LobsterAI Report: Office Agent Task Size Up 3.2x in Six Months, DAU Growth Hits 83.1%

LobsterAI has released its first China Office Agent User Behavior Report (September edition), profiling the state of office agent usage in China across three dimensions: model costs, task structure, and user growth. Key takeaway: DeepSeek V4.1 Flash's reputation for being the cheapest comes from its cache hit rate of 97% (97.3% by another reading) — the highest online — combined with cache pricing at roughly 2% of standard rates, bringing actual costs to only about 2% of list price (another reading puts it at 6.2%). Its cost advantage is driven by technical characteristics, not just price cuts.

Confirmed

The most common model pairing is the DeepSeek Flash family with GLM 5.3 Flash, at 14.0%
September's per-task token volume reached 3.2x March's, as users hand agents ever-larger jobs; design/video/creative tasks rose as a share
During the National Day holiday, general coding/IT was the fastest-growing task category (+4.0), with learning & education and personal life tasks each up +1.6
Agent usage peak shifted from 4 p.m. to 10 p.m.; @lxfater noted office agents are becoming ordinary people's "night school for coding"
The platform's DAU has accelerated for months: month-over-month growth of +19.8%, +25.1%, and +30.6% in prior months, hitting +83.1% in September

Why it matters

The report shows office agents moving from novelty to scaled everyday use: tasks are getting heavier, users are multiplying, usage extends late into the night, and model choice shows clear cost sensitivity — cache hit rate has become the key variable determining actual spend, offering empirical guidance for model vendors on pricing and engineering priorities.

2026-10-19 ~ 2026-10-19 · 6 related posts

Primary sources

LobsterAI report: DeepSeek's 97% cache hit rate makes it China's cheapest office agent model — AlchainHust · 2026-10-19 10:50 (UTC+08:00)
LobsterAI report: Agent task Token size in Sept is 3.2x that of March — xiaohu · 2026-10-19 10:07 (UTC+08:00)
LobsterAI DAU up 83.1% in September as office agents become coding night school — lxfater · 2026-10-19 10:00 (UTC+08:00)

LobsterAI report: DeepSeek cache hit rate hits 97.3%, cutting effective cost to ~6.2% of list price — lxfater · 2026-10-19
China office agent report: coding surges over Golden Week, peak usage shifts to 10pm — lxfater · 2026-10-19
[source] LobsterAI DAU up 83.1% in September as office agents become coding night school — lxfater · 2026-10-19
[source] LobsterAI report: Agent task Token size in Sept is 3.2x that of March — xiaohu · 2026-10-19
[source] LobsterAI report: DeepSeek's 97% cache hit rate makes it China's cheapest office agent model — AlchainHust · 2026-10-19

1 near-duplicate retellings: huangyun_122