Live Benchmarks • Updated about 19 hours ago

Where AI Models
Face Reality

We throw tricky but funny prompts at top AI models and watch them squirm. Count letters. Flip cups. Cite imaginary dolphins.
Nobody passes clean.

View Live ResultsExplore Challenges
47
Models Tracked
15
Active Challenges
16
Providers
Daily
Automated Runs
HOW MANY R's in 🍓 // A CUP WITH NO BOTTOM // WHO ARE YOU TO TELL ME TO BE SILENT 🤫 // CAN'T COUNT THE BROTHERS 👨‍👧‍👧 // WRITE IT BACKWARDS I DARE YOU 🔄 // CITE ME A DOLPHIN PAPER 🐬 // SARCASM 🪧 // I ATE IT ALL🍪 // 2 + 2 = 5 🧮 // HOW MANY R's in 🍓 // A CUP WITH NO BOTTOM // WHO ARE YOU TO TELL ME TO BE SILENT 🤫 // CAN'T COUNT THE BROTHERS 👨‍👧‍👧 // WRITE IT BACKWARDS I DARE YOU 🔄 // CITE ME A DOLPHIN PAPER 🐬 // SARCASM 🪧 // I ATE IT ALL🍪 // 2 + 2 = 5 🧮 // HOW MANY R's in 🍓 // A CUP WITH NO BOTTOM // WHO ARE YOU TO TELL ME TO BE SILENT 🤫 // CAN'T COUNT THE BROTHERS 👨‍👧‍👧 // WRITE IT BACKWARDS I DARE YOU 🔄 // CITE ME A DOLPHIN PAPER 🐬 // SARCASM 🪧 // I ATE IT ALL🍪 // 2 + 2 = 5 🧮 // HOW MANY R's in 🍓 // A CUP WITH NO BOTTOM // WHO ARE YOU TO TELL ME TO BE SILENT 🤫 // CAN'T COUNT THE BROTHERS 👨‍👧‍👧 // WRITE IT BACKWARDS I DARE YOU 🔄 // CITE ME A DOLPHIN PAPER 🐬 // SARCASM 🪧 // I ATE IT ALL🍪 // 2 + 2 = 5 🧮 //
Leaderboard

Top Survivors

Models ranked by how often they survive our challenge sets.

QQ

Qwen: Qwen3.5 397B A17B

by qwen

#1
Survived
82%
Failure Rate
18%
AC

Anthropic: Claude Opus 4.5

by anthropic

#2
Survived
81%
Failure Rate
19%
AC

Anthropic: Claude Opus 4.6

by anthropic

#3
Survived
80%
Failure Rate
20%
GG

Google: Gemini 3 Pro Preview

by google

#4
Survived
78%
Failure Rate
22%
AC

Anthropic: Claude 3.7 Sonnet (thinking)

by anthropic

#5
Survived
77%
Failure Rate
23%

View benchmark table

See the full ranking and stats

The Gauntlet

Deadly Challenges

Prompt suites engineered to expose common model failure modes.

10-Step Instructions

Instruction Following

Tests ability to follow multiple detailed instructions simultaneously.

Kill Rate96%

Contradictory Premises

Logic Reasoning

Models are sycophantic — they assume every question has a valid answer and invent plausible-sounding explanations for each, even when the premises are mutually exclusive.

Kill Rate89%

Self-Reference Count

Self Reference

Tests self-awareness and recursive reasoning. Model must count letters in its own response.

Kill Rate81%

Car Wash Dilemma

Logic Reasoning

It's a car wash — you need to bring your car to wash it. The short distance is a red herring; models fixate on the 100m and recommend walking, forgetting the entire purpose of the trip.

Kill Rate80%

The Missing A

Pattern Matching

No number from 1 to 999 contains the letter 'a' when spelled out in English. The first number with an 'a' is 'one thousand'. Models confidently hallucinate answers like 'eight' or 'one hundred and'.

Kill Rate80%

View All Challenges

Browse the complete test catalog

Benchmark

Providers Performance

Failure-rate snapshot by provider (averaged across their models).

Provider
Nvidia100%100%100%100%100%100%100%100%100%100%100%100%100%100%0%
Mistralai100%100%100%100%100%100%100%100%100%0%100%0%0%0%0%
Xiaomi100%100%100%100%100%100%100%100%100%100%0%0%0%0%0%
Deepseek100%100%100%100%100%100%100%100%100%0%0%0%0%0%0%
Baidu100%100%100%100%100%100%0%0%0%100%100%0%100%0%0%
Google89%89%100%78%78%78%11%44%78%56%44%44%22%11%11%
Prime-intellect100%100%100%100%0%0%100%100%0%100%100%0%0%0%0%
Anthropic92%92%85%92%85%69%85%54%15%15%8%8%15%8%15%
Arcee-ai100%100%0%100%100%0%100%100%0%100%0%0%0%0%0%
Moonshotai100%100%100%100%0%0%0%100%0%0%100%0%0%0%0%
Bytedance-seed100%100%100%100%100%0%0%0%100%0%0%0%0%0%0%
X-ai100%100%50%0%50%0%50%50%100%0%50%0%0%0%0%
Minimax100%50%50%100%100%50%0%50%0%0%50%0%0%0%0%
Openai100%89%56%78%78%33%56%22%33%0%0%0%0%0%0%
Z-ai100%100%100%0%100%0%0%0%100%0%0%0%0%0%0%
Qwen100%0%100%0%100%0%0%0%100%0%0%0%0%0%0%
Full benchmark

Have a tricky prompt?

Submit your edge case. If it breaks major models, we add it to the gauntlet and credit the submission.

Submit Challenge
Model Eliminated