Claude Opus 4.8 crushes WorkBench: 89% tasks done, 2.5% harmful actions
From 43% to 89% in two years: AI workplace agents get safer and smarter
A follow-up study to the 2024 WorkBench benchmark reveals dramatic improvements in workplace AI agents. The best agent in March 2024 (GPT-4) managed only 43% task completion and caused unintended harmful actions on 26% of tasks. Fast forward to June 2026, and the current leader Claude Opus 4.8 achieves 89% task completion with just 2.5% harmful actions — a 46 percentage point improvement in capability and a 23.5 point drop in errors.
Three standout findings emerge from the research. First, capability and safety are not trade-offs — models that finish the most tasks also cause the least unintended damage. Second, while many error classes have been eliminated, frontier models still make basic mistakes that occasionally lead to irreversible harm, such as sending an email to the wrong person. Third, open-weight models have drastically lowered costs for performance levels once exclusive to proprietary systems, while frontier model costs remain stable. The authors have released an updated benchmark with better data and code quality, new scores, and analysis spanning two years of agent progress.
- Claude Opus 4.8 completes 89% of tasks vs. GPT-4's 43% in 2024 — a 107% improvement
- Harmful action rate dropped from 26% to 2.5%, proving capability & safety now align
- Open-weight models offer previously proprietary-level performance at drastically lower costs
Why It Matters
Workplace AI agents are now both more capable and safer, but basic errors remain a risk in critical tasks.