Office AI Embraces Third-Party Models with New Usage-Based Pricing
This article examines the latest update to a major office AI workbench, which now supports third-party large language models alongside its proprietary engines and introduces a usage multiplier–based pricing mechanism. It breaks down the strategic rationale behind multi-model integration, explains how the new cost-calculation model works, and offers practical guidance for enterprises looking to optimize AI spending without sacrificing output quality.
As AI applications accelerate across enterprise environments, the office workplace has become one of the most direct proving grounds for large language model (LLM) value delivery. Recently, a leading intelligent office workbench completed a major product iteration: it moved from supporting only proprietary models to integrating a third-party LLM ecosystem, while simultaneously rolling out a consumption-based billing mechanism. This shift not only gives users greater freedom in model selection, but also signals that AI office tools are entering a new phase—moving beyond feature competition into精细化 cost optimization.
From Closed to Open: The Strategic Significance of Multi-Model Access
Historically, AI capabilities within office platforms were tightly coupled to the platform’s proprietary models, leaving users with limited options when balancing task complexity against performance expectations. The core breakthrough of this update lies in adding third-party models—such as Kimi K3.1 Pro, GLM 5.3, and DeepSeek V4.1 Flash—to the optional model list, running in parallel with the existing proprietary models. This combined approach reflects multiple technical considerations: different models excel in different capability dimensions, including semantic understanding, code generation, and long-document processing. Open model selection allows users to match the most suitable model to specific task characteristics, rather than passively accepting the inherent capability boundaries of a single engine.
The introduction of a multi-model architecture was not an overnight development. Prior to the full public rollout, the workbench had already piloted third-party model configuration for team subscription and enterprise subscription customers. In this pilot phase, enterprises were not required to make separate purchases or applications to model vendors; usage was deducted directly from existing AI credits. This integration model significantly reduced the technical barriers and procurement complexity associated with multi-model access. The subsequent full-scale rollout extended this capability to a broader user base.
From a technical architecture perspective, integrating third-party models requires solving a series of engineering challenges: model gateway design, request routing, context management, and security isolation. The model selection list visible to users may appear simple, but underneath it relies on a mature adapter layer and unified API abstraction. When users switch between models, the experience should feel seamless—behind the scenes, non-functional attributes such as model output quality evaluation, timeout control, and error retry mechanisms all require rigorous testing and tuning.
Usage Multiplier: Making Costs Quantifiable
The direct consequence of multi-model coexistence is a more complex cost structure. Different models handling the same task exhibit significant variance in parameter scale, inference compute consumption, and output token counts. If a uniform billing standard were applied, it would inevitably lead to a cost-price inversion, and would indirectly push users toward low-compute models, thereby intensifying infrastructure pressure.
The introduction of the usage multiplier is designed precisely to establish a transparent, quantifiable adjustment mechanism between user experience and resource allocation. The core logic is straightforward: on the model selection page, each model displays a specific multiplier value. This multiplier can be understood as the relative weight of account credits consumed by that model when completing a standardized task. The lower the multiplier, the fewer credits deducted per task, and vice versa.
The key point here is that the usage multiplier is not a judgment of model capability—it is a quantitative expression of task execution cost. It does not change the complexity of the task itself, nor does it affect task parameters such as prompt length or output length. However, it directly determines how quickly the available balance in a user’s account is consumed. For example, in high-frequency interaction scenarios (such as text polishing, summarization, or table parsing), if a user selects a large model with a high multiplier, the account balance will drain significantly faster than if a lightweight model were used, especially as the number of calls accumulates.
From a product design psychology perspective, this front-end cost visualization carries deep significance. In traditional models, API call costs are often hidden in backend billing statements, making it difficult for users to develop an intuitive sense of the cost per operation, which can lead to budget overruns without awareness. By displaying the multiplier directly next to each model button, users develop a clear cost awareness before clicking—establishing a direct trade-off between performance expectations and budget planning. In essence, this refines the granularity of “pay-as-you-go” pricing—already familiar in cloud computing—down to the level of individual model operations.
Two Paths for Limited-Time Promotional Strategies
Running parallel to the billing mechanism update is a limited-time promotional plan targeting specific time windows and specific models. The plan operates along two dimensions: time-based discounts and model-based multiplier reductions.
On the time dimension, during designated periods on workdays and holidays, using participating models qualifies for discounted credit consumption. This type of elastic pricing strategy is well-established in cloud computing. Its essence lies in using price signals to shift user call patterns across time, smoothing workload peaks and improving overall infrastructure utilization. For users, scheduling non-urgent batch tasks (such as document classification or meeting note summarization) during discount windows can significantly reduce operational costs, while the actual model response capability remains unchanged.
On the model dimension, two complementary models have had their usage multipliers directly reduced. One is a proprietary lightweight model, and the other is a third-party high-efficiency model. Both are positioned for high-frequency, lightweight office scenarios such as quick Q&A, information extraction, and format conversion. These task types have much higher call frequencies per unit time than long-context deep analysis tasks, which means the effect of multiplier reduction is continuously amplified with cumulative usage. The credits saved per individual call may seem trivial, but when aggregated at monthly or quarterly scale, the cost optimization effect becomes substantial.
Model Selection Logic and Practical Recommendations for Enterprises
Given the array of model options and differentiated billing rules, enterprise administrators need a clear selection framework when configuring AI tools for their teams. The first step is task profiling: identify the most frequent AI call scenarios within the team, and determine the distribution ratio between high-concurrency, low-latency scenarios and high-quality, deep-reasoning scenarios. The second step is defining a model combination strategy: rather than pursuing a single model that performs optimally in all scenarios, choose a combination based on task characteristics—for example, using lightweight models for information preprocessing and reserving high-multiplier models for complex reasoning and creative generation tasks.
Enterprises should also take full advantage of usage traceability to analyze the actual call distribution across models and continuously tune the default model settings. By correlating credit consumption data with task output quality, teams can gradually build a model cost-performance evaluation framework specific to their needs, steering AI investment toward high-return scenarios.
From a broader perspective, the trend represented by this update is clear: capability competition in office AI is shifting from single-model performance contests to a compound competition of “capability spectrum + precision pricing.” In the future, the platform that delivers optimal compute cost-performance at any task scale will be the one that builds lasting user stickiness in the deep end of intelligent office adoption. The rollout of the usage multiplier mechanism is a significant marker of this new round of competition.
