Alibaba Cloud QoderWork Launches “Peak-Valley Token” for the First Time — Nighttime Calls to Qwen3.7 as Low as 20%

The first “Peak-Valley Billing” mechanism in China’s Agent product sector has been implemented: bringing the load-shifting logic from the electricity market into the AI Token world.

On June 24, it was announced that last night (June 23), Alibaba Cloud’s desktop AI agent platform QoderWork officially launched the “Peak-Valley Token” billing mechanism. According to the announcement, Agents running between 22:00 and 08:00 the next day will automatically enjoy nighttime discounts, with the flagship model Qwen3.7-Max as low as 20% of the daytime rate, resulting in overall point consumption of only 20%-40% compared with daytime.

This marks the first appearance of a “time-based billing” mechanism in the domestic Agent sector — QoderWork has brought the familiar “peak-valley electricity prices” logic from the electric power industry into large model invocation, targeting the increasingly prominent challenge of “load-shifting utilization of compute power” in the Agent era.

1. Mechanism Explained: 22:00 to 08:00, Automatic 20%-40% Rate

According to Alibaba Cloud’s official announcement, the core rules for Peak-Valley Token are extremely simple — no need for manual switching by the user, no need to apply for quotas, no need to submit tickets. As long as the Agent task execution period falls between 22:00 each night and 08:00 the following morning, the system will automatically calculate point consumption at the nighttime discount rate.

Specific discount levels:

• Qwen3.7-Max: as low as 20% (i.e., 20% point consumption)
• Other main models: generally fall within the 20%-40% range
• Coverage: Applies to the entire QoderWork, QoderDesktop, and Qoder CLI product lines
• Trigger: Automatic recognition of task runtime period, no extra steps required

It’s worth noting that the definition of “nighttime” here is based on the actual execution period of the Agent, not the submission time. This means users can write instructions and debug Prompts during work hours in the daytime, and then schedule tasks or long-horizon tasks to push the heavy compute-intensive Token consumption to nighttime.

The official example use case is:

Set up timed tasks during the day, or submit long-horizon task instructions before bedtime — let the Agent autonomously execute the entire process at night, then review the results in the morning.

This “plan during the day, execute at night, review in the morning” workflow fits perfectly with the asynchronous, long-horizon, autonomous nature of Agent products.

2. Why “Peak-Valley Token”?

To understand the significance of this mechanism innovation, we need to go back to a basic question: What does the cost structure for large model inference really look like?

As everyone knows in the industry, behind large model inference services lies expensive GPU clusters. These clusters’ peak capacity is prepared for the daytime peak — between 09:00 and 22:00, developers, office workers, and students call them intensively, forcing cloud vendors to reserve ample GPU resources to guarantee SLA. After nightfall, call volume often drops sharply, leaving many GPUs semi-idle, while electricity, depreciation, and operations costs continue.

This “tight by day, loose by night” supply-demand mismatch is essentially identical to the “peak-valley load curve” in the power industry. The power industry addresses this with “peak-valley electricity prices” — using price signals to guide consumers to use electricity during off-peak periods. Now Alibaba Cloud has mapped this logic to the Token level:

• Daytime: GPUs in high demand, full-rate billing
• Nighttime: GPUs idle, 20%-40% discount
• Result: Cloud vendors improve cluster utilization, users get tangible savings

This is a win-win load-shifting. With Agent call volume rapidly growing exponentially, whoever can monetize idle compute first will gain a cost advantage in the price war.

3. Why Is “Nighttime Execution” Particularly Sensible in the Agent Era?

Traditional IDE programming assistants (e.g., code completion tools) have a distinct feature — the user must be online. Completion is a synchronous operation; without real-time human feedback, it’s meaningless.

But Agent tools are different. For desktop agents like QoderWork, the core usage scenarios have shifted from “second-level response” to “long-horizon tasks”:

• Have the Agent compile a week’s meeting notes into a weekly report
• Have the Agent conduct overnight research on a technical solution and provide a comparative analysis
• Have the Agent batch process hundreds of PDFs to extract structured data
• Have the Agent autonomously create, run, and debug a complete small project
• Have the Agent track multiple data dashboards and generate reports in the early morning

The common trait of these tasks is: the user doesn’t need to watch the screen in real time. After submitting a natural language instruction, the Agent will break down steps, call tools, operate local apps, and generate intermediate outputs by itself. Whether the task runs for 1 hour or 8 hours doesn’t change user experience much — as long as they get the result in the morning.

This is exactly the product logic foundation for “Peak-Valley Token”: Agent tasks inherently have a “time-insensitive” nature and can be shifted to nighttime execution without loss. This is a benefit that synchronous interaction modes like Chatbots or code completion will never enjoy.

4. Typical Usage Scenarios: Treat the Agent as Your “Night Shift Colleague”

Combining QoderWork’s capabilities, the “Peak-Valley Token” mechanism generates several imaginative usage patterns:

1. Timed Task Flows

During the day, design a complex multi-step task flow in the QoderWork workstation (such as “pull database → generate charts → draft analysis report → send email”), set the execution time to 23:00, and check the output directly the next morning at work.

2. Long-Horizon Coding Tasks

Use Qoder CLI to submit a large refactoring task, e.g., “migrate all JS files in the repo to TypeScript and add unit tests.” One sentence before bedtime; review the PR draft in the morning.

3. Batch Data Processing

Run heavy batch tasks like data cleaning, feature extraction, and vector database ingestion at night. With the same Token budget, you can process 3-5 times more data.

4. Digital Employees Working the Night Shift

The “QoderWake digital employee” in the QoderWork product family also enjoys nighttime discounts. A digital employee running 24/7 used to be costly — now the nighttime 16-hour window cost is only a fraction (20%-30%) of daytime, significantly reducing overall TCO (Total Cost of Ownership).

5. What Is QoderWork? A Product Line Review

For some users, the name QoderWork may still be unfamiliar. Here’s a brief review.

QoderWork is a desktop AI agent tool launched by Alibaba in January 2026 — users can use natural language instructions to let AI directly operate local applications and files on their computer to complete document generation, data analysis, file organization, and other tasks. It currently supports both Mac and Windows platforms.

As of now, the Qoder product family has expanded into a sizeable matrix:

• Qoder Desktop: desktop IDE form
• Qoder CLI: command-line Agent
• Qoder JetBrains plugin: IDE extension for the entire JetBrains suite
• QoderWork: desktop agent workstation
• QoderWake: digital employee
• Cloud Agents: cloud-based Agents
• Qoder mobile: mobile app

According to official disclosure, the entire product line has launched a Chinese edition, serving over 5 million users worldwide. In capability building, QoderWork has previously rolled out Design Desk, Writing Desk, and expert kits for industries like product management, legal, and marketing, as well as launched a “Consciousness” system featuring memory, reflection, and skill evolution to enhance Agent autonomy and efficiency.

This “Peak-Valley Token” looks more like a differentiated move on the “billing side” in the midst of white-hot competition on the “capability side” — it’s not just a price cut, but leveraging the asynchronous execution nature of Agents to tie discounts deeply to usage scenarios.

6. Industry Impact: Will Agent Billing Become Fully “Time-Based”?

Alibaba Cloud’s move may trigger a chain reaction across the entire Agent industry.

First, QoderWork claims this is the first Peak-Valley Token mechanism launched by an Agent product in China. Given that almost all mainstream players in the Agent field (domestic and overseas alike) are facing the challenge of high per-call costs, time-based billing is almost certain to be widely imitated.

Second, “Peak-Valley Token” will in turn shape user habits. Once users develop the muscle memory of “submitting long tasks at night,” Agent product forms will evolve toward “better suited for asynchronous, long-horizon, and automated orchestration” — perfectly aligned with the recent overseas trend emphasized by companies like Anthropic around “long-horizon agents.”

Third, such a mechanism also raises the bar for GPU cluster scheduling technologies. Cloud vendors need finer load forecasting and elastic scaling capabilities to accommodate “nighttime tasks” precisely within originally idle capacity — without causing new peak congestion when all users flock to the nighttime window.

Finally, from a business model perspective, “Peak-Valley Token” has an implicit meaning — it encourages users to migrate from “Chat mode” to “Agent mode”. Chat is instant dialogue and can’t enjoy nighttime discounts; Agent is asynchronous execution and can readily access the 20% rate. This pricing lever will steadily push users from “ask-answer” lightweight interactions toward “submit an instruction, Agent runs all night” heavy autonomous workflows.

7. Reminders and Tips

If you’re already a QoderWork user planning to try “Peak-Valley Token,” there are a few details to note:

1. Period determination is based on execution time: submission during the day is fine; what matters is whether the actual Token consumption period falls within 22:00-08:00.
2. How cross-period tasks are billed: the official statement hasn’t clarified in detail how tasks spanning the “peak-valley divide” are billed by segment — for complex tasks, it’s advisable to schedule them fully within the off-peak window.
3. Qwen3.7-Max benefits the most: 20% is the highest disclosed discount; other models fall between 20%-40%, so choose the most economical combination based on task complexity.
4. Make good use of timed tasks: splitting daytime “immediate needs” into “planning + asynchronous execution” stages is key to maximizing nighttime discounts.
5. Watch for evolution of the “Consciousness” system: long-horizon tasks demand greater Agent capabilities like memory, reflection, and error recovery — QoderWork’s proprietary “Consciousness” system is the critical safeguard for stable overnight long-horizon task completion.

8. In Closing

After the “parameter races” and “inference optimization contests” of 2023-2025, entering 2026, the competitive focus of the large model industry is quickly shifting toward application and business model layers. Agents are the most important application form right now, and billing innovations around Agents are destined to be the next year’s new battleground among cloud vendors.

Alibaba Cloud QoderWork’s “Peak-Valley Token” mechanism may look like just a “nighttime discount promotion,” but in essence it cleverly stitches together three things: the asynchronous execution nature of Agent products, the engineering reality of GPU cluster load-shifting, and users’ urgent need to reduce costs. It’s not just a simple price cut — it’s using price signals to reshape usage paradigms in the Agent era.

Whether the 20% rate will become a new industry benchmark, and whether Tencent Cloud, Volcano Engine, Baidu AI Cloud, and others will follow suit, is something we can watch and see.

---

References

• Alibaba Cloud: QoderWork Launches “Peak-Valley Token,” Off-Peak Use of Qwen 3.7 as Low as 20% - IT Home: IT Home’s first report on QoderWork’s Peak-Valley Token mechanism launch, including key points from the official announcement.
• Smart Era Channel - IT Home: IT Home’s Smart Era channel, continually tracking the latest developments in domestic AI tools and Agent products.