Abstract
In March 2024, the top agent on WorkBench, GPT-4, completed 43% of tasks and took an unintended harmful action, such as emailing the wrong person, on 26% of them. Revisiting the benchmark in June 2026, we find that the best agent to date, Claude Opus 4.8, completes 89% of tasks and incurs unintended harmful actions on only 2.5%.
Key Findings
- Capability and Safety: On WorkBench, capability and safety go hand in hand rather than trade off. Models that complete the most tasks also tend to cause the least unintended damage.
- Elimination of Error Types: While several classes of error have been completely eliminated, frontier models still make basic mistakes that can occasionally lead to irreversible harm, such as sending emails to the wrong person.
- Rise of Open-Weight Models: The emergence of open-weight models has drastically reduced costs for performance levels previously only accessible to proprietary models, while frontier costs have remained relatively stable.
We release an updated version of the benchmark with data and code quality improvements, new model scores, and an analysis of agent progress on WorkBench since 2024.
Blogger's Review: This article highlights significant advancements in the intelligent agent domain, particularly in the synergy between capability and safety. Although errors still occur, the evolution of technology is clearly reshaping the landscape of workplace intelligence. Future research should continue to focus on further reducing errors and enhancing user trust.