广告
加载中

Gartner预测2028年Agent工作流推理成本将涨超五倍

亿邦AI 2026-10-19 09:32
亿邦AI 2026/10/19 09:32

邦小白快读

EN
全文速览

AI使用成本并不会随token降价同步走低,Gartner预测到2028年单个智能体工作流的推理成本将上涨超五倍,原因是效率提升带来的节省被投入到更贵的前沿模型研发,同时复杂任务消耗的token量远高于普通聊天。

1. 对普通用户和企业而言,低成本的AI交互可能涨价,尤其是Agent类应用,完成同等任务所需token量是基础聊天机器人的5到30倍。

2. 实战建议:部署AI前要明确成功阈值,不能放任运行;优先把常规任务分配给成本更低的模型;建立成本可观测机制,追踪每笔支出的实际产出。

3. 前沿模型和通用模型成本分化,Anthropic旗下Fable模型单token成本高于Opus,消费者需关注具体场景的定价变化。

品牌商依赖AI做营销、客服和用户运营时,需警惕推理成本上涨对预算的冲击,并学会用分层路由和结果衡量机制控制AI投入。

1. 成本趋势判断:基础AI能力使用成本下降,但前沿模型成本走高,品牌商若追求更智能的Agent服务,可能面临开支翻倍风险。

2. 用户行为洞察:消费端用量增长是成本抬升的核心驱动之一,复杂工作负载的token消耗远高于普通场景,意味着品牌商如果提供深度个性化服务,AI成本会显著上升。

3. 实操措施:投入可观测性建设,将AI功能绑定明确结果指标,如自动化完成的任务量、成功闭环数;常规任务禁用前沿模型,搭建Agent管控框架;善用专有数据微调垂直模型,形成差异化优势。

卖家在布局AI客服、智能运营等Agent应用时,需要重新评估成本模型,避免被推理成本上涨侵蚀利润。Gartner的预测给出了明确风险信号和应对路径。

1. 风险提示:到2028年Agent工作流推理成本上涨超五倍,当前企业为追同行进度会选择更高级的场景,导致token消耗量激增。

2. 市场机会:垂直领域模型在特定工作负载上表现优于通用前沿模型,卖家可基于专有数据微调,构建成本更低、效果更好的解决方案。

3. 应对措施:建立自动中止机制防止token支出失控,将常规高体量任务分配给低成本模型,仅对复杂场景使用前沿模型。

4. 落地建议:在AI部署中嵌入单结果价值衡量机制,对不合格工作流快速迭代或关停,保证每一分AI投入都有明确回报。

工厂在推进智能制造和数字化升级时,使用AI Agent会面临推理成本上涨的新问题,需要从单纯追求智能转向精细控制AI投入。

1. 需求变化:完成同等任务,AI Agent所需token量是基础机器人的5到30倍,意味着产线质检、设备预测性维护等复杂场景的AI成本会显著高于简单问答。

2. 选型启示:常规任务应路由到成本效率最高的模型,而非默认调用前沿模型;高效的垂直领域模型在生产可靠性上普遍优于通用模型,工厂可基于自身工艺数据微调。

3. 实施要点:在部署AI前明确成功阈值和合规开销,未通过成本压力测试不应上线;同时投资可观测性,否则无法优化看不见的成本。

服务商面临客户AI部署成本失控的痛点,推理悖论为行业带来新机会:谁能提供有效的成本优化方案,谁就能切走更大市场。

1. 趋势判断:token成本效率提升与复杂工作负载消耗激增并存,到2028年Agent工作流推理成本涨超五倍,服务商必须理解这一分化。

2. 客户痛点:企业普遍缺乏成本管控手段,调研显示21%使用编排平台的企业没有设置自动中止机制,31%会做低配分流,可见工具和咨询需求旺盛。

3. 解决方案:围绕AI FinOps建设可观测性,提供单结果价值衡量工具;设计推理分层机制和Agent管控框架;协助客户用专有数据微调垂直模型,降低长期成本并提高任务可靠性。

平台商运营AI相关业务时,需要应对推理成本上涨带来的商户流失风险,并主动提供成本管控工具和模型路由能力。

1. 平台需求:商户在Agent应用中的AI成本可能上涨超五倍,平台若不能帮助商户控制token支出,将导致应用活跃度下降。

2. 最新做法建议:在平台内嵌入自动中止机制、推理分层配置和可观测性面板,让商户可以设定准入标准并追踪工作流产出。

3. 风向规避:关注成本支出随新模型发布出现的多轮震荡,提前设计弹性计费或套餐模式,避免在tokenmaxxing潮流导致的支出高峰后迎来紧缩期。

4. 运营管理:鼓励商户使用专有数据微调高效垂直模型,平台可提供模型托管和小样本数据训练服务,帮助商户建立差异化能力。

推理悖论揭示了AI产业供给效率提升与需求消耗膨胀之间的结构性矛盾,是值得深入研究的新课题,也为商业模式和政策设计提供新视角。

1. 产业新动向:Gartner预测到2028年Agent工作流推理成本上涨超五倍,基础研究机构将效率收益重新投入更高算力模型,导致前沿模型单token成本不降反升。

2. 量化分析:五倍预测基于Tokenomics模型,纳入超300个输入项,覆盖成本效率提升和复杂场景消耗激增两类因素,呈现非线性波动。

3. 企业行为证据:调研显示21%企业无自动中止机制,31%采用低成本模型分配,反映成本治理能力落后于技术应用。

4. 政策与商业模式启示:需要建立成本可视化和问责机制,可考虑设置AI部署最低准入标准;专有数据微调、垂直模型服务可能成为可持续的商业模式。

返回默认

声明:快读内容全程由AI生成,请注意甄别信息。如您发现问题,请发送邮件至 run@ebrun.com 。

我是 品牌商 卖家 工厂 服务商 平台商 研究者 帮我再读一遍。

Quick Summary

AI usage costs will not decline in line with token price drops. Gartner predicts that by 2028, inference costs for a single agent workflow will rise more than fivefold, because efficiency gains are being reinvested into more expensive frontier model R&D, while complex tasks consume far more tokens than ordinary chats.

1. For ordinary users and businesses, low-cost AI interactions may become more expensive, especially agent-based applications, which require 5 to 30 times more tokens than basic chatbots to complete equivalent tasks.

2. Practical advice: Before deploying AI, define a clear success threshold and do not let it run unchecked; prioritize routing routine tasks to lower-cost models; and establish cost observability to track the actual output of every expense.

3. Costs are diverging between frontier and general-purpose models. Anthropic's Fable model has a higher per-token cost than Opus, so consumers should watch pricing changes in specific use cases.

Brands relying on AI for marketing, customer service, and user operations must watch for the impact of rising inference costs on budgets and learn to control AI spending through layered routing and outcome measurement.

1. Cost trend: Basic AI capabilities are becoming cheaper, but frontier model costs are rising. Brands that pursue smarter agent services may face the risk of doubling expenses.

2. User behavior insight: Growth in consumer-side usage is a core driver of rising costs. Complex workloads consume far more tokens than ordinary scenarios, meaning that if brands provide deeply personalized services, their AI costs will increase significantly.

3. Practical measures: Invest in observability and tie AI features to clear outcome metrics, such as the number of automated tasks completed and successful closed loops; disable frontier models for routine tasks and build an agent governance framework; use proprietary data to fine-tune vertical models and create differentiated advantages.

Sellers deploying agent applications such as AI customer service and intelligent operations need to reassess their cost models to prevent rising inference costs from eroding profits. Gartner's forecast offers clear risk signals and response paths.

1. Risk warning: By 2028, inference costs for agent workflows will increase more than fivefold. Companies chasing peer progress today are selecting more advanced scenarios, driving a surge in token consumption.

2. Market opportunity: Vertical-domain models outperform general frontier models on specific workloads. Sellers can fine-tune based on proprietary data to build lower-cost, higher-performance solutions.

3. Countermeasures: Establish automatic stop mechanisms to prevent runaway token spending, allocate high-volume routine tasks to low-cost models, and reserve frontier models for complex scenarios.

4. Implementation advice: Embed a per-outcome value measurement mechanism in AI deployments, rapidly iterate or shut down underperforming workflows, and ensure every AI investment delivers clear returns.

When factories advance smart manufacturing and digital upgrades, using AI agents introduces the new challenge of rising inference costs, requiring a shift from simply pursuing intelligence to finely controlling AI spending.

1. Demand shift: Completing the same task with an AI agent requires 5 to 30 times more tokens than a basic bot, meaning AI costs for complex scenarios such as production-line inspection and predictive equipment maintenance will be significantly higher than for simple Q&A.

2. Model selection insight: Routine tasks should be routed to the most cost-efficient models rather than defaulting to frontier models; high-efficiency vertical models are generally more reliable in production than general-purpose ones, and factories can fine-tune based on their own process data.

3. Implementation points: Define success thresholds and compliant spending limits before deploying AI; do not launch without passing a cost stress test; and invest in observability, otherwise invisible costs cannot be optimized.

Service providers face clients struggling with runaway AI deployment costs. The inference paradox creates a new opportunity: whoever can deliver effective cost optimization will capture a larger market.

1. Trend assessment: Token cost efficiency gains coexist with a surge in complex workload consumption. By 2028, agent workflow inference costs will rise more than fivefold, and service providers must understand this divergence.

2. Client pain points: Companies generally lack cost control mechanisms. Research shows that 21% of enterprises using orchestration platforms have no automatic stop mechanism, and 31% use low-cost fallback routing, indicating strong demand for tools and consulting.

3. Solutions: Build observability around AI FinOps, offering per-outcome value measurement tools; design inference layering mechanisms and agent governance frameworks; help clients fine-tune vertical models with proprietary data to lower long-term costs and improve task reliability.

Platform operators running AI-related businesses need to address the risk of merchant churn from rising inference costs and proactively provide cost control tools and model routing capabilities.

1. Platform needs: Merchant AI costs in agent applications may rise more than fivefold. If platforms cannot help merchants control token spending, application activity will decline.

2. Recommended practices: Embed automatic stop mechanisms, inference layering configurations, and observability dashboards into the platform, allowing merchants to set admission standards and track workflow output.

3. Wind-shear avoidance: Watch for multiple rounds of cost volatility following new model releases, and design flexible billing or package models in advance to avoid a contraction after spending peaks driven by the token-maxing trend.

4. Operations management: Encourage merchants to fine-tune efficient vertical models with proprietary data. Platforms can offer model hosting and small-sample data training services to help merchants build differentiated capabilities.

The inference paradox reveals a structural contradiction between rising supply-side efficiency and exploding demand-side consumption in the AI industry. It is a new subject worthy of deeper research and offers fresh perspectives for business model and policy design.

1. Industry trend: Gartner predicts that agent workflow inference costs will rise more than fivefold by 2028. Basic research institutions are reinvesting efficiency gains into higher-compute models, causing frontier model per-token costs to rise rather than fall.

2. Quantitative analysis: The fivefold forecast is based on a Tokenomics model incorporating over 300 inputs, covering both cost efficiency gains and the surge in complex scenario consumption, showing nonlinear dynamics.

3. Evidence of enterprise behavior: Surveys show 21% of enterprises lack automatic stop mechanisms, and 31% use low-cost model allocation, reflecting cost governance capabilities lagging behind technology adoption.

4. Implications for policy and business models: There is a need for cost visibility and accountability mechanisms, possibly setting minimum admission standards for AI deployment. Proprietary data fine-tuning and vertical model services may become sustainable business models.

Disclaimer: The "Quick Summary" content is entirely generated by AI. Please exercise discretion when interpreting the information. For issues or corrections, please email run@ebrun.com .

I am a Brand Seller Factory Service Provider Marketplace Seller Researcher Read it again.

2026年10月16日,市场研究机构Gartner发布AI成本相关预测。当下token价格持续下行,不少企业默认AI整体成本将同步走低,Agent类应用的落地正在打破这一判断。Gartner将相关现象命名为推理悖论,在token经济性持续优化的背景下,到2028年单个智能体工作流的推理成本将上涨超五倍。将任务分配给Agent推理模型,相比基础聊天机器人交互,服务商的推理成本至少提升5倍,且随任务复杂度提升成本会进一步走高,单位经济性优化带来的空间被消耗在更高的AI投入中,目前暂未形成匹配且可预测的价值回报路径。

效率提升带来的成本空间,正在被重新投入到性能更强、成本更高的模型研发中。Gartner高级总监分析师Will Sommer提到,token生成的成本效率确实在持续提升,但基础AI实验室并未留存这部分成本节约,反而转向研发需要更高算力支撑的下一代模型,这类模型的单token成本更高。旧有AI能力的使用成本持续下降,前沿模型的使用成本反而走高,Anthropic旗下Fable模型的单token成本高于Opus,正是这一分化的典型案例。

消费端的用量增长是成本抬升的另一核心驱动。token效率提升让更多复杂工作负载具备落地可行性,这类负载的token消耗量远高于普通场景。完成同等任务,AI Agent需要的token量是基础AI聊天机器人的5到30倍。竞争压力将两端因素绑定,企业为追上同行进度,会选择更高级的应用场景,这类场景需要性能更强、价格更高的模型,对应的token消耗量也远高于上一代AI系统。

Gartner判断,到2028年相关成本不会呈线性平稳上涨,控制成本和保持竞争力之间的拉扯会推动支出上下波动,支出水平已随新模型发布出现多轮震荡。第一波Agent能力落地在今年春季推高了行业整体AI支出,其中tokenmaxxing潮流是部分推动因素,当前行业正处于紧缩开支的阶段。后续随推理能力更强的Agent、世界模型等AI技术成熟,还将出现数轮支出波动。

不同企业的成本管控水平存在明显差距。今年9月一项覆盖109名合格受访者的行业调研显示,使用编排平台的企业中,21%没有设置任何自动中止Agent token支出的机制,31%会将常规、高体量的Agent任务分配给成本更低的模型,以此压缩开支。本次五倍成本增长的预测基于Gartner的Tokenomics预测模型,该计量模型纳入超300个输入项预测长期token成本变化,同时覆盖token成本效率提升、AI应用于更复杂场景带来的token消耗量激增两类因素。具备学习能力、在企业内部跨域或在多领域间运行的AI解决方案,工作流落地需要消耗的token量会显著更高。

针对推理悖论带来的成本压力,Gartner梳理了五个方向的架构调整路径。企业可投入可观测性能力建设,AI FinOps是所有成本优化工作的核心,无法观测的成本就无法优化。在AI部署过程中嵌入单结果价值衡量机制,要求所有AI功能针对各类不确定场景做产出预测,追踪支出对应的明确结果指标,比如自动化完成的任务量、成功闭环的案例数,快速迭代或关停表现不佳的工作流以保障利润。落地推理分层机制,通过自动化流程将查询路由到成本效率最高的模型,常规任务默认禁止调用前沿模型,搭建Agent管控框架,明确自主系统内模型部署的场景、时间和方式。落实AI执行的最低准入标准,设置后即放任运行的模式不可行,企业需要提前明确成功阈值、合规开销和风险缓释成本,在场景通过token价格波动、合规支出压力测试前,不批准相关部署上线。善用专有数据,通过专有数据微调和自学习反馈循环,长期维持差异化优势和利润水平,高效的垂直领域模型在特定工作负载的可靠性和任务表现上,普遍优于通用前沿模型。

本文首发于 亿邦动力 官方网站

文章来源:亿邦动力

广告
微信
朋友圈

FAQ回顾

什么是AI推理悖论?

推理悖论是Gartner提出的现象:token生成效率持续提升、单位成本下降,但到2028年单个智能体工作流的推理成本反而将上涨超五倍。原因在于AI实验室将成本节约投入研发单token成本更高的下一代模型,同时Agent等复杂工作负载的token消耗量远高于普通场景,导致总体推理成本不降反升。

为什么Agent工作流推理成本会增加?

核心是“两端绑定”:一方面Anthropic Fable等前沿模型的单token成本高于Opus,企业为追赶竞争会选择性能更强、价格更高的模型;另一方面AI Agent完成任务所需token量是基础聊天机器人的5到30倍,单位经济性优化省下的成本被更高用量和更强模型消耗掉,因此2028年前推理成本将持续承压。

AI Agent的token消耗量比普通聊天机器人高多少?

Gartner指出,完成同等任务,AI Agent需要的token量是基础AI聊天机器人的5到30倍;将任务分配给Agent推理模型,相比基础聊天机器人交互,服务商推理成本至少提升5倍。token用量增加叠加前沿模型单token成本更高,使Agent应用落地带来显著成本压力。

企业如何降低Agent推理成本?

Gartner建议五方面调整:建设可观测性并落地AI FinOps;为AI部署嵌入单结果价值衡量,追踪自动化任务量等指标并关停低效工作流;通过推理分层将查询路由到成本效率最高的模型,并搭建Agent管控框架;设置AI执行最低准入标准;用专有数据微调和自学习反馈循环,保持差异化优势。

这么好看,分享一下?

朋友圈 分享

APP内打开

赞 +1
+1
微信好友 朋友圈 新浪微博 QQ空间
关闭
收藏成功
发送
/140 0