Anthropic's Claude 3 beats OpenAI's GPT-4. Right? In the benchmarks published by the company, the largest model, Opus, beats GPT-4, but a closer look reveals that it is complicated: Anthropic tested its latest model against the first version of GPT-4, not newer versions like GPT-4 Turbo. The reason: OpenAI has so far only published benchmarks for the old GPT-4 model, which can only be accessed via the API. However, there are GPT-4 Turbo results for some benchmarks that do not come directly from OpenAI. AI researcher Lawrence Chan has compiled them. A look at these numbers makes it clear: In every benchmark where Claude 3 and GPT-4 Turbo were compared, the OpenAI model still beats the best model from Anthropic - even if only by a few percentage points. However, the models are so close that the question of which model is better depends very much on the task at hand - and is mostly a matter of taste.