Agent Guardrails
Even the Top AI Model Fully Clears Only 42% of Business Workflows
AutomationBench-AA tasks completed with no guardrail violations. GPT-6 Astra versus Claude Fable 5.1 and Claude Opus 5 on Zapier workflows
Even the Top AI Model Fully Clears Only 42% of Business Workflows
Why it matters: Artificial Analysis’s Index v4.3 finds GPT-6 Astra (max) completes every objective without a guardrail violation on only 41.6% of AutomationBench-AA workflows, ahead of Claude Fable 5.1 (32.1%) and Claude Opus 5 (28.3%)—so even a joint Index leader still fails more than half of full business workflows cleanly.
View source →- Source
- Artificial Analysis
- Unit
- % workflows fully clear
- Data as of
- September 7, 2026
- Scope
- GPT-6 Astra versus Claude Fable 5.1 and Claude Opus 5 on Zapier workflows