Weekly Generative AI Trends: Model Performance, Cost, and the Rise of AI Assistants

    This weekly review covers how generative AI competition is shifting from answering questions to completing real tasks. It examines a new large Chinese model, low-cost lightweight reasoning models built for AI agents, the move toward always-on personal assistants, and usage data from a public model routing platform. It also outlines how teams can evaluate models and design AI systems that are reliable and cost-controlled.

    1. Competition Shifts from Answering Questions to Completing Tasks

    Recent changes in the large model market go beyond individual models posting better benchmark scores. The more important change is that the criteria for competition are shifting.

    Early comparisons focused on static abilities such as knowledge Q&A, mathematical reasoning, code generation, and text comprehension. As models were connected to office systems, development workflows, and business platforms, users began to focus on questions closer to production environments:

    • Can the model follow context changes across long-running tasks?
    • Can it keep calling tools and handle errors?
    • Can it pass information between multiple software applications?
    • Can it reduce the total call cost of complex tasks?
    • Can it meet enterprise requirements for data security, stability, and auditability?

    As a result, the quality of a single response no longer fully reflects a model’s value. For enterprises, a more meaningful metric is the “task success rate per unit cost.” In other words, can a model complete a goal within an acceptable latency and budget in a real workflow?

    This week’s industry activity suggests that model capability, inference cost, open release policies, and the productization of AI agents are forming a feedback loop that affects one another.

    2. Chinese Models Continue to Move Into the High-Performance Tier

    A Chinese model recently released a new post-trained version with a parameter scale in the trillion range. It focuses on improving complex scenarios, including long-horizon coding, 3D modeling assistance, and financial research. Public information shows the model scored 55 points on a comprehensive intelligence benchmark, about 11 points higher than its previous generation, and it outperformed some other high-performing models released in the same period.

    Note that a post-trained version does not necessarily mean the base model’s parameters changed by the same amount. Post-training usually includes supervised fine-tuning, preference optimization, reinforcement learning, tool-use training, and inference strategy optimization for specific tasks. Its value lies in the following areas:

    1. Better execution stability on complex tasks
      The model does more than generate a piece of code. It needs to break down requirements, plan steps, verify execution, and correct its work based on error feedback. Post-training can improve behavioral consistency across multi-turn tasks.

    2. Stronger performance in specialized domains
      Tasks such as financial research, engineering design, and 3D modeling often require a model to process structured data, professional documents, and tool outputs at the same time. Simply expanding the pretraining corpus does not fully solve this. The model also needs training tailored to specific workflows.

    3. Better use of information in long contexts
      A long context window does not guarantee effective understanding. What matters is whether the model can find key constraints in a large volume of earlier information, and avoid missing dependencies or repeating actions in long tasks.

    However, higher benchmark scores do not mean a model has an absolute advantage in every production environment. When selecting a model, enterprises should also check the following factors:

    • Combined input and output token pricing
    • Actual latency in long-context scenarios
    • Concurrency limits and stability during peak demand
    • Support for structured output and tool calling
    • Whether generated results are easy to audit and trace
    • How the model handles sensitive data, permission boundaries, and unusual inputs

    Public pricing shows that inputs and outputs for this class of high-capability model cost about 13 yuan and 68 yuan (RMB) per million tokens, respectively. Compared with overseas models in a similar tier, the cost per task may be about 1.5 to 2 times higher. Capability gains and cost control clearly remain in tension.

    Development teams should not make decisions based on a single intelligence score. A more practical approach is to build a task-level evaluation set. For example, teams can score tasks such as code repair, data analysis, report generation, and browser operations on success rate, average number of calls, and total cost.

    3. Low-Cost Inference Brings AI Agents into Practical Use

    Another notable change is that a lightweight reasoning model designed for agent scenarios has further reduced usage costs and opened its model weights. Public information shows the model’s input and output prices are about $0.15 and $0.60 per million tokens. Its cost per agent task has dropped about 70% compared with the previous generation, and its online inference speed is nearly three times faster.

    The value of this type of model is not that every single response reaches the highest quality. Its value is that it suits high-frequency calls. An agent task may go through the following process:

    1. Understand the user’s goal
    2. Read the existing context
    3. Select a tool
    4. Call the tool and parse the result
    5. Decide whether a retry is needed
    6. Adjust parameters and run again
    7. Summarize the results and generate the final response

    If every step uses an expensive model, overall costs rise quickly. For this reason, enterprises often adopt a tiered architecture:

    • Use high-capability models for task planning, complex judgment, and final review
    • Use low-cost models for classification, information extraction, simple rewriting, and tool parameter generation
    • Replace stable, repetitive processes with rules or dedicated programs instead of model calls
    • Cache historical results to reduce repeated inference

    Cross-application office work and computer operation mean that a model completes actions through a browser, a desktop environment, or business interfaces. The main technical challenges of this capability include the following:

    1. Uncertainty in the Operating Environment

    Page layouts may change, button labels may differ, and network requests may fail. An agent needs visual understanding, page element localization, and error recovery. It cannot simply click fixed coordinates.

    2. Permission and Security Controls

    Actions such as sending emails, changing financial data, or submitting approvals should not depend entirely on the model’s own decisions. Systems should enforce permission allowlists, require secondary confirmation, block sensitive actions, and log every operation.

    3. Task State Management

    Long-running tasks need to record the current step, completed actions, pending items, and reasons for failure. It is recommended to save task state in structured storage instead of relying only on the conversation context.

    4. Recoverability After Failure

    An agent should not end a task after a single failed call. Engineering systems need to distinguish between retryable errors, parameter errors, permission errors, and business rejections, and apply a different handling strategy to each type.

    A safer agent workflow can be summarized as follows:

    1. Receive goal
    2. ↓
    3. Break down the task and identify risks
    4. ↓
    5. Select model and tools
    6. ↓
    7. Execute a single step
    8. ↓
    9. Verify the result
    10. ├─ Success: proceed to the next step
    11. ├─ Retryable: adjust parameters and retry
    12. └─ High risk: pause and request human confirmation
    13. ↓
    14. Generate results and audit records

    4. Personal Assistants Become a New Competitive Focus for Leading Vendors

    After several large AI service providers launched similar features, a leading model vendor also began testing a personal assistant product that runs continuously. Unlike instant chat, this type of product attempts to handle information such as email, calendars, and financial analysis over the long term. It can keep executing tasks while the user is offline.

    This shows that personal assistants are evolving from a “chat window” into a continuously running task agent. Key capabilities include:

    • Reading and organizing information from multiple sources
    • Recognizing the user’s long-term preferences
    • Scheduling tasks based on time and priority
    • Automatically generating summaries, reminders, and to-do lists
    • Escalating unusual situations
    • Keeping a searchable history of operations

    However, deploying a personal assistant is much harder than deploying a standard Q&A system. It has to manage ongoing state and highly sensitive data. Email, calendars, bills, and financial information are all high-value data. A single incorrect operation can have consequences far greater than a wrong answer.

    For this reason, enterprises designing similar systems should prioritize a “suggest first, limited automation” approach. For example:

    • Automatically organize emails, but require confirmation before sending a reply
    • Automatically suggest calendar arrangements, but let the user approve the final schedule
    • Automatically analyze bills, but do not perform fund transactions directly
    • Automatically generate financial summaries, but clearly label data sources and calculation methods

    In addition, the system should provide memory management features. Users should be able to view, edit, and delete long-term memories. A personal assistant becomes genuinely usable only when users can understand what the system has remembered, why it performed a given action, and how to undo its results.

    5. Public Usage Data Shows Developer Demand Is Rising Again

    Recent open data from a public model routing platform shows that total token usage over the past seven days was about 184.01 trillion, up 9.3% from the previous period. Text requests totaled about 7.018 billion, up 4.7%. This suggests that demand for generative AI from developers and enterprise users is recovering.

    By model, usage of a Chinese model’s old and new versions combined reached about 49.55 trillion tokens, up about 35.5%. A newly launched lightweight reasoning model received about 14.13 trillion tokens in calls. A high-performance general-purpose model received about 8.26 trillion tokens, up about 36.5%. Another preview model saw its usage fall about 79.6%.

    These data points reflect several trends:

    1. New Models Typically Receive Fast Trial Traffic

    After a new model launches, developers run compatibility tests, performance tests, and migration assessments. Short-term growth in calls does not necessarily mean long-term retention. To judge whether a model is truly adopted, teams need to track repeat call rates and the share of production traffic over several weeks.

    2. Pricing Is a Key Prerequisite for Scaling Agents

    Ordinary conversations involve a limited number of calls, but agents generate many rounds of inference and tool calls. Only after prices fall are developers willing to place models into automated workflows, which in turn drives much higher token consumption.

    3. Switching Costs Between Models Are Falling

    As unified calling interfaces, model routing, and standardized message formats become more common, applications can connect to several model types at once. Developers can choose models dynamically based on task difficulty, latency, and cost. This will further intensify market competition.

    6. How Enterprises Can Build a More Reliable Model Evaluation System

    Because models are upgraded frequently, enterprises should not rely only on general leaderboards published by vendors. They should build an evaluation system that fits their own business.

    At a minimum, testing should cover five dimensions:

    1. Task success rate: Does the model actually complete the goal, rather than producing text that only looks plausible?
    2. Factual accuracy: Are key conclusions sourced, and does the output contain claims that cannot be verified?
    3. Call cost: This includes input, output, retries, tool calls, and manual review costs.
    4. Response stability: Measure latency and error rates across different times and concurrency levels.
    5. Security and compliance: Assess data isolation, permission controls, log auditing, and the handling of sensitive information.

    In engineering terms, teams can wrap model calls in a unified service and add the following components around it:

    • Request routing
    • Token usage tracking
    • Result caching
    • Timeouts and retries
    • Sensitive content detection
    • Quality evaluation
    • Monitoring and alerting
    • Cost analysis

    For the underlying infrastructure, object storage can hold evaluation data, message queues can handle asynchronous tasks, container platforms can scale services elastically, and logging and monitoring systems can track call quality. For scenarios that require domestic deployment or enterprise-grade security, teams can also evaluate the general AI and cloud infrastructure capabilities offered by Baidu AI Cloud. The actual architecture should still be designed independently, based on data sensitivity levels, concurrency requirements, and business workflows.

    7. Conclusion: From a Model Race to a Task System Race

    This week’s trends show that the large model industry has entered a stage that puts more emphasis on engineering deployment.

    On one hand, high-capability models continue to improve at complex coding, specialized research, and long-running tasks. Some Chinese models are approaching or exceeding established high-performance models on certain comprehensive benchmarks. On the other hand, low-cost reasoning models reduce the cost per call and improve response speed, which makes large-scale agent use practical.

    At the same time, personal assistants are shifting from experimental features to continuously running task systems. The future competition will focus less on which model has the higher benchmark score and more on the following:

    • Who can complete a full workflow at a lower cost
    • Who can handle errors and long-running tasks more reliably
    • Who can balance automation with human confirmation
    • Who can provide clear permission, audit, and data governance capabilities

    For developers, the priority in the next stage is not to chase every new model. It is to build a model application architecture that can be swapped, evaluated, and monitored. Generative AI becomes a sustainable productivity tool only when model capability is combined with business workflows, tool systems, security mechanisms, and cost management.