The state of AI image generation.

Summary

GPT Image 2.6 Sunburst is the overall leader at 1410 Elo. It also leads eight of the ten categories. Its strongest advantages appear in complex compositions, architecture, portraits, and surreal visual synthesis.

Overall leaderboard: leading group

Read the close placements sensibly: third and fourth differ by only 9 Elo. Positions five through eight span just 12 Elo.

Key discoveries

• The overall winner is not the best choice for every deliverable. GPT Image 2.5 Sunburst leads both text-in-image work and graphic design. The latter category exposes substantial edge and background defects in several newer isolated-asset outputs.
• Semantic accuracy separates leaders from merely attractive images. Correct relationships, actions, material transformations, and object counts matter more than abundant texture.
• Specialists can outperform their overall position. Seedream 5.5 is close to the anime category leader. Muse Image is especially competitive for portraits and hands.
• Coverage can radically change a recommendation. A high category rating based on one successful image is not dependable evidence of broad capability.

Quick conclusions

• Quality-first, intricate scenes: GPT Image 2.6 Sunburst.
• Typography and polished design work: GPT Image 2.5 Sunburst.
• Broad, relatively fast alternatives: Grok Imagine 2.0, Muse Image, and Nano Banana 2.1, with task-specific differences.
• Rapid concept exploration: Nano Banana 2 Lite, provided the output receives a correctness check.

General Analysis & Useful Insights

1. The leader's advantage is substantial—but not universal

GPT Image 2.6 Sunburst leads its nearest overall rival by 91 Elo, equivalent to approximately a 63% expected comparison win rate under the Elo calibration—not an observed win percentage.

There is still uncertainty. Its supplied rating interval, rounded to whole numbers, is 1356–1471 Elo, versus 1291–1367 Elo for GPT Image 2.5 Sunburst. The intervals overlap slightly. The point-estimate lead is meaningful, but the exact ordering is not statistically guaranteed by these intervals alone.

Its strongest work combines three qualities:

• Complete interpretation: requested subjects and actions remain visible.
• Selective detail: intricate content supports the focal story rather than overwhelming it.
• Integration: lighting, materials, scale, and contact points belong to the same scene.

The evaluation describes these strengths in the Busy market, Machiya architectural cutaway, and Miniature city inside a snail shell, each scored 9/10.

However, the same model scores only 4/10 on Horse riding an astronaut, because it depicts conventional riding. Excellent rendering does not eliminate relational errors.

2. Model profiles differ more than the overall table suggests

• GPT Image 2.5 Sunburst: the strongest broad design alternative. It leads graphic design and text-in-image Elo, while remaining strong in surreal imagery and environmental illustration. It completed 98 of 100 requests.
• GPT Image 2.6 Flare: strong scenes, uneven fine anatomy. It ranks second in architecture and complex scenes, but its lower hands and anime ratings reflect consequential local failures. The Hand holding a red apple scores 3/10 for an extra digit, despite polished surfaces.
• Grok Imagine 2.0: a balanced practical contender. Its 1229 Elo is only 4 points above Muse Image. It performs particularly well in difficult mixed requirements and disciplined graphic assets, with 99/100 outputs scored.
• Muse Image: strong people and tactile interactions. Its portrait and anatomy standings are more impressive than its broader architectural performance. It also delivered a scored output on every prompt.
• FLUX 3 Image: a useful technical and typography specialist. It places third in text-in-image Elo and shares third in hands and anatomy. However, ten unscored attempts and uneven coverage of named-style requests matter operationally.
• Nano Banana 2.1: a strong faster-generation option. It combines 1199 Elo with 99/100 scored outputs. Its strengths include anatomy, selected graphic systems, and several highly successful individual illustrations.
• Seedream 5.5: photographic restraint and strong selected illustrations. Its portrait coherence score is 9.0/10, but adherence is less reliable. The Businesswoman headshot scores 5/10 because the required glasses are missing.

3. Better-looking is not always more correct

A recurring distinction is surface quality versus task completion.

In the ultra-hard category, Flux 1.1 Pro Ultra averages 7.9/10 for technical quality but only 3.9/10 for adherence. In surreal prompts, Recraft V3 averages 7.8/10 technically but 4.2/10 for adherence. These are diagnostic gaps, not alternative rankings.

The common errors fall into five groups:

This is a qualitative association, not a computed correlation: as relational and functional requirements increase, visual polish alone becomes a weaker guide to success.

4. Version progression is task-dependent

Overall improvement does not imply improvement everywhere.

GPT Image 2.6 Sunburst surpasses GPT Image 2.5 Sunburst overall, yet trails it in graphic design: 988 versus 1261 Elo. Written evaluations repeatedly identify perimeter contamination, background problems, and over-rendering in isolated assets such as the Coffee-brand logo, Banking icon set, and Weather app icon.

These recurring defects suggest a delivered-asset problem. The data does not establish whether the underlying cause is generation, background handling, or another processing step.

Likewise, Midjourney v7 trails Midjourney V6.1 overall, 667 versus 725 Elo. Its atmosphere and decorative ambition often remain attractive, but exact text, clear anatomy, and specific subject relationships are frequent weaknesses in this benchmark. That is not a universal judgment about artistic preference.

5. Availability deserves its own decision rule

The strongest complete-coverage models near the top include Muse Image and Qwen-Image-3.0-Pro, both with 100/100 scored outputs. By comparison, MAI-Image-2.6 has 85/100 and Seedream 5.0 Pro has 87/100.

The sharpest warning appears in the Ghibli category: MAI-Image-2.6 ranks second at 1486 Elo, but has only one scored output out of ten. GPT Image 2 also has only one scored output there. Neither should be treated as a dependable category-wide recommendation from this evidence.

Moderation blocks cluster around some named-style and character requests, while other failures are technical. Error messages describe this run's behavior; they are not evidence of universal provider policy or legal restrictions.

6. Latency can change the practical shortlist

Recorded successful generations for GPT Image 2.6 Sunburst commonly take roughly two to three minutes. GPT Image 2.6 Flare often operates in roughly 30–55 seconds. Nano Banana 2 Lite frequently records about 2.5–4 seconds.

Two matched-task illustrations are useful:

• Muse Image produced its 10/10 Professional headshot with glasses in 9.95 seconds. The corresponding 10/10 Professional headshot with glasses from GPT Image 2.6 Sunburst took 132.39 seconds. Equal integer scores do not establish equal pairwise quality, but this shows why a faster specialist can be attractive.
• Nano Banana 2 Lite produced a 9/10 Robot painting its self-portrait in 3.21 seconds.

Timing is not a controlled inference benchmark. Several Midjourney V6.1 records are near one millisecond, while many ChatGPT 4o records are almost exactly five seconds. Other models have very long outliers. These anomalies make simple speed rankings unsafe. No cost-efficiency conclusion is possible without pricing.

Best Models by Use Case / Category

Category leaderboard at a glance

Ordered by category Elo. Close ratings should be read as a competitive group, not a decisive podium.

People, portraits, and anatomical interactions

Choose GPT Image 2.6 Sunburst for quality-first portrait assignments. Its category advantage is supported by natural facial surfaces, accurate physical traits, and emotionally specific expressions. The Heterochromia headshot and Bride smiling through tears both score 10/10.

For faster professional portraits, Muse Image is a strong specialist. It sits only 25 Elo behind second-place GPT Image 2, making those alternatives close by the stated threshold.

Hands remain a separate challenge. The leader's 40-point advantage over Muse Image is relatively modest. Nano Banana 2.1 and FLUX 3 Image share the next rating tier, but their strengths differ:

• Nano Banana 2.1 handles the Handshake and Hand holding an apple strongly, both at 9/10.
• FLUX 3 Image succeeds on the Extended-limb yoga pose, scored 9/10, but struggles with the High-five, scored 5/10.

Special case: MiniMax-H3-Image, at 1161 Elo in anatomy, delivers strong individual Running action and Front-and-back mirror portrait results. These are useful niche examples, not evidence that it leads the category.

Text, branding, and reusable design assets

Use GPT Image 2.5 Sunburst as the first quality-oriented choice for text integrated into scenes. Its advantage involves more than spelling: hierarchy, material plausibility, and supporting composition also matter. The Movie poster and Technology magazine cover each score 9/10.

FLUX 3 Image is a strong alternative for selected product and display typography. Its Birthday cake inscription scores 10/10, while its Printed T-shirt scores 9/10.

For graphic design, the top three—GPT Image 2.5 Sunburst, GPT Image 2.5 Flare, and Grok Imagine 2.0—are separated by only 5 Elo. Treat them as effectively tied and choose by deliverable and workflow.

Useful task-level distinctions:

• Flat icon families and mascots: Grok Imagine 2.0 has strong Banking icons and HelperBot mascot examples, both scored 9/10.
• Botanical typography: Nano Banana 2.1 and MiniMax-H3-Image produce strong Vine-built GROWTH lettering and Vine-built GROWTH lettering, respectively.
• Repeating decorative surfaces: FLUX 3 Image produces a strong Art Deco pattern. Even a highly rated candidate still needs an actual tiled-edge test.

Production advice: inspect edges, contrast at target size, exact copy, and unwanted text. Vector-like appearance is not proof of editable vector geometry.

Anime, cartoons, and film-specific environmental illustration

The anime leader, GPT Image 2.6 Sunburst, is only 25 Elo ahead of Seedream 5.5. Both belong on the shortlist. Coverage is incomplete: they scored 7/10 and 8/10 prompts, respectively. Seedream 5.0 Pro, at 1331 Elo, is also close to the runner-up but has only 7/10 scored outputs.

Seedream 5.5 is particularly compelling for polished character-and-environment illustrations. Its Samurai under cherry blossoms and Cat-and-dog adventure each score 10/10. Its Ramen-shop scene, however, scores 6/10 because the central eating interaction is less resolved.

For the dedicated Ghibli category, GPT Image 2.6 Sunburst leads, with eight scored outputs. GPT Image 2.5 Sunburst is a strong, better-covered alternative at nine scored outputs. The leader's Enchanted bathhouse scores 10/10.

For complete coverage in this run, consider Muse Image at 1321 Elo and Nano Banana 2 Lite at 1298 Elo. They are close and both scored 10/10 prompts. The latter is especially attractive for fast exploration. Do not mistake MAI-Image-2.6's second-place rating for demonstrated breadth: it is supported by only one scored category image.

Crowded scenes and surreal concept fusion

GPT Image 2.6 Sunburst is the strongest default for dense, multi-action compositions. Its complex-scenes lead over the runner-up is 242 Elo. The evaluation highlights readable interactions in the Family kitchen, Varied-activity classroom, and Night festival.

GPT Image 2.6 Flare is the next category-rated option and often much faster. Grok Imagine 2.0, at 1378 Elo, is also close to third-place GPT Image 2.5 Sunburst at 1388 Elo.

For surreal prompts, shortlist both Sunburst variants. Successful images make the transformation structural rather than decorative. Strong examples include GPT Image 2.6 Sunburst's Planet-shaped dessert and GPT Image 2.5 Sunburst's Cloud-made elephant.

A useful faster exception is GPT Image 2.6 Flare's 9/10 Waterfall made of galaxies. Its category rating remains third, so this is a task-specific success rather than a reason to reorder the shortlist.

Difficult semantics, technical text, and communication

GPT Image 2.6 Sunburst leads ultra-hard Elo, but Grok Imagine 2.0 is a strong alternative, particularly for the Branded mathematics lecture and Robot painting its self-portrait, both scored 9/10.

The category also exposes sharp exceptions: GPT Image 2.6 Flare's Horse riding an astronaut scores 9/10, compared with the overall leader's 4/10 on that task. Choose from the category ranking, then inspect whether the exact relationship was understood.

The ASL thank-you portrait requires extra caution. The wording's chest-position instruction conflicts with the usual chin-originating outward movement of the sign. A still image also cannot fully demonstrate motion. High scores here should not be treated as linguistic validation; educational material needs review by a fluent ASL signer.

Likewise, plausible-looking equations require mathematical checking. Readability alone does not establish correctness.

Architecture, interiors, and explanatory drawings

GPT Image 2.6 Sunburst leads architecture by 225 Elo over second-place GPT Image 2.6 Flare. It is especially useful when a brief requires both attractive presentation and explicit functional explanation.

Its Machiya cutaway, Underground bunker section, and Space habitat each score 9/10. These examples explain why it wins: rooms, circulation, systems, and requested design features are more clearly connected.

For straightforward interior photography, Seedream 5.5 deserves a task-level trial. Its Scandinavian living room scores 10/10 and Roman bathhouse scores 9/10. Its lower overall category standing reflects weaker results on other architectural requirements, not an inability to produce excellent rooms.

For glass and structural visualization, Muse Image, FLUX 3 Image, and Qwen-Image-3.0-Pro provide strong individual Glass skybridge, Glass skybridge, and Glass skybridge examples.

Final workflow recommendation: use Elo to build the shortlist, completion history to select a dependable route, and prompt-specific inspection to approve the result. For print, sign language, equations, historical reconstruction, or engineering diagrams, add the relevant domain validation before publication.