Doubao Work Adds Third-Party Models with Usage Multipliers and Limited-Time Discounts
Doubao Work has expanded beyond its two in-house models by adding Kimi K3.1 Pro, GLM 5.3, and DeepSeek V4.1 Flash. The update also introduces usage multipliers and time-limited discounts, giving users clearer cost signals, more control over model selection, and new ways to optimize workloads across different task types.
From a Single In-House Model Strategy to Multi-Model Routing
The capability ceiling of AI office assistants is shifting from improving a single model to coordinating multiple models more effectively. On November 12, Doubao Work announced a major product update. In addition to its two existing in-house models, the platform added three categories of third-party models and introduced a dynamic pricing mechanism based on usage multipliers.
This is more than a simple expansion of the model list. It represents a structural change in how an office productivity platform manages and routes models.
1. Expanding the Model Matrix: From a Closed System to Open Routing
Previously, Doubao Work offered only two in-house models, positioned for lightweight responses and advanced reasoning respectively. While this closed model strategy helped maintain a consistent user experience, it also created a clear limitation: different tasks require very different capabilities.
Simple summarization and data extraction are fundamentally different from complex multi-turn reasoning and long-document analysis. These use cases are better served by models with different architectures and parameter scales.
Following the update, Kimi K3.1 Pro, GLM 5.3, and DeepSeek V4.1 Flash became available as the first group of third-party models. Each model has a distinct focus:
- DeepSeek V4.1 Flash: Designed for frequent, lightweight tasks, with fast response times and suitability for high-volume text processing.
- Kimi K3.1 Pro: Stronger long-context understanding for deep reading and complex instruction decomposition.
- GLM 5.3: Differentiated by general semantic understanding and coherent multi-turn conversations.
Together, these models create a three-level spectrum covering lightweight, medium-complexity, and demanding workloads. Users can switch the underlying model from the same task interface according to task complexity, without needing to manage the integration details themselves.
In effect, the design moves model selection closer to the user, rather than leaving all routing decisions to the platform’s backend.
Before the general rollout, Doubao Work had already made third-party model configuration available to team and enterprise subscription customers on October 29. These customers did not need to purchase separate services from model providers or apply for individual API access. Usage was deducted directly from their existing AI credits.
This early-access approach lowered the cost of enterprise experimentation while allowing the platform to collect real-world workload and stability data before the broader release.
2. Usage Multipliers Make Cost Differences Visible and Predictable
Once multiple models are available, an important question arises: how should the platform account for the fact that different models require different levels of computing resources?
Doubao Work addresses this by displaying a usage multiplier for each model on the model-selection page. A lower multiplier means that completing the same task consumes fewer credits.
The key value of this design is that it places cost information directly in the decision-making interface. Users no longer need to guess which model is more economical based on experience or intuition. Instead, they can make a quantitative trade-off using the visible multiplier.
From a technical perspective, a usage multiplier is essentially a pricing weight. It does not change the complexity of the task itself. Instead, it reflects the difference in computational cost when different models process comparable workloads.
This type of mechanism is common in cloud services. However, exposing the underlying cost logic so directly in a productivity assistant designed for non-technical users is relatively uncommon.
The mechanism also creates several broader effects:
- Improved budget visibility: Team managers can use multiplier differences to estimate the monthly credit consumption of different workflows in advance.
- Better task-to-model matching: Users are encouraged to consider which model is appropriate for each task instead of sending every request to the most powerful model, helping conserve computational resources.
- More constructive competition among providers: Differences in usage multipliers may encourage model providers to reduce inference costs rather than relying solely on larger parameter counts.
3. Limited-Time Discounts Target High-Frequency Workloads
The update introduces limited-time discounts in two ways.
First, time-based discounts. During designated periods on working days and holidays, participating models can be used at a discounted credit consumption rate. This peak-and-off-peak pricing approach can help shift batch-oriented, non-real-time workloads away from busy periods.
For example, large-scale text-cleaning tasks could be scheduled during discounted periods to reduce costs without disrupting business operations.
Second, lower multipliers for specific models. The usage multipliers for Doubao 2.1 Lite and DeepSeek V4.1 Flash have been reduced, allowing users to consume fewer credits for the same task.
Both models are positioned toward frequent, lightweight office workloads, such as meeting-note organization, email classification, and information extraction. These tasks typically have low consumption per request but high usage frequency. As a result, a lower multiplier can produce meaningful savings over time.
The strategic intent behind the discounts is clear: reduce the cost of lightweight models enough to encourage long-term, high-frequency usage while using time-based discounts to smooth the overall load curve and reduce resource pressure during peak periods.
4. Practical Guidance for Building a Model Routing Strategy
When working with multiple models and usage multipliers, teams can consider the following configuration principles:
- Prioritize lightweight models by default: Use DeepSeek V4.1 Flash or Doubao 2.1 Lite as the default model for routine tasks such as summarization, extraction, and classification.
- Route by task complexity: At the API-call layer, establish simple rules. For example, automatically switch to Kimi K3.1 Pro or GLM 5.3 when the input exceeds a certain length or requires multi-step reasoning.
- Run batch jobs during discounted periods: Schedule tasks that do not require real-time results, such as report generation and historical data cleaning, during discount windows.
- Review multiplier estimates against actual consumption: A multiplier is an expected pricing coefficient. Actual usage may fluctuate based on input and output length, concurrency, and other factors. Teams should use logs or analytics tools to compare estimated and actual consumption each month.
5. Industry Perspective: Model Diversity Is Reshaping Workflow Design
As integrating more models becomes standard practice, the real competitive advantage will no longer be the number of models a platform can access. Instead, it will be the quality of its framework for routing, pricing, and experience management.
Usage multipliers are only the most visible part of this framework on the billing side. Behind them are deeper infrastructure concerns, including latency allocation, concurrency priorities, and context-caching strategies.
For enterprise users, the positive development is that model-selection authority is moving closer to business teams. In the past, the choice of model was often made by IT departments or architects. Today, everyday users can see cost differences in the interface and make trade-offs within their normal workflows.
This broader participation in model decisions could become one of the most important trends in office software over the next two years.
As more models are added and pricing systems become more granular, Doubao Work may evolve from a simple productivity entry point into a model-routing hub. If that happens, the design principles behind its usage multiplier system could become a useful reference for the wider industry.
