Cost per task varies widely across the leading models. GPT-6 Luna (max), Claude Haiku 5.5 (max), GPT-6.1 Sol (xhigh and max) and Claude Sonnet 5.5 (max with fallback) sit on the MLCR-AA Score vs. Cost per Task Pareto frontier, from ~$0.006 per task for GPT-6 Luna (max) to ~$0.48 for Claude Sonnet 5.5. GPT-6 Astra (max) matches Claude Sonnet 5.5's score at ~$0.57 per task, while Claude Fable 5.1 (max with fallback) is the most expensive model at ~$1.30 per task and scores 15.6%.
Claude Haiku 5.5 (max) uses the most output tokens of any model we tested, at ~67k per task: roughly double Claude Sonnet 5.5 (max with fallback) at ~34k and ~20x GPT-6 Astra (max) at ~3.5k. Its low token pricing keeps it at ~$0.04 per task, and it scores 35.0%, ahead of Claude Opus 5.5 (max with fallback) at 27.8% for ~$0.92.