It's telling that even with virtually unlimited access to Grok 4.1, xAI and Cursor ultimately went with Opus 4.5 instead. Says a lot about the actual gap between different models in real-world performance. The choice hints at something deeper—test-time compute alone doesn't solve everything. Sometimes raw inference power takes a backseat to whatever edge another model brings to the table.