0.3% 幻觉率登 Nature Medicine,医疗 agent 应用生态的"可信"拐点到了?

Researchers presented 100 real patient questions to six AI systems, including frontier large language models and specialized clinical AI tools. Twelve clinicians then evaluated the responses through a double-blind review process.

A 0.3% Hallucination Rate in Nature Medicine: Has Medical AI Reached Its Trustworthiness Tipping Point?

Bilingual Article|English + Chinese
The complete English version appears first.
中文完整版位于文章后半部分,请继续向下阅读。

In July 2026, a study published in Nature Medicine introduced an intriguing number into the medical AI debate.

Researchers presented 100 real patient questions to six AI systems, including frontier large language models and specialized clinical AI tools. Twelve clinicians then evaluated the responses through a double-blind review process.

The result appeared encouraging: all six systems produced relatively low rates of hallucinated medical information.

But another set of findings deserves even more attention.

A separate systematic study examining hallucinations across real-world clinical data reported rates ranging from 0.3% to 76.3%, depending on the type of medical text, prompting strategy, and model used to classify the outputs.

The same metric—“hallucination rate”—varied by more than 250 times.

A 0.3% rate sounds like evidence that medical AI may finally be crossing a trustworthiness threshold.

But then what should we make of 76.3%?

The answer is that a hallucination rate is never just a number.

Its meaning depends on the system, the task, the conversation history, the evaluation method, and the clinical environment in which the model operates.


The Context Behind 0.3% Matters More Than the Number

The 0.3% result should not be interpreted as a universal hallucination rate for medical AI.

It was measured under a specific combination of conditions:

  • A particular type of clinical text;

  • A defined prompting strategy;

  • A selected model configuration;

  • A specific classification method;

  • A controlled evaluation process.

In other words, 0.3% represents performance under favorable and clearly bounded conditions.

That is still meaningful.

It demonstrates that medical AI can produce highly factual results when the task, context, evidence sources, and evaluation criteria are carefully controlled.

But it does not mean every medical chatbot, diagnostic agent, or clinical workflow has reached the same level of reliability.

A model answering one structured question under controlled conditions is very different from an agent that must follow a patient through a long consultation, interpret multiple test results, retrieve evidence, update records, and make recommendations over many conversational turns.

The number should therefore be treated as a signal of possibility—not a blanket guarantee of clinical safety.


A Broader Evaluation Reveals a More Complicated Reality

Another Nature Medicine study offered a more comprehensive view.

Researchers developed the Medieval evaluation framework and invited 27 experienced clinicians—with an average of 25.9 years in practice—to assess seven frontier large language models.

The evaluation included:

  • 685 real and simulated clinical cases;

  • 4,795 generated responses;

  • Multiple forms of medical reasoning and communication.

The findings were far less uniform.

Between 3% and 19% of model responses were rated unacceptable, and evaluators identified 40 hallucinated outputs.

The best-performing models approached the level of an early-career physician in some tasks.

That is an impressive achievement.

It is also far from complete clinical reliability.

A 3% error rate may appear small in a conventional software product. In healthcare, the consequences depend entirely on where the error occurs.

Three errors in 100 routine administrative summaries may be manageable with human review.

Three errors in 100 medication recommendations, radiology interpretations, or high-risk triage decisions may be unacceptable.

This is why medical AI cannot be judged through one aggregate score.

Clinical risk is unevenly distributed.

The same numerical error rate can have completely different consequences across different tasks.


Multi-Turn Conversations Can Magnify Hallucinations

An even more concerning signal came from research published in npj Digital Medicine.

A clinical retrieval-augmented generation system recorded a hallucination rate of approximately 5% when no conversation history was included.

After only ten rounds of prior dialogue, the hallucination rate reportedly increased to around 40%.

The conversation history grew from zero to ten turns.

The hallucination rate increased eightfold.

This matters because healthcare interactions are rarely single-turn exchanges.

A real medical workflow may include:

  • Initial symptoms;

  • Previous diagnoses;

  • Medication history;

  • Laboratory results;

  • Imaging findings;

  • Follow-up questions;

  • Changes in patient condition;

  • Clarifications from doctors;

  • Updated treatment decisions.

The agent must maintain a consistent understanding over time.

Every new round introduces another opportunity for:

  • Earlier errors to propagate;

  • Irrelevant details to contaminate the context;

  • Conflicting information to be merged incorrectly;

  • The model to over-weight an earlier assumption;

  • Evidence to become disconnected from the final conclusion.

A system that appears trustworthy in a one-question benchmark may become significantly less reliable in a real clinical conversation.

That is one of the most important lessons for medical-agent developers:

Trustworthiness must be evaluated across the full workflow, not only on isolated answers.


Medical AI Is Moving From “Can It Work?” to “Can We Trust It?”

In 2026, the healthcare industry’s attention is shifting quickly from foundation models to agents.

At medical-information conferences, vendors are increasingly presenting full agent portfolios covering:

  • Intelligent diagnosis support;

  • Clinical documentation;

  • Patient services;

  • Hospital operations;

  • Evidence retrieval;

  • Pharmaceutical research;

  • Follow-up care.

China’s 2026 Government Work Report called for deeper expansion of the “AI Plus” initiative and faster deployment of intelligent agents.

In May, three national authorities jointly issued policy guidance on the standardized application and innovative development of AI agents.

The document defined agents as intelligent systems with capabilities such as:

  • Autonomous perception;

  • Memory;

  • Decision-making;

  • Interaction;

  • Execution.

The industry is no longer discussing only whether a model can answer medical questions.

It is discussing systems that can remember, plan, call tools, update records, and take action.

That expands both the value and the risk.

A hallucinated sentence in a chat interface is dangerous.

A hallucinated conclusion passed automatically into a medical record, prescription workflow, or follow-up plan is more dangerous still.

The transition from model to agent changes the trust question from:

Is this answer correct?

to:

Is the entire system capable of operating safely over time?


The Medical-Agent Ecosystem Is Already Taking Shape

Tencent Health has introduced a full-stack AI solution connecting users, hospitals, pharmaceutical companies, and research institutions through scenario-specific agents.

Its NGES 2026 architecture combines the Chonglou foundation model with MedAgent Suite, providing approximately 11 standardized skill modules that pharmaceutical companies can assemble according to their needs.

Ewell Health’s MedClaw initiative integrates evidence-based medical agents with a multi-agent framework, allowing physicians to invoke coordinated agent capabilities through a more unified interface.

These systems may support workflows such as:

  • Medical-record generation;

  • Guideline retrieval;

  • Literature analysis;

  • Quality inspection;

  • Diagnostic support;

  • Patient education;

  • Pharmaceutical research.

The agents are beginning to operate.

But one question remains unresolved:

Who is responsible for what they produce?


Hospitals Are Responsible Not Only for Adoption, but for Consequences

Healthcare institutions cannot treat AI deployment as a normal software-purchasing decision.

China’s trial rules on healthcare-institution data security and personal-information protection emphasize the primary responsibility of medical-institution leaders, as well as oversight responsibilities concerning third-party platforms.

This means that when a hospital introduces an AI product, it may assume responsibility for more than technical selection.

It must also consider:

  • Whether the system is compliant;

  • Whether data access is necessary and proportionate;

  • Whether model outputs are traceable;

  • Whether third-party providers meet security obligations;

  • Whether high-risk decisions remain under human supervision;

  • Whether failures can be audited and corrected.

A vendor cannot simply say, “The model generated the answer.”

The hospital cannot simply say, “The vendor supplied the model.”

In clinical environments, accountability must be designed into the workflow.

This is why low hallucination rates alone cannot define trustworthiness.


A Trustworthiness Tipping Point Requires More Than Fewer Hallucinations

Baichuan Intelligence has reported reducing the factual hallucination rate of its new medical model, Baichuan-M4, to approximately 3.3%.

Baichuan founder Wang Xiaochuan has summarized three rigid requirements for medical foundation models:

  • Low hallucination;

  • Strong evidence;

  • The ability to ask clarifying questions.

These three requirements capture the problem well.

Low hallucination

The model must avoid inventing facts, test results, clinical references, or diagnoses.

Strong evidence

The output should be connected to verifiable medical guidelines, literature, patient data, or clinical findings.

The ability to ask questions

When information is incomplete, the model should not guess confidently. It should recognize uncertainty and request the missing details.

Many general-purpose models remain inconsistent across all three dimensions in serious medical settings.

But even reducing hallucinations to 3.3% would not solve every problem.

A trustworthy medical-agent ecosystem must also be:

  • Economically sustainable;

  • Legally compliant;

  • Operationally stable;

  • Technically auditable;

  • Clinically supervised;

  • Continuously available.

That brings the discussion to a less visible but equally important layer: computing resources and Token costs.


Who Pays for the Tokens Behind Medical Agents?

Every agent call consumes resources.

A hospital workflow may involve:

  • Speech-to-text;

  • Record generation;

  • Record-quality inspection;

  • Evidence retrieval;

  • Diagnostic suggestions;

  • Risk scoring;

  • Image analysis;

  • Follow-up communication.

The system may use one model to generate an initial output, a second to verify it, and a third to retrieve supporting evidence.

To lower hallucination risk, hospitals may add:

  • Retrieval-augmented generation;

  • Multi-model verification;

  • Longer context;

  • Repeated checks;

  • Human-in-the-loop review;

  • Specialized medical tools.

Each safety layer adds cost.

A large tertiary hospital may generate more than 100,000 medical documents per day through intelligent systems.

Tencent’s medical stack includes multiple agents operating across different scenarios.

Every agent reads context, produces output, calls tools, and may validate other agents.

Token usage can grow rapidly.

For major hospitals, these costs may be absorbed through dedicated cloud contracts and larger IT budgets.

For smaller hospitals, community clinics, independent medical developers, and regional providers, the same architecture can become economically inaccessible.

This creates a serious paradox:

The institutions that most need AI to improve efficiency may be least able to afford the infrastructure required to deploy it safely.


Unexhub Is Addressing the Token Infrastructure Layer

Unexhub is attempting to restructure the global distribution of Tokens and AI computing resources.

Its core goal is to build a computing-power trading center focused on Token aggregation and distribution.

The platform connects distributed computing and model supply with demand from enterprises, developers, and AI applications.

In a medical-agent environment, one institution may need:

  • A low-cost model for transcription;

  • A balanced model for medical-record drafting;

  • A specialized model for evidence retrieval;

  • A multimodal model for medical imaging;

  • A premium reasoning model for high-risk review;

  • Backup capacity when one provider becomes unavailable.

Connecting each supplier separately creates substantial technical and commercial overhead:

  • Multiple APIs;

  • Separate accounts;

  • Different pricing rules;

  • Independent rate limits;

  • Fragmented invoices;

  • Different settlement cycles;

  • Additional failover logic.

A unified Token distribution and routing layer can simplify that structure.


Faster Settlement Helps Smaller AI Suppliers Survive

For small and medium-sized model and computing suppliers, cash flow can become a critical constraint.

A supplier may provide computing capacity today but wait weeks or months before receiving payment.

Meanwhile, it must continue paying for:

  • GPUs;

  • Electricity;

  • Cloud resources;

  • Bandwidth;

  • Operations;

  • Engineering staff.

Long settlement periods can place significant pressure on otherwise capable suppliers.

Unexhub’s financial settlement and clearing model is designed to support faster distribution of usage revenue.

The basic principle is:

When computing power is consumed, the corresponding supplier income can be settled more quickly.

This can help suppliers:

  • Shorten the cash-conversion cycle;

  • Reduce working-capital pressure;

  • Reinvest in infrastructure sooner;

  • Maintain more stable service;

  • Expand supply with less financial risk.

Eligible early suppliers may also receive promotional platform commission rates as low as 0%.

The value is not merely “zero commission.”

It may allow early participants to:

  • Keep more initial revenue;

  • Reduce channel costs;

  • Reach enterprise demand faster;

  • Test different pricing strategies;

  • Recover computing investments sooner.


API Bidding Could Make Token Pricing More Flexible

Traditional model APIs are usually sold at fixed list prices.

Users pay the same rate even when computing capacity is idle.

Unexhub’s planned API bidding mechanism allows prices to respond more dynamically to supply and demand.

During off-peak periods, developers may gain access to underutilized resources at lower prices.

During peak demand, suppliers may earn higher returns for making scarce capacity available.

For developers and medical institutions, this may provide:

  • Lower prices for non-urgent workloads;

  • More flexible scheduling;

  • Access to more model suppliers;

  • Less dependence on one vendor;

  • Better control of large Token budgets.

For suppliers, it may provide:

  • Higher utilization of idle resources;

  • Greater exposure to demand;

  • Better peak-period pricing;

  • Faster settlement;

  • Lower distribution costs.

This is not simply a temporary promotion.

It attempts to turn Token pricing into a market mechanism.


Trustworthiness Also Requires Intelligent Model Routing

A medical workflow should not use the same model for every step.

Medical workflowSuitable resource
Speech transcriptionFast, low-cost model
Medical-record draftingBalanced language model
Record-quality reviewSpecialized verification model
Guideline retrievalEvidence and search agent
Medical-image analysisMedical multimodal model
High-risk diagnostic reviewPremium reasoning model
Patient follow-upLow-cost conversational model

This architecture improves both cost and safety.

Low-risk work can be handled by economical models.

High-risk conclusions can be escalated to more capable models, independent verification agents, and human clinicians.

Dynamic routing can consider:

  • Clinical risk;

  • Task complexity;

  • Required latency;

  • Data type;

  • Model availability;

  • Price;

  • Regional requirements;

  • Compliance constraints.

A Token marketplace is therefore not merely a source of cheaper APIs.

It can become a resource-allocation layer supporting a safer agent architecture.


Unexhub’s Aggregation and Millisecond-Level Routing

Unexhub is designed to bring together AI suppliers of different sizes.

Its ecosystem may include access to major domestic and international models such as:

  • Gemini;

  • ChatGPT;

  • Claude;

  • DeepSeek;

  • GLM;

  • Specialized medical and industry models.

Individuals, startups, and small or medium-sized suppliers may also contribute high-quality models or computing capacity.

Requests can be routed in milliseconds according to:

  • Capability;

  • Price;

  • Latency;

  • Availability;

  • Region;

  • Task type;

  • Service quality.

For medical systems, this matters.

A low-risk administrative summary should not consume the same premium resource as a complex diagnostic review.

Routing can help preserve high-end capacity for situations where it delivers the greatest clinical value.


Paying by Actual Token Usage Lowers the Entry Barrier

Usage-based billing allows developers and institutions to pay according to actual Token consumption rather than purchasing large packages from several providers.

That can offer:

  • A lower starting budget;

  • Easier testing across models;

  • Gradual scaling;

  • Less unused prepaid capacity;

  • Faster model switching;

  • Spending that better reflects real workload volume.

Under selected supply conditions and promotional plans, effective API pricing may be as low as approximately 30% of standard market rates.

For medical-agent deployments, the value is not simply a cheaper individual call.

The same budget may support:

  • More consultations;

  • More generated medical records;

  • More evidence searches;

  • More independent verification;

  • Longer agent workflows;

  • Wider adoption among smaller institutions.

Naturally, affordability is only one requirement.

Healthcare users must also assess:

  • Data protection;

  • Privacy;

  • Reliability;

  • Compliance;

  • Auditability;

  • Permission management;

  • Service continuity;

  • Failure recovery.

Low prices create value only when the infrastructure remains secure and dependable.


Where PoloAPI Fits Into the Deployment Stack

Unexhub addresses the deeper market infrastructure for Token aggregation, distribution, bidding, and settlement.

At the practical application-access layer, a multi-model platform such as PoloAPI can help developers and enterprises connect several major models through one interface.

PoloAPI provides access to models including:

  • Gemini;

  • Claude;

  • ChatGPT;

  • DeepSeek;

  • Other domestic and international AI services.

For medical developers and enterprise AI teams, this can reduce the need to separately maintain multiple:

  • API integrations;

  • Authentication systems;

  • Billing accounts;

  • Error-handling rules;

  • Failover routes.

Practical benefits may include:

  • Unified model access;

  • Centralized usage and cost visibility;

  • Easier switching between models;

  • Backup routing;

  • High-concurrency support;

  • Enterprise customization.

Under selected resource plans, effective prices may be as low as approximately 30% of official rates.

For teams operating long-running or high-frequency agents, this can allow the same budget to support more calls, more users, and more independent verification.

PoloAPI also provides enterprise services covering model selection, API integration, permission configuration, concurrency planning, and deployment.

Its service experience includes large enterprises, publicly listed companies, state-owned organizations, and international customers.

The roles are complementary:

  • Unexhub focuses on Token-market infrastructure and supplier settlement;

  • PoloAPI provides a more direct multi-model access and enterprise-integration layer.

Both can help lower the cost and technical barriers of building a multi-model agent ecosystem.


The Tipping Point May Not Be at the Model Layer

When these signals are viewed together, a clearer picture emerges.

Medical agents are turning the clinical environment into a coordinated workflow.

Record generation, quality control, diagnostic assistance, guideline retrieval, and patient follow-up may each be handled by specialized agents.

Governments are encouraging deployment.

Technology companies are building products.

Hospitals are conducting trials.

But the real trustworthiness tipping point may not arrive when hallucination rates fall from 3% to 2%.

It may arrive when the whole system becomes:

  • Economically sustainable;

  • Legally deployable;

  • Operationally stable;

  • Clinically accountable;

  • Fully traceable.

A 0.3% hallucination rate is an important signal.

It shows what AI can achieve under carefully controlled conditions.

The 76.3% result at the opposite end is equally important.

It reminds us that “trustworthiness” is not a static property attached permanently to a model.

It depends on:

  • The text type;

  • The prompt;

  • The conversation history;

  • The evidence source;

  • The evaluation model;

  • The workflow;

  • The deployment environment.

Trustworthiness belongs to the system, not just the model.


The Bottom Line

Medical AI may indeed be reaching a turning point.

But the curve that is turning is not simply model capability.

It is the industry’s definition of trust.

Trust no longer means only:

The model rarely invents facts.

It now also means:

  • The model cites evidence;

  • The agent asks when information is missing;

  • The workflow remains consistent across multiple turns;

  • High-risk outputs receive independent review;

  • Actions can be audited;

  • Data remains protected;

  • Human responsibility is clear;

  • The economics remain sustainable.

Medical agents are rewriting the clinical workflow.

Unexhub is attempting to rewrite the Token-distribution and settlement logic beneath that workflow.

PoloAPI and other multi-model platforms can simplify practical access and deployment.

One layer determines what the agents can do.

Another determines how the necessary models and computing resources reach the institutions that need them.

The real industrial tipping point lies where those layers meet.

Has the turning point arrived?

Yes—but it is not simply a turning point in model intelligence.

It is a turning point in how the healthcare industry defines, evaluates, and communicates trust.


中文版

《自然·医学》0.3% 幻觉率:医疗智能体应用生态的“可信”拐点到了吗?

2026 年 7 月,一项发表于《自然·医学》的研究,为医疗 AI 领域带来了一个耐人寻味的数字。

研究人员选取 100 个真实患者问题,交由六个 AI 系统回答,其中包括前沿大语言模型与专业临床 AI 工具。随后,12 名临床医生通过双盲方式对结果进行评估。

最终结果颇为乐观:六个系统产生医学幻觉的比例整体处于较低水平。

但真正值得警惕的,是另一组同期研究数据。

一项针对真实临床数据的系统性研究发现,不同大语言模型的幻觉率可能在 0.3% 至 76.3% 之间波动。

具体结果取决于:

  • 医疗文本类型;

  • 提示词策略;

  • 模型配置;

  • 分类模型;

  • 评估方法。

同样一个“幻觉率”指标,最高值与最低值竟相差 250 多倍。

0.3% 听起来像是医疗 AI 已经跨过“可信拐点”的信号。

但 76.3% 又该如何解释?

答案是:幻觉率从来不只是一个孤立数字。

它的真实含义,取决于任务、上下文、对话历史、评估方式,以及模型所处的临床环境。


0.3% 的语境,比数字本身更重要

0.3% 并不是通用医疗 AI 的统一幻觉率。

这一结果是在一组明确限定的条件下测得的:

  • 特定临床文本;

  • 特定提示策略;

  • 特定模型设置;

  • 特定分类方法;

  • 受控评估流程。

换句话说,0.3% 代表的是模型在相对有利、边界清晰的条件下所能达到的表现。

这依然非常有意义。

它证明,当任务范围、上下文、证据来源和评估标准被严格控制时,医疗 AI 确实可以给出高度事实化的结果。

但这并不等于所有医疗聊天机器人、诊断智能体和临床工作流,都已经拥有相同可靠性。

在受控条件下回答一个结构化问题,与持续跟进一位患者的完整诊疗流程完全不同。

后者可能需要模型:

  • 理解长期病史;

  • 处理多轮对话;

  • 分析检验结果;

  • 调取循证材料;

  • 更新病历;

  • 持续调整建议。

因此,0.3% 更适合作为“能力上限已经出现”的信号,而不是可以覆盖所有医疗场景的安全保证。


更全面的评估,揭示了更复杂的现实

另一项发表于《自然·医学》的研究,给出了更完整的图景。

研究团队开发了 Medieval 评估框架,并邀请 27 名具有丰富经验的临床医生参与评价。这些医生平均从业时间达到 25.9 年。

评估范围包括:

  • 685 个真实及模拟临床案例;

  • 7 个前沿大语言模型;

  • 4,795 条模型回复;

  • 多种临床推理与沟通任务。

结果并不完全乐观。

约有 3% 至 19% 的模型回复被评为“不合格”,评估过程中还识别出 40 条幻觉内容。

表现最好的模型,在部分任务上已接近职业医生早期水平。

这是非常值得肯定的进步。

但它距离完整的临床可靠性仍有明显距离。

在普通软件产品中,3% 的错误率或许并不算高。

在医疗场景中,错误的严重程度取决于它发生在哪个环节。

100 份常规行政摘要中出现 3 个问题,可以通过人工审核纠正。

100 条用药建议、影像判断或高风险分诊结果中出现 3 个错误,却可能难以接受。

因此,医疗 AI 不能只看一个平均准确率。

不同临床任务的风险分布完全不同。

相同的错误比例,可能带来完全不同的后果。


多轮对话可能放大幻觉

一项发表于《npj Digital Medicine》的研究,带来了更加值得警惕的结果。

某临床 RAG 系统在没有对话历史时,幻觉率约为 5%。

当系统仅增加 10 轮历史问答后,幻觉率便上升至约 40%。

对话历史从 0 轮增加到 10 轮。

幻觉率扩大了 8 倍。

这对医疗智能体尤其关键,因为真实诊疗几乎从来不是单轮问答。

一条临床工作流可能涉及:

  • 初始症状;

  • 既往诊断;

  • 用药历史;

  • 检验结果;

  • 影像发现;

  • 医生追问;

  • 患者补充;

  • 病情变化;

  • 治疗方案调整。

智能体必须在整个过程中保持一致理解。

每增加一轮对话,都会增加新的风险:

  • 早期错误被不断放大;

  • 无关信息污染上下文;

  • 冲突内容被错误合并;

  • 模型过度依赖最初假设;

  • 最终结论与原始证据脱节。

一个在单轮问答测试中看似可信的系统,进入真实多轮临床对话后,可靠性可能迅速下降。

因此,医疗智能体开发者必须认识到:

可信度必须在完整工作流中评估,而不能只看孤立的一次回答。


医疗 AI 正从“能不能用”走向“能不能信”

2026 年,医疗行业的关注重点正在从基础大模型迅速转向智能体。

在各类医疗信息化大会中,厂商展示的已经不再只是一个医学问答模型,而是覆盖不同场景的智能体矩阵:

  • 智能诊疗;

  • 临床文书;

  • 患者服务;

  • 医院运营;

  • 循证检索;

  • 药物研发;

  • 随访管理。

2026 年政府工作报告明确提出,要深化拓展“人工智能+”行动,推动智能体加快落地。

5 月,三部门联合印发智能体规范化应用与创新发展相关政策文件,将智能体定义为具备以下能力的智能系统:

  • 自主感知;

  • 记忆;

  • 决策;

  • 交互;

  • 执行。

行业已经不再只讨论模型能否回答医学问题。

而是在讨论一套能够持续记忆、规划、调用工具、更新记录并采取行动的系统。

这会同时放大智能体的价值与风险。

聊天界面中出现一句幻觉内容已经足够危险。

如果错误结论被自动写入病历、用药流程或患者随访计划,后果则更加严重。

从模型走向智能体以后,“可信”的问题已经由:

这条回答是否正确?

升级为:

整套系统能否长期安全运行?


医疗智能体生态已经开始形成

腾讯健康推出了面向用户、医疗机构、药企和科研机构的 AI 全栈解决方案,以全场景智能体连接医疗健康创新流程。

NGES 2026 以“重楼”基座模型与 MedAgent Suite 为核心,向药企提供约 11 个可以灵活组合的标准 Skill 模块。

轻松健康的“正源方 MedClaw”,则将循证医学智能体与多智能体框架进行结合,让医生能够通过更加统一的入口调用多智能体协作能力。

这些系统可以承担:

  • 病历生成;

  • 指南检索;

  • 文献分析;

  • 质量审查;

  • 诊断支持;

  • 患者教育;

  • 药物研究。

智能体已经开始运转。

但一个问题始终没有消失:

谁来为智能体的输出负责?


医院不仅要对“引入”负责,还要对后果负责

医疗机构不能把引入 AI 当作一次普通的软件采购。

《医疗卫生机构数据安全和个人信息保护管理办法(试行)》强调医疗机构负责人的主体责任,以及医疗机构对第三方平台合规情况的监督责任。

这意味着,医院引入 AI 产品时,不仅要考虑技术表现,还要判断:

  • 系统是否符合监管要求;

  • 数据权限是否必要、适度;

  • 模型输出能否追溯;

  • 第三方供应商是否履行安全义务;

  • 高风险决策是否保留人工监督;

  • 出现错误后能否审计与纠正。

厂商不能简单地说:“答案是模型生成的。”

医院也不能简单地说:“产品是供应商提供的。”

临床责任必须被设计进整条工作流。

因此,单纯降低幻觉率,还不足以构成真正的可信生态。


“可信拐点”需要的,远不止更低的幻觉率

百川智能表示,新一代医疗模型 Baichuan-M4 的事实性幻觉率已降至约 3.3%。

王小川将医疗领域对大模型的刚性要求概括为三点:

  • 低幻觉;

  • 强证据;

  • 会提问。

这三个条件准确抓住了核心问题。

低幻觉

模型不能编造事实、检验结果、医学文献或诊断结论。

强证据

模型输出应当连接到可验证的诊疗指南、医学文献、患者数据或检查发现。

会提问

当现有信息不足时,模型不能自信地猜测,而应识别不确定性并主动补充询问。

在严肃医疗场景中,许多通用模型仍然无法稳定满足这三项要求。

但即使幻觉率已经降至 3.3%,仍然有一个更加基础的问题没有解决。

可信医疗智能体还必须做到:

  • 经济上可持续;

  • 法律上可落地;

  • 运行中足够稳定;

  • 结果能够审计;

  • 临床责任清晰;

  • 服务持续可用。

这就把讨论带到了一个经常被忽视的层面:算力和 Token 成本。


医疗智能体背后的 Token,由谁买单?

每一次智能体调用都会消耗资源。

一条医院工作流可能同时涉及:

  • 语音转写;

  • 病历生成;

  • 病历质检;

  • 循证检索;

  • 诊断提示;

  • 风险评分;

  • 医学影像分析;

  • 患者随访。

为了降低幻觉风险,医院还可能增加:

  • RAG 检索;

  • 多模型复核;

  • 更长上下文;

  • 重复校验;

  • 人工审核;

  • 专业医学工具调用。

每增加一层安全机制,都会增加成本。

大型三甲医院的智能系统,每天可能生成超过 10 万份医疗文书。

腾讯医疗体系中的多项智能体能力,也可能在不同工作流中同步执行。

每个智能体都要读取上下文、生成结果、调用工具,还可能验证其他智能体的输出。

Token 使用量会迅速增长。

大型医院可以通过专属云合同和更高 IT 预算消化部分成本。

但对中小医院、基层医疗机构、独立医疗开发者和区域医疗服务商来说,相同架构可能根本无法负担。

这形成了一个现实悖论:

越需要 AI 提升效率的机构,越可能缺乏安全运行智能体所需的基础设施预算。


Unexhub 正在解决 Token 基础设施问题

Unexhub 正尝试重构全球 Token 与 AI 算力资源的分发方式。

其核心目标,是建立面向 Token 聚合与分发的算力交易中心。

平台将分散的模型与算力供给,与企业、开发者和智能应用的需求连接起来。

在医疗智能体环境中,一家机构可能同时需要:

  • 低成本语音转写模型;

  • 均衡型病历生成模型;

  • 专业循证检索模型;

  • 医疗影像多模态模型;

  • 高风险复核推理模型;

  • 主通道异常时的备用资源。

如果分别接入每一家供应商,医院和开发团队就需要长期管理:

  • 多套 API;

  • 不同账户;

  • 各自的定价体系;

  • 独立限流规则;

  • 分散账单;

  • 不同结算周期;

  • 额外故障切换逻辑。

统一的 Token 分发与路由层,可以显著降低这一复杂度。


快速结算,帮助中小供应商持续供给

对中小模型和算力供应商而言,现金流往往是最严重的限制之一。

供应商今天提供算力,收入却可能要数周甚至数月后才能结算。

在等待回款期间,它们仍然要持续承担:

  • GPU 成本;

  • 电力;

  • 云端资源;

  • 带宽;

  • 运维;

  • 工程人员成本。

过长的回款周期,会让具有技术能力的中小供应商也难以持续经营。

Unexhub 的金融结算清算体系,强调更快地分配调用收入。

基本原则是:

算力被使用后,相应收益可以更快回到供应端。

这可以帮助供应商:

  • 缩短现金回收周期;

  • 缓解周转资金压力;

  • 更早投入新增算力;

  • 提升服务稳定性;

  • 降低扩张风险。

符合条件的早期供应商,在推广阶段还可享受最低 0 佣金的平台费率。

这项权益的价值不只是“零佣金”。

它还可能让早期参与者:

  • 保留更多初期收益;

  • 降低渠道成本;

  • 更快接触企业需求;

  • 测试更多定价方案;

  • 更快回收算力投入。


API 竞价,让 Token 价格更加灵活

传统模型 API 大多采用固定公开价格。

即使算力处于闲置状态,用户也需要按照统一费率调用。

Unexhub 计划中的 API 竞价机制,可以让价格更加灵活地反映供需关系。

在闲时,开发者可以用更低价格调用未充分利用的算力。

在需求高峰,供应商则可以因提供稀缺资源而获得更高回报。

对开发者与医疗机构而言,可以获得:

  • 更低的非紧急任务调用价格;

  • 更灵活的任务调度;

  • 更多模型供应选择;

  • 更低的单一厂商依赖;

  • 更可控的大规模 Token 预算。

对供应商而言,则可能获得:

  • 更高的闲置资源利用率;

  • 更广泛的需求来源;

  • 更合理的高峰期价格;

  • 更快的结算;

  • 更低的渠道分发成本。

这并不只是一次短期折扣。

它试图让 Token 定价逐渐成为一种市场机制。


可信医疗还需要智能模型路由

医疗流程并不需要在每一个环节都调用同一种模型。

医疗流程更适合的资源
语音转写快速低成本模型
病历起草均衡型语言模型
病历质检专业验证模型
指南检索循证与搜索智能体
医学影像分析医疗多模态模型
高风险诊断复核高端推理模型
患者随访低成本对话模型

这种架构可以同时改善成本与安全性。

低风险工作交给更经济的模型。

高风险结论升级到更强模型、独立复核智能体和临床医生。

动态路由可以综合考虑:

  • 临床风险;

  • 任务复杂度;

  • 延迟需求;

  • 数据类型;

  • 模型可用性;

  • 价格;

  • 地区限制;

  • 合规要求。

因此,Token 交易市场不只是提供更便宜的 API。

它还可以成为支持安全智能体架构的资源分配层。


Unexhub 的模型聚合与毫秒级路由

Unexhub 计划聚合不同规模的 AI 供应商。

平台生态可以覆盖海内外主流模型,例如:

  • Gemini;

  • ChatGPT;

  • Claude;

  • DeepSeek;

  • GLM;

  • 医疗及其他行业专业模型。

个人开发者、初创团队和中小供应商,也可以提供优质模型与算力资源。

请求可以根据以下条件进行毫秒级路由:

  • 模型能力;

  • 价格;

  • 延迟;

  • 可用性;

  • 地区;

  • 任务类型;

  • 服务质量。

这对医疗系统尤为重要。

低风险行政摘要,不应该消耗与高风险诊断复核相同的高端算力。

路由机制可以把优质资源保留给真正需要它们的场景。


按实际 Token 消耗付费,降低使用门槛

按量计费允许开发者和机构根据实际 Token 消耗付费,而不必分别向多个供应商购买大量固定套餐。

这能够带来:

  • 更低的起步预算;

  • 更方便地测试多个模型;

  • 根据业务逐步扩容;

  • 减少预付资源浪费;

  • 更快速地切换模型;

  • 让支出与真实工作量匹配。

在部分资源与推广方案下,实际 API 价格最低可达到市场标准价格的约 30%。

对医疗智能体而言,价值不只是单次调用更便宜。

同一预算可能支撑:

  • 更多患者咨询;

  • 更多病历生成;

  • 更多循证检索;

  • 更多独立复核;

  • 更长的智能体工作流;

  • 更广泛的基层医疗部署。

当然,可负担性只是其中一个条件。

医疗机构还必须评估:

  • 数据安全;

  • 隐私保护;

  • 服务可靠性;

  • 合规性;

  • 审计能力;

  • 权限管理;

  • 服务连续性;

  • 故障恢复。

只有价格优势与安全、稳定的基础设施同时存在,才能创造真正价值。


PoloAPI 在应用接入层的角色

Unexhub 解决的是更底层的 Token 聚合、分发、竞价与结算问题。

在实际应用接入层,PoloAPI 等多模型平台可以帮助开发者和企业,通过统一接口连接多款主流模型。

PoloAPI 可提供包括以下模型在内的接入能力:

  • Gemini;

  • Claude;

  • ChatGPT;

  • DeepSeek;

  • 其他国内外 AI 服务。

对医疗开发团队与企业 AI 部门而言,这可以减少分别维护多套系统的工作:

  • API 接口;

  • 鉴权方式;

  • 计费账户;

  • 错误处理规则;

  • 备用路由。

其实际权益可以包括:

  • 统一模型接入;

  • 集中查看调用量与费用;

  • 更方便地切换模型;

  • 备用通道;

  • 高并发支持;

  • 企业级定制。

在部分资源方案下,实际调用价格最低可达到官方价格的约 30%。

对长期、高频运行智能体的团队而言,同样预算可以支撑更多调用、更多用户和更多独立验证。

PoloAPI 还提供模型选型、API 集成、权限配置、并发规划和部署等企业服务。

相关服务经验已覆盖大型企业、上市公司、国有机构及海外客户。

两者可以形成互补:

  • Unexhub 聚焦 Token 市场基础设施与供应商结算;

  • PoloAPI 提供更加直接的多模型接入和企业集成能力。

这类基础设施可以共同降低建设多模型智能体生态的经济与技术门槛。


真正的拐点,可能不在模型层

将这些线索放在一起,未来图景开始变得清晰。

医疗智能体正在把诊室改造成一条协同运行的工作流。

病历生成、质量控制、诊断辅助、指南检索与患者随访,都可能由不同专业智能体承担。

政策正在推动。

厂商正在建设产品。

医院正在开展试点。

但“可信”的真正拐点,或许并不是模型幻觉率从 3% 降至 2% 的那一天。

它可能出现在整个系统同时具备以下条件时:

  • 经济上可持续;

  • 法律上可落地;

  • 运行上足够稳定;

  • 临床责任清晰;

  • 全流程能够追溯。

0.3% 的幻觉率是一个重要信号。

它证明在严格控制的条件下,AI 可以达到非常好的事实表现。

76.3% 的另一极端同样重要。

它提醒我们,“可信”并不是永久附着在某个模型上的固定属性。

它取决于:

  • 文本类型;

  • 提示策略;

  • 对话历史;

  • 证据来源;

  • 评估模型;

  • 工作流程;

  • 部署环境。

可信属于整个系统,而不只是模型本身。


写在最后

医疗 AI 或许确实已经走到一个拐点。

但发生变化的,不只是模型能力曲线。

更重要的是,行业正在重新定义“可信”。

可信已经不再只是:

模型很少编造事实。

它还意味着:

  • 输出能够引用证据;

  • 信息不足时智能体会主动追问;

  • 多轮流程中保持逻辑一致;

  • 高风险结论接受独立复核;

  • 所有操作可以被审计;

  • 患者数据得到保护;

  • 人类责任边界清晰;

  • 整套系统在经济上可以持续。

医疗智能体正在重写诊室流水线。

Unexhub 正试图重写这条流水线底层的 Token 分发与结算逻辑。

PoloAPI 等多模型平台,则可以简化真实应用中的模型接入与部署。

一层决定智能体能够做什么。

另一层决定模型与算力如何抵达真正需要它们的医疗机构。

行业真正的拐点,就在这两层交汇之处。

拐点到了吗?

到了。

但转弯的并不只是模型智能。

而是医疗行业对“可信”二字的定义、评估与表达方式。