China's office agents: 20% of users burn 87% of the compute
A million-user desktop agent just opened its logs, and the numbers say agent adoption is nothing like the chatbot curve. Also: OpenAI takes its newsroom AI program to Ukraine.
China's first office-agent behavior report is out, and usage is extremely top-heavy. Published in Beijing on September 7, the report draws on real user data from LobsterAI, an open-source desktop office agent from a major Chinese tech company that passed one million users in August. The headline finding: the top 20% of users consumed 87.4% of the compute, with the top 5% alone accounting for 53.5% of token consumption. That is not a broad, shallow user base dabbling with a chatbot — it is a power-user curve where a small group has rewired how it works.
The second finding is the one that should travel: people are handing agents heavier work. Average task size grew 3.1x over five months, with a 53% month-over-month jump in August alone. Nearly 60% of token consumption now happens outside normal office hours — 34.3% on weekday evenings, 25.6% on weekends — and 12.2% of model calls aren't initiated by a human at all. A single scheduled instruction drives an average of 64 model calls, roughly six times the 10.6 calls of a manually started task. Coding and debugging lead the task mix at 38.5%, followed by document writing at 10.1% and data and spreadsheets at 8.8%, which suggests the people most changed by this are the ones who never wrote code before.
Two details matter beyond the usage curve. DeepSeek V4 Flash takes 52.2% of model calls — over 75% once its vision variant is counted — making the agent layer the fastest route to real workloads for new Chinese models, which reach roughly 50% user penetration within three days of launch. And the geography has escaped the tier-one bubble: Beijing leads, but the top five cities account for only about 27% of users, while 62.79% sit outside the top ten and overseas users make up 12.75%, led by the US at 2.73%.
The honest caveat is that this is one vendor's self-reported data from one product, and "incomplete report" is in its title. But the shape of it — heavy tasks, off-hours, unattended runs — matches what agent builders everywhere have been claiming without evidence. If it holds, the economics of agents look less like seat-based software and more like compute rationing: a handful of users will consume most of the capacity, and pricing will have to follow.
OpenAI, WAN-IFRA and Ukraine's regional press association launched an AI program for the country's newsrooms. Announced as a joint press release on September 7, the initiative pairs two components: a Newsroom AI Masterclass Series that began August 5, and a Newsroom AI Catalyst giving hands-on support to ten Ukrainian publishers as they build and pilot AI projects. Participants get API credits alongside the training. The Catalyst launches September 17.
The framing from AIRPPU CEO Oksana Brovko is worth taking literally — the pitch is that AI should "give journalists back time to do journalism," in a market where war, economic pressure and disrupted daily life have hammered independent outlets. This is OpenAI's latest move in a news-industry partnership strategy that now runs through WAN-IFRA globally, and it lands as the same company fights publishers over training data elsewhere. Worth watching whether the ten Catalyst pilots produce anything the wider industry can copy, or whether it stays a wartime resilience program.
What to watch: whether other labs follow OpenAI into direct newsroom funding, or leave the field to philanthropy.
Do power users getting six times the value out of agents change how you think about per-seat pricing? Tell us in the comments.
