Doubao Work Adds Third-Party Models with a New Usage Multiplier System

    Doubao Work now supports third-party models including DeepSeek V4.1 Flash, Kimi K3.1 Pro, and GLM 5.3, allowing users to select models by task. The update also introduces model-specific usage multipliers, making credit consumption more transparent and helping teams balance quality, speed, and cost across AI-powered workflows.

    From a Single Model to a Multi-Model Ecosystem

    Over the past year, large language models have become deeply integrated into workplace applications. Tasks such as document summarization, email drafting, data analysis, and code assistance have moved from experimentation into everyday production use. At the same time, one limitation has become increasingly clear: no single model performs best across every task.

    Some models excel at long-context understanding and rigorous reasoning. Others are stronger at code generation, while lightweight models may offer faster responses and lower costs. In this environment, keeping users locked into a single in-house model can limit the potential of AI-powered productivity tools.

    Doubao Work’s latest update directly addresses this issue. While retaining its proprietary models, the platform now offers additional third-party options, including DeepSeek V4.1 Flash, Kimi K3.1 Pro, and GLM 5.3. Users can switch between first-party and third-party models based on the task at hand.

    For example, a team can use a model with stronger coding capabilities for software development tasks and a model with more reliable long-context understanding for meeting summaries—all within the same workflow. They no longer need to use one model for every type of work.

    The Real Challenge of Multi-Model Access: Different Cost Profiles

    Once multiple models are available, another issue becomes unavoidable: different models consume different amounts of resources.

    Two models may complete the same task while requiring different levels of compute. If a platform continues to deduct the same amount of credits for every model, users may be charged equally for lightweight and resource-intensive processing. This approach is neither transparent nor economically rational, and it makes it harder for users to choose models based on cost.

    Doubao Work’s new usage multiplier system is designed to address this problem. A usage multiplier can be understood as a pricing coefficient applied to different models when they perform comparable tasks. A higher multiplier means that more account credits are consumed for the same level of usage, while a lower multiplier means fewer credits are deducted.

    The multiplier does not change the complexity of the task itself. Instead, it determines how quickly the user’s available credits are consumed.

    Its primary value is that it makes cost information available before a task is executed. Previously, users often had to rely on experience or guesswork when deciding which model would be more economical. They might only discover that their usage had become expensive after reviewing their monthly balance or bill. With multipliers displayed in the model selection interface, users can see each model’s cost characteristics in advance and make a deliberate trade-off between quality, speed, and cost.

    This type of pricing design is not entirely new, but its implementation in workplace AI has specific implications. Office tasks are typically small and frequent: generating meeting minutes, drafting batches of email replies, or analyzing spreadsheet data. The credit cost of an individual task may be minimal, but the total can become significant when multiplied by the number of employees and monthly usage frequency.

    Usage multipliers make the unit cost more predictable and give organizations a more precise basis for planning their AI budgets.

    From Team Plans to Broader Availability

    Doubao Work did not open multi-model access to all users at once. In late October, the platform first made third-party model configuration available to customers on team and enterprise subscriptions.

    During this initial phase, organizations did not need to purchase additional access from model providers or apply for separate permissions. Usage was deducted directly from their existing AI credits. In other words, enterprise customers could use more models within their existing subscription framework without going through an additional procurement process.

    The update in mid-November expanded the capability to a wider user base. Individual users and smaller teams could also see third-party model options in the model selection list and use the same credit deduction logic as team subscribers.

    This gradual rollout served two purposes. It prioritized enterprise requirements while giving the platform time to validate the stability of multi-model routing and the accuracy of its usage accounting before expanding access more broadly.

    From a product strategy perspective, a “enterprise first, then broader release” approach is reasonable. Enterprise customers typically have higher requirements for reliability, security, and cost control. Serving this group first helps the platform strengthen its underlying mechanisms. Once those mechanisms are mature, wider availability can be introduced with lower operational risk.

    Limited-Time Incentives to Encourage Adoption

    Any new selection and pricing mechanism faces a cold-start challenge: users need both time and motivation to adjust their behavior. Doubao Work therefore introduced limited-time incentives intended to reduce the cost of trying the newly available models.

    Time-based discounts. During designated periods on weekdays and holidays, participating models are available at a discounted credit consumption rate. This structure aligns with common workplace usage patterns, since many AI assistant tasks are performed during working hours. Users can try the new models while completing routine work, with the discount reducing the initial cost of experimentation.

    Lower multipliers for selected models. The update also lowered the usage multipliers for Doubao 2.1 Lite and DeepSeek V4.1 Flash. For comparable tasks, users consume fewer credits when using these models.

    Both models are positioned toward frequent, lightweight workplace tasks such as quick question answering, document summarization, and information extraction. These tasks generally have a low cost per request but a high frequency of use. As a result, a lower multiplier can produce substantial savings for users who make repeated calls over time.

    Together, the two incentives create a clear behavioral effect. Time-based discounts reduce the marginal cost of trying a new model, while lower multipliers make lightweight models more economical for frequent tasks. The overall goal is to encourage users to develop a “choose the model for the task” workflow.

    The Technical Foundation: Routing and Usage Accounting

    From an engineering perspective, supporting multiple models involves more than adding several API endpoints. At least two core capabilities are required.

    First is model routing and orchestration. The platform needs a unified request entry point that can forward requests to different model services based on the user’s selection. This layer must account for differences in response time, service availability, rate limits, and fallback or degradation strategies.

    For example, DeepSeek V4.1 Flash is positioned as a lightweight model and may typically respond more quickly. However, during periods of high concurrency, the platform still needs to prevent excessive load on one model from affecting the overall user experience.

    Second is usage measurement and accounting. This is the central component of the new multiplier system. After each request is completed, the platform must accurately record the number of tokens or other usage units consumed, apply the multiplier associated with the selected model, and deduct the corresponding amount from the user’s account.

    This requires precise metering and near-real-time credit updates. For enterprise users, the platform also needs to provide detailed usage records to support financial reconciliation and internal cost allocation.

    These capabilities may appear straightforward, but implementing them reliably is difficult. With multiple models running in parallel and the user base continuing to grow, the platform must handle billing accuracy, transparent statements, and fault tolerance for abnormal or failed requests. Each area requires careful system design and substantial engineering effort.

    Practical Recommendations for Developers and Businesses

    For developers and organizations using, or considering, a multi-model workplace platform, several practices are worth adopting.

    1. Choose models by task rather than by default preference.
    The value of a multi-model platform is not simply the number of available models. It comes from assigning each task to the model best suited to it.

    Organizations can map common workflows to recommended models. For example, lightweight models may be sufficient for routine internal communication summaries, while complex contract reviews or technical proposals may benefit from models with stronger reasoning capabilities.

    2. Monitor credit consumption and configure alerts.
    Cost visibility becomes more important when users can freely switch between models. Organizations should ensure that teams can monitor credit usage in real time and set appropriate alert thresholds.

    This is especially important when individual users begin making frequent calls to resource-intensive models. Without monitoring, that behavior could result in unexpected credit consumption toward the end of a billing period.

    3. Use limited-time incentives as a testing window.
    For organizations that have not yet established clear model-selection standards, promotional periods can provide a relatively low-cost opportunity to experiment.

    Different roles—such as administrative staff, developers, and marketing teams—can test different models against their typical workloads. The resulting feedback can then inform a longer-term model allocation strategy.

    4. Maintain security and compliance controls.
    Using multiple models may send data to a broader set of third-party services. Organizations should review each provider’s data handling policies carefully.

    For workflows involving confidential or sensitive information, model-selection rules should be applied more cautiously. Access to additional models should not replace existing data governance and security requirements.

    Conclusion

    Doubao Work’s transition from a single proprietary model to a broader ecosystem of first-party and third-party models, together with its move from uniform credit deductions to model-specific usage multipliers, reflects a deeper transformation in workplace AI products.

    The industry is shifting from simply providing model capabilities to managing the economics of model usage.

    As model capabilities become more comparable across common tasks, platforms that help users select the right model at the right time and at a reasonable cost may gain a meaningful advantage. A usage multiplier may appear to be a billing detail, but it represents a clear product philosophy: make every unit of AI capability count where it matters most.