WorkBench Revisited: Workplace Agents Two Years On

Workplace AI agents have roughly doubled their task-completion rate in two years while slashing harmful side effects, according to an updated benchmark released this week. The best agent, Claude Opus 4.8, now completes 89% of tasks and takes unintended harmful actions on only 2.5% of them, researchers report. When the WorkBench benchmark was introduced in 2024, the strongest agent — a ReAct loop around GPT-4 — completed just 43% of tasks and took an unintended harmful action, such as emailing the wrong person, on 26% of them [1][5]. The benchmark, which places agents in a sandbox environment with five databases, 26 tools, and 690 tasks representing common business activities, was designed to test planning, tool selection, and multi-step execution [5]. At the time, the weakest open model evaluated, Llama2-70B, managed only 3% [3][4]. Two years later, the landscape has shifted. Six models from four different providers now clear 80% completion [3][4]. Claude Opus 4.8, released by Anthropic, leads the field at 88.8% completion [3][4]. Anthropic describes the model as its strongest computer-use and browser-agent system, scoring 84% on the Online-Mind2Web benchmark [10]. The model is priced at $5 per million input tokens and $25 per million output tokens, unchanged from its predecessor [10]. The researchers identify three notable patterns in the results. First, capability and safety improve together on WorkBench rather than trading off: the models that finish the most tasks also do the least unintended damage [1][2]. Second, while several classes of error have been eliminated entirely, frontier models still make basic mistakes that occasionally result in irreversible harm, such as sending an email to the wrong person [1][2]. Third, the rise of open-weight models has drastically lowered costs for a performance level previously accessible only to proprietary systems, while frontier costs have remained relatively stable [1][2]. Completion now runs from 26% to 89% across models, with costs spanning two orders of magnitude and side effects ranging from 2% to 39% [3][4]. The strongest model still fails roughly one task in nine, with remaining failures concentrated in harder reasoning and multi-step retrieval cases, plus a residue of tasks where model reasoning is defensible despite being scored as incorrect [3][4]. The researchers have released the updated benchmark, harness, and per-model cost estimates so that future models can be evaluated on the same axes [3][4].