“Initial results from Harvey's legal benchmark show that while Anthropic's models are strong, there are specific areas where OpenAI's models or open-source models perform better.”