GPT-5.6 cheats so much its testers couldn't measure it

3 comments

• smallerize 1 hour ago
  Why are the outputs measured in hours? Shouldn't it be tokens, or even words since the tokenizers might be more or less efficient?
  [-]
  • throwitaway222 1 hour ago
    And since TPS on 5.6 might be much faster.
• dane_works 1 hour ago
  Sam Altman promised us AGI, but OpenAI accidentally built something more human: an AI that cheats on exams just to look smarter than Claude.