GPT Image in Detail

[ikaijua]

GPT Image in Detail

GPT Image 2.5 is OpenAI's latest image generation model, open to all ChatGPT users. One line: text-to-image, edit on command — only what you specify changes.

What it is. OpenAI's image model line; 2.5 is the current generation with two API variants, Flare and Sunburst.
Signature capability. Precision editing (change only what you point at), multi-round edit consistency, up to ~50% lower latency.
Price. Free tier usable in ChatGPT; API at $8 image input / $5 text input / $30 image output per 1M tokens.
Note. No independently reproduced benchmark yet — speed/quality claims are mostly official (Manus's 2-4x speedup is the partner corroboration).

Floors below follow “Product intro → Eval & FAQ → Resources”.

Table of Contents

Product intro

1. Official headline upgrades
2. API dual models: Flare vs Sunburst

Eval & FAQ

3. Leaderboards & third-party takes
4. FAQ

Resources

5. Prompt & template library
6. Sources

Clicking a TOC item jumps to the floor; if it does not respond in-page, open the link in a new tab.

[ikaijua]

Official headline upgrades

• Image quality: more natural lighting and textures; better preservation of identifiable features (e.g. faces) when generating from a reference image
• Precision editing: only the parts you specify change — the rest stays intact, even with complex subjects and backgrounds
• Multi-round edit consistency: repeated edits in a long conversation don't corrupt each other; quality doesn't degrade with each round
• Speed: up to ~50% lower generation latency vs Images 2.0
• Intelligence & style: stronger complex-visual-instruction understanding; more accurate image content with real-world information; supports complex layouts (including transparent backgrounds)

New in ChatGPT

How to start: just describe the image in the ChatGPT chat box; type @Sketch to start from a hand drawing, or use templates for a one-click start.

• Sketch: type @Sketch in ChatGPT to hand-draw a rough draft — the model turns it into a finished image, and you can layer style descriptions on top
• Templates: one-click starting points for posters, product shots and other popular formats
• Image comments: annotate directly on the image to mark edit locations for pinpoint edits
• Prompt sharing: share an image with the prompt that generated it, so others can reuse it with their own assets

[ikaijua]

API dual models: Flare vs Sunburst

	GPT-Image-2.5 Flare	GPT-Image-2.5 Sunburst
Model ID	gpt-image-2.5-flare	gpt-image-2.5-sunburst
Positioning	Default choice, speed-first	Edit-precision-first
Best for	Everyday generation, batch tasks	Production ad assets, refined product shots
Price	Same as GPT-Image-2: $8 image input, $5 text input, $30 image output (per 1M tokens)	Same rates, but slower and more token-hungry — higher real cost per image

How to choose: use Flare for everyday traffic; hand Sunburst the jobs where "edits must hold up to scrutiny" (refined product shots, ad assets) — see FAQ.

• Speed: ~50% lower latency than GPT-Image-2; ~2-4x faster in internal evals (independently corroborated by early customer Manus)
• New tiers: quality adds xhigh / max; size ceiling raised to 3840x2160 (4K-class, ~8.3MP)
• Transparent background: background=transparent outputs transparent PNG directly
• Compatibility: gpt-image-2 still callable; existing integrations unaffected
• Ecosystem: Adobe integrates both models in Firefly (first OpenAI image models there); Simon Willison upgraded open-source tool openai_image.py with reference-image support
• Billing unit: tokens are the base billing unit — the smallest chunk of processed content; images are billed in image tokens

[ikaijua]

Leaderboards & third-party takes

• LMArena multi-image-edit board (updated 2026-09-07; LMArena is a public model-battle evaluation platform; Elo is battle rating, higher is better):
  • 1: gpt-image-2.5-sunburst (Elo 1535±9)
  • 2: gpt-image-2.5-flare (Elo 1501±8)
  • Passed: gpt-image-2 (medium) (1454), ByteDance Seedream 5.0 Pro (1414), Meta Muse Image (1405), Nano Banana 2 (1368)

Real-user tests (mixed results)

Positive

• @thesoragirls (targeted-edit test): changed only a candle area, fixed a petal — petal fully preserved, the rest changed as asked
• Yana Welinder (fashion founder): 2.5 is a clear step up for fashion design; output more finished, less "flat" than old models

Negative

• @genel_ai (Japanese animator): single edits are stable, but stacking edits can still show jaggies and texture drift

• Mark Kretschmann (AI/software engineer): forest-scene noise/artifact tests were unstable; Images 2.0 often looked more realistic in several cases

• Overall: 2.5 shines at edit control and complex-instruction execution; realism varies by image — no blanket win over 2.0

Practitioner takes

• Axultan Alimkulov (Higgsfield AI product lead): Flare better understands "what should stay the same" — editing a single element preserves original character, composition and visual identity
• @gkxspace (serial founder): solves the old "first image great, second/third unstable" problem; e-commerce outfit changes, brand-asset extensions and sequential illustrations finally have real use

Honest caveats & safety

• Caveat (OrcaRouter analysis): as of 2026-09-09 there is no independently reproduced benchmark for 2.5 — speed/quality claims are mostly official, with only Manus's 2-4x speedup as partner corroboration. Suggested: route everyday traffic to Flare, hand Sunburst the "edits must hold up" jobs
• Safety: official internal unsafe-generation rates Flare 1.41% / Sunburst 1.09% (below GPT-Image-2's 1.64% baseline); all outputs carry C2PA metadata (digital provenance records) and invisible watermarks

[ikaijua]

FAQ

Can free users use 2.5?
The official announcement says it's open to all ChatGPT users, including the free tier (desktop / mobile / web, per multiple media reports); exact quotas per account aren't disclosed — follow what the client shows.

Flare or Sunburst?
Start with Flare for everyday traffic (default, fast, cheap); hand Sunburst only the production-grade jobs where "edits must hold up to scrutiny" (refined product shots, ad assets). This is the third-party (OrcaRouter) recommendation.

Main improvements over GPT Image 2?
More precise editing (edit exactly what you point at), multi-round edits that don't drift, faster generation (up to ~50% lower latency), better reference-image feature preservation.

Do generated images carry provenance?
All outputs carry C2PA metadata and invisible watermarks; official internal unsafe-generation rates are Flare 1.41% / Sunburst 1.09%, below GPT-Image-2's 1.64% baseline.

Can I output transparent-background images?
Yes. API: background=transparent outputs transparent PNG directly; ChatGPT supports complex layouts including transparent backgrounds.

How is the API billed?
Same as GPT-Image-2: $8 image input, $5 text input, $30 image output (per 1M tokens); the new xhigh/max tiers consume more tokens, so per-image cost is higher (reference: high quality ≈ $0.21 / 1024x1024 image).

[ikaijua]

Prompt & template library

• freestylefly/awesome-gpt-image-2
• wuyoscar/GPT-Image2-Skill

[ikaijua]

Sources

• OpenAI official: Introducing ChatGPT Images 2.5
• Third-party analysis:
  • OrcaRouter: Flare vs Sunburst breakdown
  • Unite.AI launch coverage
  • AlphaSignal dispatch
• Leaderboard: LMArena Image Edit board
• Chinese version of this post: GPT Image in Detail (Chinese)

[tk1475]

Nice writeup. A suggestion for the Prompt & template library floor: https://github.com/tk1475/style-contract, an open source agent skill for keeping a series of GPT Image 2 renders in one consistent style (exact hex palette, finish, composition rules, and a pre-flight check), with 5 presets and their full prompts. Disclosure: I made it.