Mistral 3 Large Delivers 92% of GPT-5.2 Performance at 15% of Cost

Mistral released Large 3 (675B parameters, MoE) achieving 92% of GPT-5.2 benchmark performance at 15% of the cost. OpenAI's GPT-5.2 offers 400K-token context; Meta's Llama 4 pushes to 10M tokens via MoE. March 2026 has produced major model releases approximately every 72 hours.