Claude Sonnet 5.5
17:11 min $1.94
Overall ranking
Every creation was put head to head with every other creation in its challenge, and three AI judges picked the better one each time. The rating works like a chess rating: 1000 is the average model, and there is no maximum. Select a model to see everything it made.
MR-Bench rating
From 4,920 judge votes. Higher is better; a model 200 points ahead wins about 3 votes in 4. The line around each dot is the range its rating very likely falls in.
Not ranked yet, still waiting for some of their creations: Seed 2.1 Turbo.
The 8 challenges
Pick a challenge to see every model's attempt, best rated first. Select one to run it live, read its code and see how it did in its head-to-head matchups.
3D model / three.js
Pigeon
A lifelike 3D city pigeon that bobs, blinks and pecks.
Build a detailed, realistic 3D model of a common city pigeon (rock dove) in three.js, standing on a weathered stone ledge, and present it like a museum showcase. The bird: - Model it entirely from three.js geometry built in code: no model files, no image textures. Aim for a recognisable, lifelike pigeon, not a cartoon. - Proportions: a plump, rounded body; a small head on a short, thick neck; a tail about a third of the body's length; short legs. - Colours and materials: blue-grey body and wings; two distinct dark bars across each folded wing; a dark band at the tip of the tail; an iridescent green-and-purple neck that shimmers as the light moves (use MeshPhysicalMaterial with sheen or iridescence); orange eyes with black pupils and a thin grey eye-ring; a dark grey beak with a white cere at its base; pink-red legs and feet with four toes each. - Build it from many shaped parts (body, head, beak, cere, eyes, neck, wings made of layered feather shapes, tail feathers, legs, toes) with smooth shading. Animation (loops forever, no input needed): - The head bobs forward and back as if walking in place; the bird blinks every few seconds, turns its head to look around, and every few seconds bends down to peck at the ledge. - Feathers ruffle subtly and the chest rises and falls with breathing. Scene and presentation: - Soft daylight: a warm key light, a cool fill light and real-time soft shadows on the ledge. - The stone ledge is textured with code (vertex colours or noise) and slightly worn at the edges. - Background: a soft, out-of-focus city-at-dusk gradient. - The camera slowly orbits the bird; OrbitControls let the visitor drag to look around and scroll to zoom, and the slow orbit pauses while they drag. - A small, tasteful caption in a corner reads: Columba livia - rock pigeon.
What a great answer looks like
- Instantly recognisable as a pigeon, with believable proportions
- Correct colouring: blue-grey body, dark wing bars, iridescent neck, orange eyes, pink feet
- Natural idle animation: head bob, blinking, pecking, looking around
- Lighting, shadows and the stone ledge look polished
- Runs smoothly and the camera controls work
GPT-6 Astra
4:28 min $0.61
Muse Spark 1.3
49 s $0.05
GPT-6.1 Sol
7:26 min $0.22
Claude Fable 5.1
11:45 min $2.60
GPT-6 Luna
2:24 min <$0.01
Claude Haiku 5.5
7:42 min $0.05
Kimi K3
1:00 min $0.18
Claude Opus 5.5
17:25 min $2.17
Gemini 3.8 Flash
2:01 min $0.08
Gemini 3.1 Pro
3:08 min $0.08
GLM-5.3 Flash
17:50 min $0.03
MiniMax M3
11:01 min $0.10
Grok 4.7
55:02 min
Seed 2.1 Turbo
9:30 min $0.10
Mistral Large 4
5:40 min $0.07
Nemotron 3 Ultra
4:50 min $0.02
Composer 2.5
32 s
Qwen 3.8 Flash
15:51 min $0.03
MiMo V2.6 Pro
19:36 min $0.04
DeepSeek V4 Pro
2:37 min $0.05
Speed and cost
How long each model took to finish a challenge, and what that would cost through its maker's paid API. Shorter and cheaper are better.
Time per challenge
Typical (median) time from sending the brief to receiving the finished page.
Cost per challenge
Average cost of one answer at the model's published API price, thinking included.
Rating for the money
Each logo is a model. Higher means a better rating; further left means cheaper. The best value sits in the top left.
- Claude Sonnet 5.5: rated 1399, $0.99 per challenge
- GPT-6.1 Sol: rated 1352, $0.15 per challenge
- Claude Opus 5.5: rated 1308, $1.35 per challenge
- GPT-6 Astra: rated 1183, $0.71 per challenge
- Claude Fable 5.1: rated 1106, $2.10 per challenge
- Grok 4.7: rated 1069, $0.33 per challenge
- GPT-6 Luna: rated 1046, <$0.01 per challenge
- Muse Spark 1.3: rated 1042, $0.06 per challenge
- MiMo V2.6 Pro: rated 1018, $0.04 per challenge
- Gemini 3.8 Flash: rated 1012, $0.12 per challenge
- Kimi K3: rated 995, $0.28 per challenge
- GLM-5.3 Flash: rated 986, $0.02 per challenge
- Claude Haiku 5.5: rated 983, $0.05 per challenge
- Mistral Large 4: rated 915, $0.12 per challenge
- Qwen 3.8 Flash: rated 879, $0.03 per challenge
- MiniMax M3: rated 807, $0.07 per challenge
- Gemini 3.1 Pro: rated 801, $0.09 per challenge
- DeepSeek V4 Pro: rated 772, $0.06 per challenge
- Nemotron 3 Ultra: rated 709, $0.03 per challenge
Every rating
The full table: each model's rating in every challenge, where 1000 is average. Brighter blue means a higher rating. Select a rating to open that creation, or a column title to sort.
Times and costs are per challenge. Costs use each model's published API price.
How we test
The same five steps for every model, so the differences you see come from the models themselves.
Step 1: The same brief for everyone
Every model got the same 8 detailed briefs and the same instructions, word for word, with its thinking set to high.
Step 2: Run safely and recorded
Each creation runs in a locked-down window in a real browser. We record a short video and note its frame rate and any errors. 11 answers that crashed or stayed blank were generated again, and are marked.
Step 3: Every pair, head to head
Within each challenge, every creation is paired with every other creation. Nobody gets an easier set of opponents.
Step 4: Three blind judges
Claude Opus 5.5, GPT-6 Astra and Grok 4.7 each pick the better of every pair, thinking at extra-high effort. They never see model names, and the order is swapped between judges.
Step 5: A rating with no ceiling
The votes become a rating, like in chess: beating strong rivals counts for more. 1000 is the average of today's field, and future models can climb past any number.
Test settings
The same for every model, so no one gets an advantage.
- Reasoning effort
- High, for every model
- Judges' reasoning effort
- Extra-high, for all three
- Judge votes
- 4,920, on 1,640 pairs
- Answer length limit
- 128,000 tokens, or the model's own maximum
- Window size
- 1280 x 800 pixels
- 3D library
- three.js 0.170.0
- Tested
- Oct 10, 2026
The instructions every model got
Sent word for word before each challenge's brief.
You are an expert creative front-end developer taking part in a benchmark. Reply with one complete, self-contained HTML document and nothing else: no explanations and no Markdown code fences. Start with <!DOCTYPE html> and end with </html>.
Rules:
- Put all CSS and JavaScript inline in the document.
- Do not load any external files, with one exception: tasks that use three.js must load it with exactly this import map, and may import add-ons from "three/addons/":
<script type="importmap">
{ "imports": { "three": "https://cdn.jsdelivr.net/npm/three@0.170.0/build/three.module.js", "three/addons/": "https://cdn.jsdelivr.net/npm/three@0.170.0/examples/jsm/" } }
</script>
- Do not use images, image textures, web fonts, audio files or 3D model files: create everything with code.
- Do not use localStorage, sessionStorage, cookies or network requests.
- The page must fill the browser window, adapt when the window is resized, and start animating immediately on load, without any click or key press. Only scroll the page if the task is a web page with several sections.
- The result is shown in a 1280 x 800 window and must also work on a phone.Checking the judges
Blind judging only works if the judges are fair, so we check every vote. All three judges agreed on 80.7% of the pairs.
Claude Opus 5.51,640 votes
- Sided with the other two
- 95.2%
- Picked the one shown first
- 47.7%
- Picked Anthropic's modelsthe other judges: 73.4%
- 76.3%
GPT-6 Astra1,640 votes
- Sided with the other two
- 97%
- Picked the one shown first
- 49.8%
- Picked OpenAI's modelsthe other judges: 71.1%
- 77.9%
Grok 4.71,640 votes
- Sided with the other two
- 88.5%
- Picked the one shown first
- 51.6%
- Picked xAI's modelsthe other judges: 57.6%
- 61.4%
A fair judge picks the first-shown creation about half the time, and doesn't favour its own company's models more than the other judges do. Judges never see names, so a gap there would be taste, not loyalty.
Questions
What is MR-Bench?
A test of how well AI models build things. We gave 21 AI models from 15 companies the same 8 challenges, such as a 3D pigeon, a self-solving Rubik's cube and a playable game, and compared what they made.
Who decides the ratings?
Three AI judges from three different companies: Claude Opus 5.5, GPT-6 Astra and Grok 4.7. Within each challenge, every creation is paired with every other one, and each judge picks the better of the two after seeing four moments of each, without the models' names. That came to 4,920 votes.
What does the rating mean?
It works like a chess rating. 1000 is the average model of today's field. Beating a strong rival raises a rating more than beating a weak one, and a model 200 points ahead of another wins about 3 votes in 4. The small ± number shows how sure we are.
Why not a score out of 100?
A score out of 100 runs out of room: once models get close to the top, there is nowhere left to go. A rating has no maximum, so when better models arrive next year they can simply rate higher, and today's numbers keep their meaning. There is no minimum either: a creation that lost almost every matchup in a challenge can end up near zero, or even below it.
Can a judge favour its own company?
The judges never see who made what, so they can't knowingly favour anyone. We still check: under How we test you can see how often each judge picked its own company's models compared with the other two judges.
Can I trust AI judges?
They are consistent and they give a reason for every vote, but no judge is perfect. That is why every creation is on this page: open any of them, run it live, read the judges' reasons and decide for yourself.
Why does a creation look broken?
Because that is what the model wrote. Answers are shown exactly as they came back, with no fixes. Some models got the same brief again after their answer crashed or stayed blank, and the page says how many tries it took. The viewer lists any errors it hit while running.
What do the costs mean?
What one answer would cost at the model's published API price, including the hidden thinking it did before answering. Prices change often, so treat them as a guide.
Can I try the challenges myself?
Yes. Every brief has a copy button, and the instructions sent before each brief are under How we test. Paste them into any AI and compare.