Harvey legal benchmark counts only hallucination-free answers: Grok 4.7 leads, GPT-6 Astra third
A new legal benchmark that filters out answers with material hallucinations puts Grok 4.7 first at 9.4%, followed by Muse Spark at 8.9% and GPT-6 Astra at 8.6%, with per-task costs compared across models.
Legal AI is heating up fast. 🚨 Harvey just released a tougher benchmark that only counts answers free of material hallucinations. Grok 4.7 leads at 9.4%. Meta’s Muse Spark sits at 8.9%. OpenAI’s GPT-6 Astra follows at 8.6%. Most models look strong until the hallucination filter hits. Muse Spark’s raw score of 26.7% collapses once errors are removed. GPT-6 models keep almost all of their points and show the fewest mistakes overall. Grok delivers the top result for about $9.50 per task. Muse Spark costs roughly $4.20. Several Claude models run $18–22. Startups like Harvey are building the tools lawyers actually use. The race is no longer about volume of output. It is about which systems can draft contracts and memos without inventing dates or citations. That gap is now measurable, and the companies closing it are scaling quickly.