Summary of Key Points
Through two days of practical testing of Tencent WorkBuddy and Alibaba Qianwen Office’s AI services, the author conducted a “computing power analysis.” The analysis revealed that the seemingly substantial subsidies offered by these platforms are partly due to differences in model costs. Heavy users consume significantly more computing power than the cost of their subscriptions, posing a “loss bomb” for the platforms. The high cost of Agents is primarily attributed to tasks such as “remembering information” (retransmitting context) and scheduling scheduled tasks. The key to controlling costs lies in the Harness layer, which optimizes task paths and routes. The subscription model essentially relies on the assumption that users will not make full use of the services; in the future, the industry is likely to shift towards pay-per-use or result-based pricing models. Technological cost reduction will be crucial for retaining users.
1. Are Subsidies Excessive? They May Be a “Deceptive Trick” Due to Model Cost Differences
Tencent WorkBuddy offers a subsidy of 5.6 times the cost, while Alibaba Qianwen Office offers a subsidy of 2.2 times. Although it may seem like Tencent is more generous, this difference is due to the costs of their respective models:
- Tencent uses the third-party Kimi K3 model, with an average cost of $24 per million tokens.
- Alibaba uses its own developed Qwen3.8-Max model, which costs only $13.2 per million tokens—82% less expensive.
- There is a minimal performance difference between the two models (Qwen3.8-Max scores 71.48 points, Kimi K3 scores 70.68 points, a difference of 0.8 points), indicating that Alibaba’s cheaper model achieves similar results.
Conclusion: Tencent’s higher subsidy is partly due to the higher cost of its third-party model; if both used the same model, the subsidy gap would not be so significant.
2. Heavy Users: The Platform’s Least-Wanted “Loss Bomb”
As a heavy user who engages in scheduled tasks, long document revisions, and brand collaboration, the author’s estimated monthly computing power cost amounts to approximately $4,300, which far exceeds the subscription fee for the premium version of Qianwen Office ($158 per month):
- For Qianwen Office: 5,000+ tokens were consumed in two days, requiring 70,000–80,000 tokens per month, but the premium version only provides up to 64,000 tokens.
- For WorkBuddy: 7 million tokens result in 1,500–1,900 tokens, and the premium version costs $140 but still does not cover the actual usage.
The Platform’s Dilemma: Most users purchase subscriptions but do not use the services extensively (for example, Microsoft 365 Copilot is only used 20–30% of the time). However, heavy users cause significant losses for the platform—this is the paradox of the subscription model: those who are most willing to pay for premium packages consume the most and incur the highest costs.
3. What Makes Agents Costly? Not “Thinking,” but “Remembering”
The cost of Agents is not related to generating new content but rather to remembering previous information:
- Retransmitting Context: Each round of dialogue requires feeding all historical interactions (system commands, tool definitions, past conversations) back into the model. Generating new content accounts for only 5% of the effort; 95% is repetitive input. The longer the task, the higher the cost of remembering information (the 30th round costs ten times as much as the first round).
- Scheduled Tasks: A Fixed Expense: The author’s two scheduled tasks (daily briefings and duty reports) accounted for half of the total consumption. Poor path optimization can lead to significant cost differences (the first briefing took 111 minutes and cost 2 million tokens, while the second took only 20 minutes and cost 700,000 tokens).
In Simple Terms: Most of your money goes towards ensuring the AI doesn’t forget previous instructions, not towards generating new ideas.
4. The Harness Layer: The “Hidden Hand” in Cost Control
Controlling Agent costs is not a financial issue but an engineering one, particularly in theHarness layer (which optimizes task paths and routes outside of the model):
- Path Memory: The AI learns which paths are ineffective and avoids them, reducing costs by two-thirds.
- Task Routing: Simple tasks are assigned to cheaper models, while complex ones are handled by more expensive models (for example, Microsoft CodeAct reduced token consumption by 64%).
- Negative Example: The author experienced unnecessary resource usage due to a poorly designed Harness layer that lacked a fail-safe mechanism.
Advantage: TheHarness layer accumulates data over time, improving performance (e.g., the AI learns to deliver content in Word files by noon).
5. The Subscription Model is a Gamble; Future Trends Point Towards Pay-Per-Use
The subscription model assumes users will not fully utilize the services, but heavy users make it unsustainable. Industry trends are shifting towards:
- Pay-per-Use: Billing based on actual usage, similar to utilities (Qianwen Office’s token system is a gamified version of this).
- Pay-Per-Result: Charging for completed tasks (e.g., writing a briefing or revising a document).
Core Logic: Subsidies cannot fill the cost gap; only by reducing costs (e.g., from $4,300 per month to $430) can platforms retain users. The company that succeeds in this will gain a competitive advantage.
Final Conclusion
- For Content Creators/Researchers: Using scheduled reports and long document revisions, spending a few dozen dollars on computing power to produce a finished briefing is more cost-effective than hiring an intern.
- For Enterprises: Implement fail-safe mechanisms first to prevent unnecessary resource consumption before considering scaling up.
- For Heavy Users: Purchase the premium version and prepare token packs in advance; however, if subsidies stop, the monthly cost of $4,300 may be unaffordable for many.
In Summary: The success of AI service models does not depend on the amount of subsidy but on who can effectively reduce costs.