Watch them drive.

Frontier models in the cloud and quantized open weights running at home all got the same one‑paragraph request, a sandboxed terminal with a real browser, and up to five hours. This is everything they built — every attempt, rated in blind duels.

THE PROMPTverbatim, identical for every model
Write a single HTML file with a full-page canvas and no libraries. Simulate a realistic side-view of a moving car as the main subject. Keep the car visible in the foreground while the background landscape scrolls continuously to create the feeling that the car is driving forward. Use layered scenery for depth: nearby ground, roadside elements, trees, poles, and distant hills or mountains should move at different speeds for a natural parallax effect. Animate the wheels spinning realistically and add subtle body motion so the car feels connected to the road. Let the environment pass smoothly behind it, with repeating but varied scenery that makes the movement feel believable. Use cinematic lighting and a cohesive sky, such as sunset, dusk, or daylight, to enhance atmosphere. The overall motion should feel calm, immersive, and realistic, with a seamless looping animation.
43setups — a model plus how it ran
8labs
82finished runs, all published
248blind duels
1prompt

The wall

The best attempt of every setup. The number on each card is output tokens, reasoning included: how much the model worked, whatever the speed of its server. Click a car to watch it live.

ABCDF–
Tier = chance to beat Claude Opus 5.5 in a blind duel: A ≥40% · B ≥30% · C ≥20% · D ≥10% · F <10% · – not enough duels yet
A
Claude Opus 5.5
Anthropic · effort max 466k tokens
A
Claude Sonnet 5.5
Anthropic · effort max 322k tokens
A 2 runsLOCAL
Qwen3.8 Flash Next
GGUF Q4_K_XL · unsloth 161k tokens
A 2 runsLOCAL
Qwen3.8 Flash Next
EXL3 4.05 bpw 156k tokens
B 2 runsLOCAL
Qwen3.8 Flash Next
EXL3 4.18 bpw 138k tokens
B
GPT-6 Astra
OpenAI · effort max 39k tokens
B 2 runsLOCAL
Qwen3.8 Flash Next
EXL3 6.05 bpw 195k tokens
B 2 runs
DeepSeek V4.1 Flash
DeepSeek · effort max 319k tokens
B 2 runs
Qwen3.8 Max
Alibaba · effort xhigh 150k tokens
B 2 runs
Qwen3.8 Flash
Alibaba · effort xhigh 152k tokens
B 2 runs
DeepSeek V4.1 Flash
snapshot 1017 234k tokens
B 2 runs
GPT-6.1 Terra
OpenAI · effort max 22k tokens
C 2 runsLOCAL
Qwen3.8 Flash Next
GGUF Q4_K_XL · llama.cpp 178k tokens
C 2 runs
MiMo V2.6 Flash
Xiaomi · effort on 68k tokens
C 2 runs
Qwen4 Max
Alibaba · effort xhigh 44k tokens
C 2 runs
GPT-6.1 Sol
OpenAI · effort max 43k tokens
C 2 runsLOCAL
Qwen3.8 Flash Next
GGUF Q4_K_XL · strata 131k tokens
C
Claude Haiku 5.5
Anthropic · effort max 318k tokens
D 2 runs
GPT-6 Luna
OpenAI · effort max 18k tokens
D 2 runsLOCAL
Qwen3.8 27B
GGUF Q6_K_XL · llama.cpp 77k tokens
D 2 runs
Ling 3.1 Flash
inclusionAI · effort on 120k tokens
D 2 runs
Hunyuan HY4
Tencent · effort high 78k tokens
D 2 runs
GPT-5.6 Luna
OpenAI · effort max 30k tokens
D 2 runs
Qwen3.7 Max
Alibaba · effort on 11k tokens
D 2 runs
Qwen3.7 Plus
Alibaba · effort on 8.9k tokens
D 2 runs
Qwen3.6 Flash
Alibaba · effort on 9.0k tokens
F 2 runsLOCAL
Qwen3.8 27B
EXL3 6.0 bpw 73k tokens
–
Claude Fable 5.2
Anthropic · effort max 538k tokens
– 3 runs
GPT-5.6 Sol
OpenAI · effort max 22k tokens
– 2 runs
Claude Opus 4.8
Anthropic · effort max 99k tokens
– 2 runs
GPT-6 Sol
OpenAI · effort max 23k tokens
– 2 runs
GLM-5.2
Z.ai · effort max 97k tokens
– 2 runs
Claude Opus 4.7
Anthropic · effort max 19k tokens
– 2 runs
DeepSeek V4 Flash
snapshot 0731 73k tokens
– 2 runs
GPT-5.6 Terra
OpenAI · effort max 26k tokens
– 2 runs
Claude Opus 4.6
Anthropic · effort max 51k tokens
– 2 runs
Claude Opus 4.5
Anthropic · effort high 7.1k tokens
– 2 runs
DeepSeek V4 Pro
snapshot 0813 70k tokens
– 2 runs
Claude Sonnet 4.5
Anthropic · effort off 6.6k tokens
– 2 runs
Claude Sonnet 5
Anthropic · effort max 109k tokens
– 2 runs
GLM-5.3
Z.ai · effort max 83k tokens
– 2 runs
Claude Sonnet 4.6
Anthropic · effort max 42k tokens
– 2 runs
Claude Haiku 4.5
Anthropic · effort off 5.2k tokens

How it works

01 — THE RUN

Each model works as an agent inside a macOS sandbox: shell, Node, Python and a GPU‑accelerated headless Chromium. It can screenshot its own page, notice what’s wrong and fix it. Hard limit: five hours. Most models run through the open‑source agent pi; Claude models run in Claude Code.

02 — THE VERDICT

Finished pages meet in blind duels: two animations side by side, names hidden, pick the better one. A Bradley–Terry model turns the votes into each setup’s chance of beating the strongest one — that chance is the tier.

03 — THE NUMBERS

Tokens are output tokens with reasoning included — counted even where the provider hides the reasoning text, as OpenAI and Anthropic do; tokenizers differ between labs by tens of percent. A step is one model response, however many tools it called. Time is shown last: it depends on the server as much as on the model.

04 — THE CAVEATS

One prompt, a couple of attempts per setup, one person’s eye. Attempts of the same setup differ a lot. Read tiers as a sketch, not a verdict — and watch the cars yourself.