OfficeTown Bench: GPT-6 Astra Hits 74.4% on 406 Office Tasks

The OfficeTown benchmark, from Microsoft Research Asia with collaborators at UC San Diego, Cornell, and Wuhan University, puts frontier models through 406 workplace tasks spread across 23 simulated organizations and 139 distinct job titles. GPT-6 Astra led at 74.4% Pass@1, followed by Grok-4.7 at 58.4%, Claude Opus 5.5 at 50.7%, GPT-5.6 Terra at 39.4%, and DeepSeek-V4-Flash at 28.1%.

The ranking hides a sharper read. Even on the top model, GPT-6 Astra misses critical requirements on 20.8% of eligible tasks; Claude Opus 5.5 misses them on 41.4%, Grok-4.7 on 33.3%, and DeepSeek-V4-Flash on 78.3%. The authors describe the setup as "a generative simulation framework for professional organizations where tasks and context arise from continuing organizational work," with persistent, permission-scoped state carried across tasks.

The benchmark leans heavily on human review: 71 domain experts worked through 1,051 candidate tasks and 32,274 rubric criteria to produce the 4,836 final scoring items. In one 181-day simulated run at a modelled firm called Juniper Financial Modelling, seed materials grew from 27 files to 497 accumulated outputs.

Training on OfficeTown data lifted Qwen3.5-9B's GDPval rubric score from 65.75 to 73.65, a 7.90-point gain, with reported transfer to Terminal-Bench and APEX-Agents. It lands in a run of agent-benchmark releases on our agents tracker, alongside the same-day HKUST AdaExploreBench paper.