Lumenfall Arena

Leaderboards

All leaderboards
Text-to-Image
Vote All Results
Image Editing
Vote All Results
Text-to-Video
Vote All Results
Image Upscaling
Vote All Results
Image-to-Video
Vote All Results
Text-to-Vector
Vote All Results
Head to head
Esc

Models · slot A

↑↓ to navigate ↵ to pick
Arena / Models / OpenAI

OpenAI

Model evolution · 2022 – 2026
Models
9
Challenges
31
Takes
221
Evolution oldest → newest · 9 versions

How OpenAI models have evolved

One model per node, oldest to newest, each with a sample of its work. These models do different jobs, like image, video, and upscaling, so the chip on each node names its mode. Compare versions within the same mode.

  1. 2022
    DALL-E 2
    Text-to-Image
    Apr 2022
  2. 2023
    DALL-E 3
    Text-to-Image
    Oct 2023
  3. 2025
    GPT Image 1
    Image Editing
    Mar 2025
  4. 2025
    no take
    Sora 2
    Sep 2025
  5. 2025
    Sora 2 Pro
    Text-to-Video
    Sep 2025
  6. 2025
    GPT Image 1 Mini
    Text-to-Image
    Oct 2025
  7. 2025
    GPT Image 1.5
    Text-to-Image
    Dec 2025
  8. 2026
    GPT Image 2
    Text-to-Image
    Apr 2026
  9. —
    no take
    Latest
    GPT-5.5
    unreleased
Smarter every release 9 models

Intelligence over time

How OpenAI's models have grown more capable over time. Each point is one of their models at its release date and arena Elo, oldest to newest.

  • DALL-E 2
  • DALL-E 3
  • GPT Image 1 Mini
  • GPT Image 1.5
  • GPT Image 2
Head to head

Every challenge, side by side

Each model's take on every challenge in this mode. Modes use different challenges and prompts, so pick one mode at a time; each view stands on its own.

Text-to-Image 17 Image Editing 10 Text-to-Video 4
Challenge 4 challenges ↓ · 1 model →
Sora 2 Pro
Sep 2025
Text-to-Video

The Will Smith Spaghetti Challenge

1/1 · 1 take
Text-to-Video

The Will Smith Spaghetti Challenge

“Will Smith eating spaghetti” has become the unofficial benchmark of generative video for one simple reason: it is deceptively simple yet brutally revealing.

This single prompt exposes weaknesses that flashy action scenes or beautiful landscapes often hide. A model can generate stunning visuals yet still fail spectacularly when asked to make a human being convincingly eat.

Prompt
“A medium shot of Will Smith sitting at a cozy, dimly lit Italian restaurant, twirling a forkful of spaghetti with a playful grin on his face. The camera slowly dollies in as he takes a bite, capturing the rich, steaming sauce. Warm candlelight, cinematic lighting, hyperrealistic, 4K. Smooth motion, consistent face, detailed hands, natural spaghetti physics.”
Judged on
  • Will Smith being Will Smith
  • Spaghetti physics
View full challenge
Text-to-Video

Neon Rain Reverie

1/1 · 1 take
Text-to-Video

Neon Rain Reverie

This prompt is exceptionally difficult because it combines complex fluid dynamics (rain, splashing, clinging wet fabric), advanced material simulation (flowing silk + hair in wind), and atmospheric lighting; three areas where even top 2026 models still frequently produce artifacts or unrealistic behavior.

Prompt
“Hyper-realistic cinematic video of an elegant young woman in a flowing white silk dress dancing gracefully in heavy pouring rain at night on a neon-lit Tokyo street. Her long wet hair whips dramatically in the wind, the dress clings and flows with realistic fabric and water physics, raindrops splash and create perfect reflections of pink and blue neon signs on the wet pavement. Subtle emotional expression of freedom mixed with melancholy on her face, water droplets on skin and eyelashes catching the light. Smooth dynamic orbiting camera with slight cinematic handheld feel, dramatic volumetric lighting with god rays piercing through the rain, photorealistic, 8K, film grain, shallow depth of field, anamorphic lens flare.”
Judged on
  • Water & Fabric Physics
  • Lighting & Reflections
  • Body Movement & Emotion
View full challenge
Text-to-Video

The Soul Gauntlet

1/1 · 1 take
Text-to-Video

The Soul Gauntlet

This is one of the hardest remaining frontiers in 2026 video generation; testing whether models can convey genuine human emotion through subtle facial acting, realistic tear physics, and micro-expressions. While many models can create beautiful faces, very few can deliver emotionally convincing performances without looking uncanny or robotic.

Prompt
“Extreme cinematic close-up of a beautiful young woman experiencing deep, raw emotion. Her expression slowly shifts from quiet sorrow to intense cathartic crying — realistic skin texture with visible pores, subtle muscle twitches, glistening tears forming in her eyes and rolling down her cheeks, red-rimmed eyes with natural blinking and micro-expressions of pain and release. Soft dramatic side lighting with gentle rim light highlighting the tears, very shallow depth of field, slight emotional camera push-in during the emotional peak, photorealistic, 8K, intricate skin and eye details, filmic color grading, subtle film grain.”
Judged on
  • Facial Micro-Expressions & Emotion
  • Skin
  • Eye & Tear Realism
  • Emotional Arc & Timing
View full challenge
Text-to-Video

The Rubik's Gauntlet

1/1 · 1 take
Text-to-Video

The Rubik's Gauntlet

This prompt is one of the hardest single tests for 2026 SOTA video models because it simultaneously demands extreme fine-motor precision at high speed, long-term physical consistency (the cube must genuinely solve without morphing), and complex multi-element rendering (hyper-detailed skin, sweat, glossy reflections, and dynamic camera movement). Areas where even top models still frequently break down.

Prompt
“Hyper-realistic cinematic close-up of a professional speedcuber solving a 3x3 Rubik's Cube at world-record pace. His hands move with insane precision and blistering speed — fingers flying across the glossy colored faces in a complex sequence of advanced algorithms, rapid twists, and smooth layer turns. The cube rotates with perfect realistic physics, slight motion blur on fast turns, and flawless color consistency as it progresses toward a solved state. Subtle sweat glistening on skin, visible veins, hyper-detailed fingerprints and nail textures. Intense focused facial expression with micro-expressions of concentration in shallow depth of field. Dramatic cinematic side lighting with strong specular highlights and reflections dancing across the cube surfaces and skin. Smooth slow orbiting camera that circles the hands and cube, capturing every intricate finger movement from dynamic angles. Photorealistic, 8K, subtle film grain, anamorphic lens flare, moody intense atmosphere, 24fps.”
Judged on
  • Finger Dexterity & Anatomy
  • Cube Physics & Color Consistency
  • Cinematic Quality & Details
View full challenge