Three new frontier model families are available in the model picker. The coding results below come from CursorBench 3.2, an evaluation of agents on ambiguous, multi-file tasks. Scores and average cost per task are a July 9, 2026 snapshot and may change.
GPT-5.6 Sol, Terra, and Luna. GPT-5.6 Sol Max ranks third overall with a 67.2% score, behind Claude Fable 5 Max at 70.5% and Claude Fable 5 Extra High at 68.4%. GPT-5.6 Terra Max scores 64.9% at $2.73 per task, roughly half of Sol Max’s $5.22. Choose Luna when you want the faster, lower-cost GPT-5.6 option.
Grok 4.5. Grok 4.5 High scores 66.7% and ranks fourth overall, half a point behind GPT-5.6 Sol Max, at an average of $1.51 per task, about 3.5 times less than Sol Max. The benchmark flags this result because an earlier snapshot of its test codebase was unintentionally included in Grok 4.5’s training data, and the score impact is unclear.
Muse Spark 1.1. Use Meta’s upgraded model for coding, debugging, long-running agentic tasks, and multimodal reasoning across supported inputs.