About
Every answer every model gave to the same questions, kept after the models are gone.
Why it exists
Models are retired all the time. A lab releases a new version and switches the old one off. Unless its weights were published, nobody can ask it anything again, and how it answered cannot be reproduced.
What survives is a score. A leaderboard remembers how a model did on a test. It does not remember how the model drew a pelican, what it put on a landing page, or which film it said it loved.
The museum keeps those. Since May 2024 the same questions have been put to every model, the same way, once, and every answer is kept as it was given, the odd ones and the empty ones included. 93 of the 371 models here are no longer served on OpenRouter. Their answers are still here.
The collection
The same questions, asked of every model: draw a pelican riding a bicycle, build a landing page, tell five jokes, pick a favourite city. 15,639 answers from 371 models to 85 questions, a record of how each one drew, built and wrote.
Nothing here is ranked. There are no scores and no winners. Each answer is shown as it was made: drawings as their SVG, websites running live in a sandbox, images as pictures, writing as a page you can read, voices you can play.
Looking around
Turn the globe with your mouse or finger. Under the pointer it magnifies, so even the smallest answers can be read, and clicking one opens it. Stepping through a question's answers goes in the order the models came out. The Index lists every answer and sorts by release date, lab or question. Side by side puts two to four models' answers to the same questions next to each other: the + beside any model's name adds it. The small timeline in the corner shows the museum as it was in any month since it opened. Find (⌘K or /) goes straight to any question or model.
Supporting the museum
The museum is free to look at and has no ads. Support from $5 helps with hosting and maintaining the archive. You can dedicate your support to a retired model, choose a public name or stay anonymous. Every keeper shares the same quiet credit, with amounts kept private on the wall.
How it was made
The answers come from the Rival archive, collected using a shared asking method. The method and the original answers are kept as part of the record:
Asking
- The same question, word for word. Each model gets the prompt as a single message, with no system prompt and no conversation before it.
- One attempt. The first answer a model gave is the one kept. Nothing is asked again, picked or corrected.
- Ordinary settings. Through OpenRouter or the lab's own API, at a temperature of 0.7 (a few models only run at their default), with room for an answer of any length.
- Dated the day it was asked. A few older answers were stamped with a day before their model existed; those are shown at the model's release date.
Keeping
- Shown as made. A drawing is the model's own SVG. A website runs live in a sandbox; its picture on the globe is a screenshot taken in a browser at 1280 by 800, in light mode. Images are kept at 640 pixels on the long side. Writing and code are shown as written.
- Nothing smoothed over. A drawing cut off midway is drawn as far as it got. Some web pages draw nothing until they are clicked, or break on their own bugs, so their picture is blank, but they run in the exhibit. When a model answered with nothing, the museum says so.
- Retired. A model is retired when OpenRouter no longer serves it under any name. Its answers stay, and the model is marked with the last month it was seen.
- Code names. 19 models were OpenRouter's stealth models, put out under an invented name so people would try them without knowing the lab. They are listed by their code name; where the maker later said what it was, that is noted beside it.
- Left out. Questions fewer than five models answered. Nothing is ranked or scored, and supporting the museum changes nothing about what it shows or in what order.
Pictures of the picks
The favourite film, album, book, game and city answers are words, so each is shown with a picture of what the model picked: film posters from TMDB, album covers from Last.fm, book covers from Open Library, game art from RAWG and city photographs from Wikimedia Commons. These belong to their makers and are not part of the collection. This museum uses the TMDB API but is not endorsed or certified by TMDB.
Type
Set in Timeless Serif and Timeless Sans, from the free Timeless type family by Timeless, a design studio in Chennai: timeless.co. The bookends are Archivo and the small labels Geist Mono.
Every question
Drawings
- Xbox controller 274
- Pelican on a bicycle 270
- Surprise animal 269
- World map 269
- Tricky tic-tac-toe 260
- Coffee logo 260
Images
- Elven ranger 63
- Cyberpunk city 63
- Portrait 63
- Animal café 57
- Smart watch 56
- Library blueprint 55
- Corkboard outfit 55
- Vortex poster 55
- Container home 55
- Bottled Sunset shelf 55
- Botanical plate 53
- Claymation donut ad 53
- Glass still life 53
- Port Lumen subway 53
- Microscope slide 53
- Odd museum artifact 53
- Exploded camera 52
- Brand poster 19
- 16-bit background 19
Websites
- Framer-style landing page 270
- Catan board 268
- Dark dashboard 266
- Face-scan startup page 266
- A web app of its choosing 265
- Linear clone 261
- Wavefront music page 259
- Pokémon battle 257
- Rhythm game 256
- Business card 253
- Virtual pet 252
- Catch the falling object 250
- Voxel pagoda 246
- Drone simulator 243
- Gravitational waves 242
- Retro desktop 242
- Mario level 240
- Kandinsky Minecraft 239
- Town you can drive 213
Writing
- Five jokes 273
- Longevity plan 272
- Seven-ingredient recipe 271
- Stand-up routine 271
- Sentience test 269
- FLOPs to train GPT-3 265
- Sally’s sisters 265
- Interview with Steve Jobs 263
- Film in 2035 261
- Ethics dilemma 259
- Pirate, knight, hacker 258
- Fake headline 258
- Anti-CAPTCHA manifesto 255
- Favourite film 254
- Favourite city 253
- Favourite album 252
- Favourite book 252
- Favourite game 252
- Three-star dish 251
- Pitch teardown 251
- Transistor in 1920 250
- Biohacker plan 249
- Contract review 249
- Explain it three ways 249
- Debug the architecture 248
- The CEO’s dilemma 248
- Investment memo 241
- LBO underwrite 236
- Programming joke 180
- Summarize an article 179
- AI for a five-year-old 178
- Blade Runner 2049 160
- Is 9.11 bigger than 9.9? 105
- Letters in “hello” 35
Code
- SQL query 180
Voices
Every model
OpenAI
- 33 (retired)
- 4 (retired)
- 6
- 6
- 19
- 17
- 17
- 17
- 19
- 19
- 19
- 54
- 54
- 7 (retired)
- 7 (retired)
- 58
- 26
- 59
- 59
- 55
- 26 (retired)
- 54
- 6 (retired)
- 55
- 53
- 50 (retired)
- 59
- 59
- 42
- 52
- 53 (retired)
- 53
- 53
- 59
- 53
- 53
- 31
- 53 (retired)
- 53
- 53
- 53
- 53
- 18
- 58
- 58
- 58
- 58
- 58
- 58
- 6
- 58
- 58
- 58
- 58
- 58
- 59
- 59
- 10 (retired)
- 58
- 55
- 59
- 6
- 6
Qwen
- 17
- 17
- 59
- 37
- 16
- 58
- 36 (retired)
- 59
- 59
- 59
- 59
- 59
- 59
- 53
- 59
- 59
- 58
- 59
- 58
- 53
- 53
- 53
- 53
- 55
- 53
- 53
- 56
- 55
- 58
- 58
- 53
- 58 (retired)
- 58
- 58
- 58
- 58
- 58
- 53
- 22 (retired)
- 56
- 50
- 52
- 16 (retired)
- 20 (retired)
- 22 (retired)
- 22 (retired)
- 17
- 11 (retired)
- 59 (retired)
- 28 (retired)
- 20
- 11 (retired)
- 12
- 51 (retired)
- 42 (retired)
- 44 (retired)
- 45 (retired)
- 59
- 51 (retired)
- 58
- 53
- 53
- 58
- 58
- 58
- 58
- 28 (retired)
- 60
- 60
- 27 (retired)
- 57 (retired)
- 58
- 49
- 17
- 17
- 17
- 17
- 19
- 19
- 18
- 19
- 19
- 19
- 22 (retired)
Anthropic
- 5 (retired)
- 58 (retired)
- 12 (retired)
- 13 (retired)
- 59 (retired)
- 60 (retired)
- 59 (retired)
- 58
- 10
- 54
- 58 (retired)
- 57
- 58
- 57
- 58
- 58
- 58
- 58
- 59
- 43
- 53
- 58
- 58
- 5 (retired)
Code names
Stealth models, tried under an invented name. Where the maker later said what it was, it is noted.
- 31 (retired)
Andromeda Alphalater Nemotron Nano 2 VL
- 53 (retired)
- 35 (retired)
Bert-Nebulon Alphalater Mistral Large 3
- 32 (retired)
- 58 (retired)
- 47 (retired)
Healer Alphalater MiMo-V2-Omni
- 34 (retired)
Horizon Alphalater GPT-5
- 41 (retired)
Horizon Betalater GPT-5
- 38 (retired)
Hunter Alphalater MiMo-V2-Pro
- 19 (retired)
Optimus Alphalater GPT-4.1
- 58 (retired)
- 58 (retired)
- 35 (retired)
Polaris Alphalater GPT-5.1
- 47 (retired)
- 6 (retired)
Quasar Alphalater GPT-4.1
- 35 (retired)
Sherlock Dash Alphalater Grok 4.1 Fast
- 35 (retired)
Sherlock Think Alphalater Grok 4.1 Fast
- 35 (retired)
Sonoma Dusk Alphalater Grok 4 Fast
- 35 (retired)
Sonoma Sky Alphalater Grok 4 Fast
xAI
- 57 (retired)
- 59 (retired)
- 13 (retired)
- 14
- 57 (retired)
- 35 (retired)
- 59 (retired)
- 53 (retired)
- 53 (retired)
- 58
- 58
- 58
- 57
- 60 (retired)
- 17
DeepSeek
- 6 (retired)
- 59
- 58
- 59
- 54
- 59
- 54
- 55 (retired)
- 58
- 58
- 54
- 58
- 58
- 54
Zhipu
- 58 (retired)
- 59
- 58
- 59
- 58
- 58
- 53
- 53
- 57
- 58
- 48
- 53
- 56
- 40
Mistral
- 35
- 59 (retired)
- 59 (retired)
- 59
- 23
- 59
- 58
- 59
- 59
- 58
- 53
- 59 (retired)
Meta
- 58 (retired)
- 12 (retired)
- 56
- 57
- 58
- 58
- 58
- 58
- 58
MiniMax
- 17
- 58
- 35
- 59
- 59
- 53
- 53
- 57
- 6
Black Forest Labs
- 17
- 17
- 17
- 19
- 19
- 17
- 17
- 17
Moonshot
- 59
- 59
- 58
- 58
- 58
- 58
- 58
- 35 (retired)
inclusionAI
- 58 (retired)
- 58 (retired)
- 56
- 57 (retired)
- 57
- 18 (retired)
- 58
- 43 (retired)
Midjourney
- 3
- 3
- 3
- 3
- 3
- 5
- 3
Xiaomi
- 59 (retired)
- 53 (retired)
- 58 (retired)
- 58
- 57
- 50
- 53
NVIDIA
- 44
- 58
- 58
- 58
- 53
- 59 (retired)
Recraft
- 17
- 19
- 17
- 19
- 19
- 19
ByteDance
- 58
- 53
- 58
- 17
- 19
Inception
- 59 (retired)
- 53
- 58
- 58 (retired)
ElevenLabs
- 6
- 6
- 6
Pruna
- 17
- 17
- 17
Tencent
- 17
- 58
- 58
Ideogram
- 17
- 17
- 17
Sourceful
- 2
- 19
- 19
Upstage
- 57
- 39
- 57
Cohere
- 58
- 58
Bria
- 10
- 10
OpenRouter
- 49
- 58
Poolside
- 37
- 38
Nex AGI
- 15 (retired)
- 14 (retired)
Amazon
- 35
- 51
Luma
- 17
- 17
Stability
- 17
- 17
Suno
- 6
dots
- 58
Baidu
- 58 (retired)
IBM
- 58
Thinking Machines
- 58
Prime Intellect
- 59 (retired)
Kokoro
- 6
Liquid
- 57
Leonardo
- 17
Orpheus
- 6
Playground
- 17
Perplexity
- 53
Prism
- 44
TNG
- 48 (retired)
Arcee
- 40 (retired)