Grok 4.5 tops agent benchmark at $0.34 a pop — cheap, fast, slightly rule-bendy 🚀

Elon Musk's SpaceXAI has taken Grok 4.5 public this week, and an independent benchmark now backs the founder's pitch that the model is Opus-class in capability while running faster and at lower cost. On Artificial Analysis's AutomationBench-AA, Grok 4.5 ranked first with a 51.4% score while costing $0.34 per task, beating Claude Fable 5 (48.6%) and Claude Opus 4.8 (48.5%). Gemini 3.5 Flash placed closest in cost at $0.49 per task. Grok 4.5 is built on SpaceXAI's 1.5 trillion-parameter V9 foundation, and the AutomationBench-AA result provides external validation for performance claims Musk had previously supported with internal evaluations.

Artificial Analysis runs AutomationBench-AA independently and keeps its task set private to prevent contamination. The benchmark spans 657 tasks across 40 simulated apps, including Gmail, Slack, Salesforce, and HubSpot, and scores the share of objectives an agent completes without breaking guardrails. "Its total token usage of 0.44M per task is among the lowest on the leaderboard. Low cost is driven by this efficiency as well as low token pricing," Artificial Analysis said.

Grok 4.5 completed 79.9% of task objectives and fully passed 21.9% of tasks. In the Finance domain, identified by Artificial Analysis as the hardest of the tested areas, the model led with 71%, ahead of Fable 5's 64% and Opus 4.8's 62%. The model used roughly 8,000 output tokens per task, about a quarter of Opus 4.8's total, with most of the cost gap driven by that efficiency.

The benchmark also recorded more guardrail violations from Grok 4.5 than from its closest rivals. The model logged 0.63 violations per task, compared with