Grok 4.7's Reasoning Dial: Nearly Right at an Eighth of the Tokens

Grok 4.7 exposes a reasoning-effort setting that controls how much deliberation happens before each response, and for agent workloads the practical question is which level to run. The evaluation in the AWS write-up ran an agent on three tasks at every reasoning-effort level, six times each: 72 runs, 67 answered correctly. At xhigh it got all 18 right; at low it got 15 of 18 while using about an eighth of the output tokens. The last number is the interesting one. The cheap setting gets close, and whether 15 of 18 is good enough depends on how costly a wrong answer is to catch — a property of the task as much as of the model.