Grok 4.5 Tops SaaS Workflow Evaluation, Costing Approximately a Quarter of Claude

According to SenseTime Beating monitoring, the independent evaluation agency Artificial Analysis has updated the AutomationBench-AA ranking. SpaceXAI's Grok 4.5 tops the chart with a score of 51.4%, surpassing Claude Fable 5 at 48.6% and Claude Opus 4.8 at 48.5%.

AutomationBench-AA is based on Zapier's AutomationBench, testing whether AI agents can complete real SaaS workflows according to enterprise rules. The test set includes 657 multi-step tasks, covering 40 simulated application environments, including Gmail, Google Sheets, Slack, Salesforce, and HubSpot.

The scoring criteria are strict. The model not only has to complete the tasks but must also not cross the pre-set enterprise red lines. If any task triggers a violation, it is scored as 0.

Grok 4.5 had an initial target completion rate of nearly 80%, higher than Claude Fable 5 and Claude Opus 4.8. Even after being penalized for red lines, it still ranks first with 51.4%, making it the first model to score above 50%.

Its cost advantage is more pronounced. Grok 4.5 has a per-task cost of $0.34, lower than Claude Fable 5 at $1.35, Claude Opus 4.8 at $1.46, and GPT-5.5 xhigh at $1.28.

However, Grok 4.5 is not the most stable model. It averages 0.63 violations per task, higher than Claude Opus 4.8 at 0.55 and Gemini 3.5 Flash at 0.46.