LobsterAI report: Agent task Token size in Sept is 3.2x that of March

xiaohu · x · 2026-10-19

xiaohu breaks down LobsterAI's China Office Agent User Behavior Report with notable data points:

Task scale: Token volume per single task in September reached 3.2x that of March; design/film/creative tasks rose to 17.2% of usage.
Dispatch channel: 77% of IM-based task dispatch users use WeChat, with peak dispatching during the 7-11am commute window.
Cache economics: DeepSeek V4 models hit 95-97% cache hit rates, with effective costs as low as 4.6-7.5% of list price, vs 29-35% for GLM 5.3. The takeaway: cheap unit pricing doesn't guarantee savings — cache hits matter more.
Model share: Doubao Seedream holds 57.2% of image generation and Seedance 54.4% of video generation calls; Zhipu GLM's token share jumped 166.7% MoM on its new model.
User structure: 70% of users rely on a single model; heavy users average 2.4 models, with DeepSeek Flash + GLM 5.3 Flash the most common pairing. Monthly DAU accelerated through September (+83.1% MoM).

The author's point: watching what users actually do with agents — and what they pay — is more informative than chasing benchmark scores.

More from coding & agent