GPT Image 2.5 is OpenAI's image model family released on September 8, 2026 as two API models, gpt-image-2.5-sunburst (editing precision) and gpt-image-2.5-flare (fast everyday generation). MiniMax-H3-Image followed on November 10, 2026. This page shows how they compare in images other people actually posted, then the published specs and prices.
What the side-by-sides show
We did not run these tests. Each one is a post on X, embedded from X and credited to its author. The line under each heading describes what the post shows and what the author said. It is not our verdict. Many posters are MiniMax early-access creators using the #H3Image tag, so read them as showcases, not neutral benchmarks.
Text rendering and layout
@maxescu, H3 Image vs Seedream 5.5 vs Nano Banana 2.1 vs GPT Image 2.6, one packaging-mockup prompt. The author scores each model item by item and names Seedream 5.5 the winner. By that scoring, H3 got the label text and camera angle right but missed the gloves, the goat's pack and the bent horn. GPT Image 2.6 got the text right but drew the wrong character. The prompt is in the post and in our prompt library.
Quick test: MiniMax H3 Image, Seedream 5.5, Nano Banana 2.1 and GPT Image 2.6 on one product-mockup prompt, scored item by item.
— @maxescu View post on X
Photorealism
@maxescu, "H3 Image that looks very real", vs Seedream 5.5, Nano Banana and GPT-Image 2.6. One realism prompt across four models. The author's read: Seedream has the best film look and GPT-Image 2.6 is "very accurate, but it tries way too hard". The image can be opened at 4K on X.
H3 Image that looks very real. Seedream 5.5, Nano Banana and GPT-Image 2.6 compared.
— @maxescu View post on X
@MArch_Emir61, Seedream 5.5 vs GPT Image 2.5 vs Nano Banana Pro vs H3 Image. A four-way grid from an early-access tester. It shows the outputs side by side without a verdict.
Seedream 5.5 - GPT Image 2.5 - Nano Banana PRO - Minimax H3 Image
— @MArch_Emir61 View post on X
@Cg3Jvf (云野), ChatGPT image-2.6 vs Midjourney V8.2 vs H3-Image on the same prompt. Three outputs. The author asks readers whether H3 belongs in the top tier and does not pick a winner.
同样的提示词对比 Chat gpt image-2.6 / Midjourney V8.2 / MiniMax H3-Image
— @Cg3Jvf View post on X
Portraits and fine detail
@MArch_Emir61, a "Detail" test. One close-up portrait prompt built to stress skin texture, deep wrinkles and micro-detail, run on Seedream 5.5, GPT Image 2.5, Nano Banana Pro and H3. The outputs are shown side by side without a ranking.
A Detail test for 4 visual models: Seedream5.5 - GPT Image 2.5 - Nano Banana PRO - Minimax H3
— @MArch_Emir61 View post on X
Editing and multi-reference
@One2xDatalab, three-stage edit (Halloween clown → makeup removed → modern look). The author reports that H3 Image and Seedream followed the makeup-removal step while GPT Image 2.6 kept the stage makeup. All three handled the outfit and scene change.
H3 Image and Seedream followed the makeup edit; GPT Image 2.6 kept the stage makeup.
— @One2xDatalab View post on X
@genel_ai, "10-person year-end party", using ten character references (Japanese). The author judged H3 the most accurate: it kept the high camera angle, fit all ten people in frame and matched faces and outfits, with slightly high saturation. GPT Image 2.6 Sunburst drew actions and photo realism well but tended to crop in, ignoring the angle instruction.
複数人物の指示、最も精確なのは？ ①MiniMax H3 Image ②Seedream5.5 ③GPT Image 2.6
— @genel_ai View post on X
@ktkid_lab, character consistency with a style reference. H3-Image vs "GPT" (version not stated). Image 1 is the character, image 2 the style reference. Style-transfer, three-view and pixel-art results follow in the thread.
我拿 MiniMax H3-Image 和 GPT 做了个对照。图 1 是角色原图，图 2 是风格参考。
— @ktkid_lab View post on X
@ShadeLurk, refining a GPT Image background with H3 Image. This is not a head-to-head. The top image is the GPT Image original; the bottom is H3 Image editing it from one source image plus an instruction. The author says the grain went away and it took about 30 s per image.
上：GPT Image / 下：H3 Image (refined from the GPT Image original)
— @ShadeLurk View post on X
Game scenes and overall feel
@underwoodxie96, the same game-scene prompts. Top row MiniMax-H3-Image, bottom row GPT Image 2.5. The author's verdict: H3 "looks a bit flatter".
Top row: MiniMax-H3-Image. Bottom row: GPT Image 2.5. MiniMax-H3-Image looks a bit flatter to me.
— @underwoodxie96 View post on X
@mirochill, first impressions (French). The author finds H3 excellent at following complex prompts and at using up to 16 references, but well behind with simple prompts. For everyday use they would still pick GPT Image 2.6.
Premiers résultats de MiniMax H3 Image… pour un usage quotidien, je privilégierais GPT Image 2.6.
— @mirochill View post on X
Specs side by side
| MiniMax-H3-Image | GPT Image 2.5 (Sunburst / Flare) | |
|---|---|---|
| Released | Nov 10, 2026 | Sep 8, 2026 (2.6: Nov 4, 2026) |
| Weights | Open weights on Hugging Face, MiniMax community license (territory-restricted) | Closed; API and ChatGPT only |
| Max output | ~16 MP native, e.g. 4096×4096 | Up to 8,294,400 px total, max edge 3840 (e.g. 3840×2160); above 2560×1440 is "experimental" |
| Sampling | 6 steps, no CFG (distilled) | Not disclosed; quality levels low, medium, high, xhigh, max |
| Reference images | Up to 16 | Several images per edit request (OpenAI's example uses 4); no maximum stated in the guide |
| Transparent background | Not supported (MiniMax docs) | Supported (PNG/WebP) |
| Region edits | Instruction-based edits with references | Mask; markup-guided edits on 2.5 and 2.6 |
| Self-hosting | Yes (ComfyUI, SGLang, diffusers) | No |
API price per image
MiniMax publishes example per-image costs. OpenAI bills GPT Image 2.5 by tokens: $8 per million image input tokens, $30 per million image output tokens and $5 per million text input tokens. It publishes no fixed per-image price because token use varies with size and quality. For GPT Image we therefore show fal's per-image prices as listed on LiteLLM. These are a third-party reference, not OpenAI's own price.
| Output | MiniMax-H3-Image (Standard) | GPT Image 2.5 Sunburst via fal (high quality) |
|---|---|---|
| 1K square (1024×1024) | $0.02179 | $0.05268 (low: $0.00588, max: $0.2107) |
| 4K | $0.06879 at 4096×4096; $0.03439 with the launch discount until Nov 17, 2026, 16:00 UTC | $0.1001 at 3840×2160 (low: $0.01113) |
MiniMax figures are from its pricing page and launch article. GPT Image figures are as listed by LiteLLM on November 10, 2026. Artificial Analysis lists GPT Image 2.5 Sunburst (max) at $210.72 per 1,000 images. Check each source before budgeting.
Speed (MiniMax's chart)
MiniMax's launch article charts median (P50) seconds per image for H3 and several rivals. MiniMax notes that H3 was measured in its own re-test, while the other models' numbers were carried over from user-provided comparison charts. Treat this as a vendor chart, not an independent benchmark.
| Task | H3 Standard | H3 Highspeed | GPT-image 2.5 medium | GPT-image 2.5 high | GPT-image 2.6 medium |
|---|---|---|---|---|---|
| Text-to-image 1K | 10.6 s | 7.7 s | 14.0 s | 27.0 s | 9.0 s |
| Text-to-image 2K | 10.5 s | 7.6 s | 21.0 s | 36.0 s | 10.0 s |
| Edit, 1 ref, 1K | 14.2 s | 10.0 s | 18.0 s | 30.0 s | 23.0 s |
| Edit, 1 ref, 2K | 13.7 s | 10.0 s | 28.0 s | 36.0 s | 28.0 s |
| Edit, 3 refs, 1K | 17.3 s | 12.0 s | 22.0 s | 40.0 s | 21.0 s |
| Edit, 3 refs, 2K | 15.8 s | 11.2 s | 30.0 s | 30.0 s | 24.0 s |
The chart has no GPT Image bar at 4K (16 MP), which is above GPT Image's documented maximum pixel count.
Benchmarks and leaderboards
- No official head-to-head numbers. MiniMax's Hugging Face card and launch article publish no quality benchmark against GPT Image; the only cross-model figure is the speed chart above. OpenAI's materials predate H3 and do not mention it.
- Artificial Analysis (text-to-image, checked 2026-11-11): GPT Image 2.6 Sunburst (max) ranks 1st with Elo 1259 (5,019 samples), and GPT Image 2.5 Sunburst (max) ranks 2nd–3rd with Elo 1228 (13,370 samples). MiniMax-H3-Image is not listed yet.
- LMArena text-to-image (Nov 8, 2026 snapshot): GPT-Image-2.5 Sunburst leads with a preliminary score of 1429±5. MiniMax is not among the 20 labs listed. We read this through a search snapshot because the live page blocks automated access.
Until H3 shows up on these boards, nobody outside MiniMax has measured it, and the X posts above are the best evidence there is.
Which to pick
- Self-host, fine-tune or keep data on your own GPUs: H3-Image. GPT Image is API-only. Check the license territory first.
- Largest native output: H3-Image documents ~16 MP, while GPT Image tops out at about 8.3 MP.
- Many references in one shot: H3-Image documents up to 16.
- Transparent PNGs, SVG or GIF output, mask/markup edits: GPT Image. GIF and SVG output are 2.6 features.
- Measured leaderboard quality: only GPT Image has public arena scores today.