes (30 presets incl. 4K)	gpt-image-2-vip / gpt-image-2.5-flare-vip / gpt-image-2.5-sunburst-vip (Reverse, Adobe line; the two 2.5 models take all six quality tiers, gpt-image-2-vip up to high)

All eight models are built on OpenAI’s GPT-Image 2.5 / 2 series underneath. The differences are in channel nature (official direct vs reverse-engineered), pricing model, and parameter granularity. The three official models share price and parameters; the five reverse models share price and call format.

The three reverse lines (-all / 2.5-all / -vip trio): This page’s “Reverse” column covers gpt-image-2-all, gpt-image-2.5-all and gpt-image-2-vip / gpt-image-2.5-flare-vip / gpt-image-2.5-sunburst-vip (alias gpt-image-2.5-vip = sunburst-vip). All share the $0.03/image flat price and the same call format (the three -vip models additionally support size):

• gpt-image-2-all / gpt-image-2.5-all: ChatGPT web line, ~90s generation — speed is the advantage; 2.5-all is backed by Images 2.5
• gpt-image-2-vip and the two 2.5 -vip models: Adobe line (Firefly), size locking (30 presets incl. 4K); all three accept quality (channel behavior, not a commitment): the two 2.5 models opened xhigh / max in 2026-09-10 retesting and now take all six tiers, gpt-image-2-vip goes up to high; all three return transparent backgrounds; flare-vip is the fastest with a softer look, sunburst-vip is visually close to gpt-image-2-vip
• In common: none support n; mask is whole-image regeneration on all of them, with no guarantee of touching only the masked region

For precise mask inpainting, arbitrary custom sizes beyond the 30 presets, or n > 1, use the official models (gpt-image-2.5-flare / gpt-image-2.5-sunburst / gpt-image-2).

About current speeds: -all / -vip generation is slower than at launch due to OpenAI upstream compute fluctuations — this affects all reverse-channel users, not just APIYI; our account pool and ops are healthy. Set client timeouts to 300s+ and leave more headroom for complex prompts.

Full Comparison Table

Dimension	gpt-image-2-all / 2.5-all / -vip (Reverse, cost-effective)	gpt-image-2.5-flare / sunburst / gpt-image-2 (Official)
Model name	gpt-image-2-all (fastest) / gpt-image-2.5-all (2.5 version of the same line) / gpt-image-2-vip (quality-first, size-locking) / gpt-image-2.5-flare-vip · gpt-image-2.5-sunburst-vip (2.5 versions of the same line, alias gpt-image-2.5-vip)	gpt-image-2.5-flare (speed-first) / gpt-image-2.5-sunburst (editing-precision-first) / gpt-image-2 (previous generation)
Channel nature	-all / 2.5-all: reverse-engineered ChatGPT web line the three -vip models: reverse-engineered Adobe line (Firefly; a high-quality GPT-Image 2.5 reverse line, not upscaling)	Official direct (OpenAI Images API); same price and parameters across all three
Pricing	Per-call: flat $0.03/call (all five reverse models, same price)	Token-metered: matches official; ~85% of list price after APIYI deposit bonuses
Typical cost/image	$0.03 (regardless of size / quality / model)	Measured $0.03 – $0.2 (correlates with prompt length, size, quality)
Token group	Default	Default
Token type	Per-call or Token-priority both work	Token-priority only (this model is token-billed; per-call tokens will be rejected)
Recommended endpoint	/v1/images/generations + /v1/images/edits (more stable, more upstream supply, and same code as official — just swap the model name to switch during risk-control turbulence)	/v1/images/generations + /v1/images/edits
Upload format	multipart file (edits endpoint)	multipart file (edit endpoint)
Output format	b64_json (default, raw base64 with no prefix, verified 2026-07; earlier versions included the prefix) or url (R2 CDN)	b64_json (raw base64, no prefix)
Reference image count	Multiple	Max 16 (image[])
Mask inpainting	-all / 2.5-all: ❌ not supported the three -vip models: ⚠️ accepted but whole-image regeneration — three runs on real photos measured an inside/outside change ratio of about 1 (2026-09-09); no guarantee of touching only the masked region	✅ Supported (alpha channel required)
Prompt adherence	Good	Excellent
Generation speed	-all: ~90 seconds (speed is the advantage) gpt-image-2-vip: ~90–150 seconds the two 2.5 -vip models: serial 1024² runs measured flare 22–138 s and sunburst 37–120 s, wide variance; no concurrency planning needed within 500 RPM, retry an occasional 429 with backoff or check Live Updates 📌 Currently slower than at launch — OpenAI upstream compute, not an APIYI-side issue	gpt-image-2.5-flare: fastest official, 1K low measured ~10 s (2026-09-09) gpt-image-2.5-sunburst: 1K low measured ~14 s, slower at high quality gpt-image-2: ~100-120 seconds, complex + 4K can reach 3-5 minutes
Quality tendency	-all / 2.5-all: good (same line; the two names produce the same images) -vip: on six same-size prompts judged by eye, sunburst-vip is close to gpt-image-2-vip, flare-vip is softer with fewer decorative details; Chinese headline strokes are correct on all three	Stable; both 2.5 models beat gpt-image-2, sunburst highest, and quality=xhigh / max max it out
size parameter	-all / 2.5-all: ❌ Not accepted (describe in prompt) the three -vip models: ✅ Restored (since 2026-07-22), 30 preset sizes (incl. 4K); images endpoints only, not supported on the chat endpoint. With size omitted, gpt-image-2-vip / sunburst-vip return 2048×2048 and flare-vip a fixed 1024×1536 (the default has changed upstream before) — pass a preset to lock it	✅ Any valid custom size
4K support	-all / 2.5-all: ❌ the three -vip models: ✅ 4K Detail tier (e.g. 3840x2160 / 2880x2880), no surcharge	✅ Including 3840×2160
Common output sizes	-all: 16:9 → 1672×941, 9:16 → 941×1672, 1:1 → 1254×1254 (adaptive) the three -vip models: 30 presets (10 ratios × 1K/2K/4K), see the full 30-size table; sizes outside the presets do not error but get aligned to multiples of 16 or raised to the minimum edge	8 presets + any valid custom size
quality parameter	-all / 2.5-all: ❌ rejected (do not pass) the three -vip models: ✅ measured, channel behavior, not a commitment — the two 2.5 models take all six tiers auto / low / medium / high / xhigh / max (xhigh / max opened 2026-09-10), gpt-image-2-vip up to high	✅ low / medium / high / xhigh / max / auto (xhigh / max on the two 2.5 models only)
quality tier alignment	Output tokens at 2048×1152: gpt-image-2-vip low 157 / medium 1,413 / high 5,650; the two 2.5 -vip models low 157 / medium 367 / high 1,413 / xhigh 2,511 / max 5,650 — 2.5 high = gpt-image-2-vip medium, 2.5 max = gpt-image-2-vip high, the same relationship as official 2.5 vs gpt-image-2; per-call billing, so the tier does not change the price	Output tokens at 1024²: official 2.5 low 196 / medium 439 / high 1,756 / xhigh 3,122 / max 7,024; gpt-image-2 low 196 / medium 1,756 / high 7,024 — official 2.5 high = gpt-image-2 medium; token-metered
n parameter	❌ None of the five reverse models support it (1 image per call)	✅ Supported
Transparent background	-all / 2.5-all: ⚠️ no background parameter, prompt-only, occasionally unreliable the three -vip models: ✅ background: "transparent" returns an alpha PNG in testing (not a commitment)	✅ Parameter-controlled and reliable — background: "transparent" with png / webp
Chinese prompts	✅ Native	✅ Native
Text rendering	High fidelity	High fidelity (strongest at high tier)
API docs	GPT-Image-2.5-All Overview / GPT-Image-2.5-VIP Overview	GPT-Image-2.5 / 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 the three official models (gpt-image-2.5-flare / sunburst / gpt-image-2) 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 / gpt-image-2.5-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).

⚡ Faster output

~90s generation — slightly faster than both -vip and the official version. Better real-time UX.

🔁 One codebase, swap anytime

Standard Images API format — same code as the three -vip models and the three official models; switch or fall back by changing the model name. gpt-image-2-all and gpt-image-2.5-all share price and behavior; the new name simply reflects that ChatGPT web has moved to Images 2.5.

🌏 Chinese + marketing text

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

Pick the three -vip models (Reverse, size-locking) when

🎚️ quality works (all six tiers on 2.5)

gpt-image-2.5-flare-vip / gpt-image-2.5-sunburst-vip accept all six tiers from auto to max in testing (xhigh / max opened 2026-09-10); gpt-image-2-vip goes up to high. All of it is channel behavior, not a commitment. Tier alignment: 2.5 high only equals gpt-image-2-vip medium, and 2.5 max equals gpt-image-2-vip high; the flat $0.03 does not change with the tier.

⏱️ Choosing among the three

flare-vip is the fastest with a softer look; sunburst-vip has higher quality and editing precision and looks close to gpt-image-2-vip; all three reach the same top token tier (max on 2.5, high on gpt-image-2-vip). All are slower than -all — max at 1024² measured 80–160 s — so pick them when a longer wait is acceptable.

🖼️ Locked sizes / 4K

The size parameter is restored (since 2026-07-22): 30 preset sizes (10 ratios × 1K/2K/4K). E-commerce hero shots, poster templates and 4K wallpapers come out at exact dimensions — flat $0.03/image, no 4K surcharge.

🔁 Code shared with -all

Same request structure as -all (just one extra size field) — one codebase switches across all five reverse models by swapping the model name based on your speed / quality preference.

-vip’s size only works on the /v1/images/generations and /v1/images/edits endpoints — the /v1/chat/completions chat endpoint does not support size. For arbitrary custom sizes beyond the 30 presets or precise mask inpainting, use the official models (gpt-image-2.5-flare / sunburst). Availability of this parameter follows upstream changes — see Live Updates for the latest status.

Pick the official models (gpt-image-2.5-flare / sunburst / gpt-image-2) when

🎚️ Quality tiers

All six quality tiers are available and officially committed, with tiers and token counts stable per the official spec. The reverse 2.5 -vip models have also opened all six tiers (measured 2026-09-10), but that is channel behavior with no commitment, and gpt-image-2-vip still stops at high.

🎯 Mask inpainting

Alpha-channel mask supported — precisely modify a region while preserving the rest. The reverse models only regenerate the whole image and do not guarantee touching only the masked region.

🖼️ Arbitrary custom sizes

size accepts any valid resolution (including 4K), not limited to presets. The -vip trio only guarantees the 30 presets and rewrites anything else — strict custom sizes go official.

🔌 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!)

As verified in July 2026, the official and reverse models all return raw base64 (no data: prefix) — but gpt-image-2-all used to include the prefix, so the safest shared code checks for it first:

# Universal pattern: detect the prefix before processing — works for all official and reverse models
b64 = resp["data"][0]["b64_json"]
if b64.startswith("data:"):          # handles historical prefixed responses
    b64 = b64.split(",", 1)[1]
with open("out.png", "wb") as f:
    f.write(base64.b64decode(b64))   # ✅ write file
img_tag = f'<img src="data:image/png;base64,{b64}">'  # ✅ browser render

When switching between official and reverse, 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 / gpt-image-2.5-all (no size field — composition goes in the prompt; both names behave the same):

"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

The three -vip models (gpt-image-2.5-vip / gpt-image-2.5-flare-vip / gpt-image-2-vip; size restored since 2026-07-22):

Accept 30 preset sizes (10 ratios × 1K/2K/4K) — pass size: "WIDTHxHEIGHT" directly (must be one of the 30 presets; full list in the 30-size table). All three also accept quality in testing (channel behavior, not a commitment): the two 2.5 models take all six tiers (xhigh / max opened 2026-09-10), gpt-image-2-vip up to high; the 2.5 models’ high output tokens only equal gpt-image-2-vip medium, and their max equals its high:

client.images.generate(
    model="gpt-image-2.5-vip",   # alias = gpt-image-2.5-sunburst-vip; swap in gpt-image-2.5-flare-vip / gpt-image-2-vip on this line only
    prompt="...",
    size="3840x2160",   # ✅ one of the 30 presets; images endpoints only (not chat)
    quality="max"       # optional; all six tiers on the 2.5 models, gpt-image-2-vip stops at high; flat $0.03 regardless of tier
)

Official models (size strictly honored + quality tiers; the sample uses gpt-image-2.5-flare, swap in sunburst / gpt-image-2 by changing only the model name):

client.images.generate(
    model="gpt-image-2.5-flare",
    prompt="...",
    size="2048x1152",   # ✅ output exactly this
    quality="max"       # all six tiers; xhigh / max on the two 2.5 models only; 2.5 high only equals gpt-image-2 medium
)

3. Upload / output format differences

Operation	Five reverse models (-all / 2.5-all / the -vip trio)	Three official models (2.5-flare / 2.5-sunburst / gpt-image-2)
Upload reference	multipart image file field (edits endpoint)	multipart image[] file field
Download output	Default b64_json (raw base64, verified 2026-07); explicit response_format: "url" returns an R2 CDN link (24h validity)	b64_json (raw base64, requires decode)
Multi-image fusion	Repeat the image field on the edits endpoint	image[] array, max 16

4. Cost ballpark

Scenario	Five reverse models (-all / 2.5-all / the -vip trio)	Official gpt-image-2.5-flare / sunburst	Official gpt-image-2
1024×1024 draft (low)	$0.03	~$0.006 (196 tok)	~$0.006 (196 tok)
1024×1024 medium quality (medium)	$0.03	~$0.013 (439 tok)	~$0.053 (1,756 tok)
1024×1024 high quality (high)	$0.03	~$0.053 (1,756 tok)	~$0.211 (7,024 tok)
1024×1024 xhigh / max	$0.03 (accepted on the 2.5 -vip models; gpt-image-2-vip stops at high; -all rejects quality)	~$0.094 (3,122 tok) / ~$0.211 (7,024 tok)	❌ Not supported
2048×1152 high quality	$0.03	Token-metered; high lands around gpt-image-2 medium	~$0.20+ (token-metered)
3840×2160 4K high quality	$0.03 (-vip 4K Detail tier, no surcharge; -all / 2.5-all have no 4K)	Token-metered, significantly higher than 1K	Token-metered, significantly higher than 1K
Edit / multi-image fusion	$0.03	Input tokens rise sharply, single call can hit $0.1+	Input tokens rise sharply, single call can hit $0.1+

Official figures are rough: output tokens measured on 2026-09-09 × $30 per million, excluding prompt input tokens (usually under $0.001).

Bottom line: For batch / low-quality workloads, the reverse channel isn’t always cheaper (1K low is actually less expensive on the official tier, and the two 2.5 models cost only ~$0.013 / ~$0.053 even at medium / high). The mid-to-high quality range (gpt-image-2 from medium up, the 2.5 models from xhigh up) is where the reverse channel’s $0.03 becomes the sweet spot. Pick official (token-metered) when you need quality tiers / mask inpainting / locked sizes, 4K / strict OpenAI-API field parity.

Client Settings

Setting	Five reverse models (-all / 2.5-all / the -vip trio)	Three official models (2.5-flare / 2.5-sunburst / gpt-image-2)
Timeout (conservative)	-all / 2.5-all: 300s (typical ~90s) the three -vip models: 300s (typical 120–200s)	2.5 models at low / medium: 240s; high / xhigh: 300s; max and gpt-image-2 high: 600s fallback (4K high quality realistically reaches 3-5 minutes) — see the tier table in the official overview
Retry strategy	Exponential backoff on 5xx / timeout, max 2 retries	Same
Concurrency	1 image per call — issue parallel requests for multiple	1 image per call — issue parallel requests for multiple
Request ID	request-id response header	x-request-id response header

Common to all eight models: for image edit / multi-image fusion, compress each input image to under 1.5MB (JPEG quality 80-90 / down-sized resolution). Sporadic shell_api_error / Unknown error responses are most often triggered by oversized inputs — compressing measurably improves success rate and latency. Output resolution is independent of input size — quality is set on the output side (size + quality for official; the size tier plus quality for the -vip trio; prompt phrasing for -all / 2.5-all), not by input file size.

FAQ

Should I compress input images? Does writing 4K / 8K in the prompt help?

Yes, strongly recommended. For all eight models, compress each input image to under 1.5MB (JPEG quality 80-90 / down-sized resolution): sporadic shell_api_error / Unknown error responses are most often triggered by oversized inputs, and compressing measurably improves success rate and latency.

Don’t worry about compression hurting quality — output resolution is independent of input size. The “output-side” controls differ across the three families:

• gpt-image-2-all / gpt-image-2.5-all: controlled by prompt composition phrasing (see the verified phrasing table on the -all overview page) — 4K / 8K in the prompt does not count
• The three -vip models (gpt-image-2.5-vip / gpt-image-2.5-flare-vip / gpt-image-2-vip): controlled by the size field (30 preset sizes incl. 4K), optionally with quality (six tiers on the 2.5 models, up to high on gpt-image-2-vip) — 4K / 8K in the prompt does not count either
• Official models (gpt-image-2.5-flare / sunburst / gpt-image-2): controlled by size + quality (any valid size)

Bottom line: shrinking inputs only speeds things up — quality is set by output-side configuration, not input file size.

Can the same API Key call all eight models?

Yes. All eight run on the Default group — the same API Key calls them with no extra config. Note: calling the official models requires a “Token-priority” token; -all / -vip accept either token type.

Which endpoints should I use on the reverse channel?

Use the OpenAI Images API (/v1/images/generations for text-to-image + /v1/images/edits for editing), for two reasons:

1. More stable: upstream resource supply for the Images API channel is more plentiful, so call success rates are higher
2. Compatible with the official relay for easy switching: the call method and parameter format are fully compatible with the three official models (gpt-image-2.5-flare / sunburst / gpt-image-2) — if the reverse channel hits risk-control turbulence, just swap the model name to switch to the official relay with zero code changes

There is also a chat-based endpoint (/v1/chat/completions, no longer recommended), only useful for multi-turn iterative editing or passing online image URLs directly. Note that when the image intent is ambiguous, it may return plain text instead of an image (prepend a fixed prefix like “Generate an image:” to reinforce it). For full parameters, see the -all chat-based API reference / -vip chat-based API reference.

Within the reverse channel, -all vs -vip — which to pick?

Both lines are reverse-engineered channels at the same flat price ($0.03/image), with the same call format (the -vip trio additionally supports size locking and quality, all six tiers on the 2.5 models). The difference is speed vs quality + size locking:

• Generation time: -all / 2.5-all ~90s — speed is the advantage; the -vip trio ~120–200s. Currently slower than at launch due to OpenAI upstream compute fluctuations
• Quality: -vip (Adobe line) detail rendering is sometimes higher — for showcase images when you’re not in a hurry; within the trio, sunburst-vip looks close to gpt-image-2-vip and flare-vip is softer
• Size locking: the -vip trio supports 30 preset size values (incl. 4K); -all / 2.5-all reject size — composition goes into the prompt

Decision: want fast output → -all / 2.5-all; need locked sizes / 4K → the -vip trio; need custom sizes beyond the 30 presets or precise mask → official. See the GPT-Image-2.5-VIP Overview for details.

Within -vip, the two 2.5 models vs gpt-image-2-vip — which to pick?

Same price, groups and call format (a 253-request three-arm comparison on the same channel and token on 2026-09-09 found the contract identical cell for cell). Only three things differ:

• Tiers: the 2.5 models take all six (xhigh / max opened 2026-09-10), gpt-image-2-vip up to high; same-named tiers are not equal — at 2048×1152, 2.5 high 1,413 = gpt-image-2-vip medium, and 2.5 max 5,650 = gpt-image-2-vip high. All three reach the same top token tier
• Quality and speed: flare-vip is the fastest with a softer look and fewer decorative details; sunburst-vip looks close to gpt-image-2-vip
• Default size: flare-vip is a fixed 1024×1536, the other two 2048×2048; always pass size to lock it

The alias gpt-image-2.5-vip is sunburst-vip. The full row-by-row table is in the GPT-Image-2.5-VIP Overview, section “Three -vip models compared”.

I need locked sizes / 4K — what now?

Start with the -vip trio (gpt-image-2.5-vip by default; pick gpt-image-2-vip with high for the highest token tier): the size parameter was restored on 2026-07-22 and supports 30 preset sizes (10 ratios × 1K/2K/4K) at a flat $0.03/image with no 4K surcharge. Note that size only works on the images endpoints and must be one of the 30 presets.

Go official (gpt-image-2.5-flare / sunburst / gpt-image-2, token-metered) when you need any valid size beyond the 30 presets, officially committed quality tiers (the -vip tiers are channel behavior with no commitment), precise mask inpainting (alpha-channel mask), or strict OpenAI-API field parity (zero-change migration for existing OpenAI-SDK code).

Migrating from 1.5 — which one should I pick?

• Stick with the OpenAI SDK / must match OpenAI official, or need custom sizes beyond the 30 presets: pick the official models (gpt-image-2.5-flare for text-to-image, gpt-image-2.5-sunburst for edits). Drop input_fidelity and leave the rest unchanged (background: transparent keeps working).
• Cut cost, want fast output: pick gpt-image-2.5-all (reverse, ~90s; same price and behavior as gpt-image-2-all).
• Cut cost, quality-first or need locked sizes / 4K: pick gpt-image-2.5-vip (reverse, ~120–200s, 30 preset sizes incl. 4K, all six quality tiers; gpt-image-2-vip stops at high).

Can I deploy multiple models for failover?

Yes. A common pattern: primary 2.5-all or 2.5-vip (predictable cost — pick by speed / quality preference), fallback to the official gpt-image-2.5-flare / sunburst (switch when you need quality tiers, mask, or custom sizes beyond the 30 presets). The reverse and official response shapes differ — normalize at the business layer.

The R2 CDN image link is slow — what can I do?

See Slow CDN downloads — what to do

Related Docs

• GPT-Image-2.5 / 2 Overview - Full integration docs for the three official models
• GPT-Image-2.5-All Overview - Reverse ChatGPT-web line (fastest output; gpt-image-2.5-all / gpt-image-2-all) full integration docs
• GPT-Image-2.5-VIP Overview - Reverse Adobe line (the gpt-image-2.5-vip series plus gpt-image-2-vip; size locking, quality tiers) full integration docs
• Deep dive: GPT-image-2.5 launch - The 2.5 dual-model launch
• Deep dive: gpt-image-2 launch - Official version launch
• Deep dive: gpt-image-2-all launch - Reverse-engineered version launch
• Community: Luck GPT-Image 2 ComfyUI Nodes - Multi-model ComfyUI node pack
• Community: APIYI GPT-Image 2 Skills - Multi-model AI Agent Skill pack
• Deposit promotions - Recharge bonus policy