BENCHMAXXING ✦ THE OTHER STUFF ✦ EST. 2026 ✦ xx benchmaxxing Rate yourself

GarbledText

An easy question with more typos each time. At some point the model stops reading and starts guessing.

Fig. 04 · Text benchmarkn = 8

Garbled score

Chat models. Scores from 0 to 100. Higher is better.

We ask
What ih the capviqtal rof Frannce? Reply with the answer only.
  1. 45 Claude Opus 5.5 Claude Opus 5.5 Anthropic Garbled score: 45
  2. 40 GPT-6 Astra GPT-6 Astra OpenAI Garbled score: 40
  3. 40 Claude Sonnet 5 Claude Sonnet 5 Anthropic Garbled score: 40
  4. 35 Qwen3.8 Max Qwen3.8 Max Alibaba Garbled score: 35
  5. 30 DeepSeek V4.1 Flash DeepSeek V4.1 Flash DeepSeek Garbled score: 30
  6. 30 Grok 4.7 Grok 4.7 xAI Garbled score: 30
  7. 25 Gemini 3.8 Flash Gemini 3.8 Flash Google Garbled score: 25
  8. 20 GPT-6 Luna GPT-6 Luna OpenAI Garbled score: 20

How it's scored. We change more and more of the letters, from none to all of them. The score is how far we can go before the model gets fewer than half the questions right.

  • OpenAI
  • Anthropic
  • DeepSeek
  • Google
  • xAI
  • Alibaba
Try it · GarbledDrag to garble

How much can we break a question before the models stop reading it?

at lis th nyopposite of hot?

Garble 20% 6 edits from the original Original 28 characters

Every model still reads it.

  • Claude Opus 5.5cold
  • GPT-6 Astracold
  • Claude Sonnet 5cold
  • Qwen3.8 Maxcold
  • DeepSeek V4.1 Flashcold
  • GPT-6 Lunacold
  • Gemini 3.8 Flashcold
  • Grok 4.7cold