
Bilingual Article | English First, Chinese Version Below
DeepSeek Cuts Prices by 75% While AI Infrastructure Costs Surge: The Real Cost Revolution Is Moving from Cheaper Tokens to a Tradable Compute Market
During the first half of 2026, the global AI industry appeared to be moving in only one direction: higher costs.
HBM prices reportedly surged by more than 500% within six months, while high-end GPUs remained in short supply. Amazon, Microsoft, and several major Chinese cloud providers adjusted their API prices, with increases in some categories reaching several hundred percent.
The market seemed to have reached a simple conclusion:
As AI demand grows, AI services will only become more expensive.
DeepSeek chose a different path.
Instead of passing rising infrastructure costs directly to developers, it permanently reduced the API price of its flagship V4-Pro model by 75%.
The move did more than lower one model’s price. It opened a new phase of competition in which model architecture, caching, routing, and market-based distribution all contribute to the final cost of completing an AI task.
A Permanent 75% Price Cut, with Cached Input Falling to One-Fiftieth of the Previous Level
On May 22, DeepSeek announced a permanent 75% reduction in the API price of V4-Pro.
A promotional rate that had originally been scheduled to end on May 31 was converted into the model’s permanent public price.
The new pricing structure included:
Cached input: reduced from RMB 0.10 to RMB 0.025 per million Tokens;
Uncached input: reduced from RMB 12 to RMB 3 per million Tokens;
Output: reduced from RMB 24 to RMB 6 per million Tokens.
At RMB 0.025 per million Tokens, cached input costs roughly $0.0036 per million Tokens, depending on the exchange rate.
For applications with highly repetitive prompts, this is an unusually low price for access to a high-capability model.
DeepSeek’s move was not an isolated event.
Xiaomi subsequently announced permanent reductions across the MiMo-V2.5 API family, with discounts approaching 99% in selected scenarios. Tencent Cloud followed in early June with price reductions reaching as high as 97.5% in some categories.
What began as one provider’s pricing decision quickly developed into a broader Token-price correction.
The result is a new competitive environment: infrastructure costs may be rising, but the cost of model inference can still fall when architecture, utilization, and distribution efficiency improve faster than hardware expenses.
Why Can DeepSeek Cut Prices While Hardware Costs Are Rising?
DeepSeek’s price reduction is not presented simply as a short-term subsidy.
Its more important foundation is a restructuring of inference cost.
The V4 series reportedly combines a proprietary sparse-attention mechanism with a mixture-of-experts architecture. When processing million-Token contexts, its compute consumption is said to be significantly lower than that of the previous generation, while KV Cache usage is also substantially reduced.
The model has also been adapted for Huawei Ascend hardware, giving DeepSeek more flexibility in how it deploys inference capacity and reducing its dependence on a single class of imported high-end accelerators.
Several cost advantages are therefore working together:
Fewer active parameters per request;
Lower attention cost for long contexts;
Reduced KV Cache requirements;
Improved hardware adaptation;
Higher serving utilization;
Greater economies of scale.
This is the key distinction between a temporary price war and a structural price reduction.
A subsidy can make an API cheaper for several months.
A more efficient architecture can keep it cheaper for much longer.
When model providers compete on intelligence produced per dollar rather than benchmark scores alone, Token pricing becomes a direct reflection of engineering efficiency.
Caching Turns Repetition into the Cheapest Part of the Workflow
The 75% price reduction is the visible part of the story.
Caching may have an even greater effect on enterprise spending.
A large share of production AI traffic is repetitive. Applications repeatedly send the same system prompt, knowledge-base instructions, product descriptions, policy documents, or workflow definitions.
Without caching, the model processes that information again during every request.
With caching, the reusable portion only needs to be processed once.
OpenRouter introduced Response Caching on April 30. When an identical request matches a stored response, the platform can return the previous result without generating a new model response or charging the normal Token cost.
This also reduces latency. Instead of waiting several seconds for inference, a cached response may be returned within hundreds of milliseconds, while the cache lookup itself can take only a few milliseconds.
OpenRouter also supports Prompt Caching.
Under this model, repeated prompt content is not completely free, but the read price can be far below the normal input rate.
For example, a model with a standard input price of $3 per million Tokens may charge only $0.30 per million Tokens for cached reads—a 90% reduction.
DeepSeek and Alibaba’s Qwen ecosystem also provide deeply discounted cache-read pricing for supported workloads.
The financial impact becomes clear when the same context is reused thousands or millions of times.
A customer-service system may repeat the same service policy in every conversation.
A coding agent may load the same repository instructions for every task.
A document assistant may repeatedly read the same company template and compliance rules.
Caching turns these repeated inputs from a permanent expense into one of the least expensive parts of the workflow.
Response Caching and Prompt Caching Solve Different Problems
Although both mechanisms reduce cost, they work at different levels.
Response Caching
Response Caching is useful when the entire request is identical.
If the prompt, parameters, and context are unchanged, the platform can return the stored result without running inference again.
It is suitable for:
Frequently asked questions;
Repeated classification queries;
Static product descriptions;
Standard policy explanations;
Deterministic content retrieval;
Development and testing requests.
In ideal scenarios, the model-inference cost of a cache hit can approach zero.
Prompt Caching
Prompt Caching is useful when part of the request remains the same while another part changes.
A long system prompt, handbook, codebase summary, or product catalog can be cached, while the user’s new question is processed normally.
It is suitable for:
Enterprise knowledge assistants;
Long-context coding agents;
Contract analysis;
Repeated document review;
Customer-service systems;
Multi-step agent workflows.
Prompt Caching does not eliminate the entire request cost, but it can sharply reduce the cost of repeatedly processing the same context.
When the two mechanisms are used correctly, enterprises are no longer paying full price every time their applications repeat something the model has already seen.
Does “Cached Input at One-Fiftieth” Mean Every Workload Costs 98% Less?
Not necessarily.
The lowest cached-input price applies only to the portion of a request that successfully matches the cache.
New questions, uncached context, model output, tool calls, failed requests, and application infrastructure still generate costs.
The total saving therefore depends on the workload’s cache-hit rate.
A workflow with little repeated content may see only a limited benefit.
A workflow in which 80% or 90% of the prompt remains unchanged can achieve a much larger reduction.
For example, suppose an enterprise agent repeatedly processes a 100,000-Token operating manual while each employee asks a different 1,000-Token question.
Without caching, the full 101,000 Tokens may be processed for every request.
With prompt caching, the operating manual can be reused at a deeply discounted read price, while only the new question and output are charged at standard rates.
At scale, this can reduce the cost per completed task far more than a simple 75% cut in the model’s public input price.
The most accurate conclusion is therefore:
Caching does not make every AI request nearly free. It makes repeated context dramatically cheaper.
For high-repetition enterprise workloads, total cost reductions above 90% may be achievable when low model prices, high cache-hit rates, batch execution, and intelligent routing are combined.
OpenRouter Makes Existing Prices Easier to Use
DeepSeek lowers the supplier’s price.
OpenRouter helps developers access, compare, route, and cache those model services through a more unified gateway.
Its value lies in reducing technical and operational fragmentation:
One compatible API;
Centralized usage visibility;
Multiple providers behind one interface;
Provider routing and fallback;
Prompt and response caching;
Fewer SDKs and separate billing systems.
However, the underlying model prices are still generally published by suppliers.
Developers may select the cheapest available provider, but the price itself remains a supplier quote rather than a continuously negotiated market rate.
OpenRouter therefore improves the question:
How can I buy existing model capacity more efficiently?
Unexhub addresses a different question:
How can model capacity itself become a dynamically priced and settled marketplace resource?
Unexhub: Moving Token from a Billing Unit to a Market Resource
Unexhub aims to build a Token aggregation and distribution marketplace that connects mainstream model providers, specialized AI suppliers, enterprise customers, and developers.
The platform is designed to aggregate model ecosystems such as Gemini, ChatGPT, Claude, DeepSeek, and GLM, while also providing distribution opportunities for individual developers, small and medium-sized AI companies, and providers with specialized or underused inference capacity.
Its direction goes beyond placing several models behind one API.
Unexhub aims to introduce a market layer in which Token capacity can be:
Aggregated from different suppliers;
Competitively offered;
Routed according to price and performance;
Metered according to actual consumption;
Settled more quickly after successful usage.
In this structure, Token is no longer only the number shown in a model invoice.
It becomes the unit through which supply is measured, traffic is allocated, revenue is calculated, and transactions are settled.
Faster Settlement: Giving Suppliers Better Capital Turnover
Traditional API distribution often includes long settlement cycles.
A smaller supplier may provide model capacity today but wait weeks before receiving the corresponding revenue. During that period, GPU, cloud, bandwidth, storage, and operational expenses still need to be paid.
This creates working-capital pressure, particularly for smaller suppliers that lack the balance sheets of large cloud providers.
Unexhub is designed to shorten the settlement cycle.
Once a request has been successfully completed and its usage verified, the platform can calculate the supplier’s revenue share and move more quickly toward settlement.
The practical benefits may include:
Faster recovery of infrastructure expenses;
Improved cash-flow visibility;
Lower working-capital pressure;
More funds available for capacity expansion;
Greater willingness to release idle compute to the marketplace.
Early suppliers may also qualify for promotional commission rates as low as 0%.
This gives smaller providers a lower-cost route into a broader Token distribution network without requiring them to independently build customer acquisition, global billing, routing, and clearing systems.
The value is not that every financial burden immediately disappears.
The real benefit is faster capital turnover, lower early distribution costs, and a clearer path from completed inference to realized revenue.
API Bidding: Let Prices Move with Supply and Demand
One of the more ambitious mechanisms planned by Unexhub is API bidding.
Instead of keeping one permanent price for every model and provider, suppliers can adjust their offers according to available capacity and market demand.
The concept resembles an electricity spot market.
Electricity prices rise when demand is high and available supply is tight. They fall when demand is low and excess capacity is available.
AI inference has similar cycles.
During weekday business hours, enterprise demand may be concentrated and serving capacity may become more valuable.
During late nights, weekends, or low-utilization periods, large amounts of compute may remain idle.
A market-responsive pricing system can create value for both sides.
For developers:
Non-urgent batch workloads can be moved to lower-priced periods;
Requests can be automatically matched with more economical suppliers;
Prepaid balance waste can be reduced through usage-based billing;
Selected model capacity may be available at substantially lower market prices;
Cost, latency, and quality can be balanced according to each task.
For suppliers:
Peak demand can generate higher revenue;
Idle capacity can be monetized through lower off-peak offers;
Traffic can be distributed according to real market demand;
Revenue can be calculated from successful Token usage;
Smaller providers can compete on price, performance, and specialization.
This mechanism does not mean every model will eventually have the same Token price.
Differences in quality, context length, latency, reliability, region, security, and compliance will continue to produce different price levels.
What changes is how those prices are discovered.
They are no longer based only on a supplier’s permanent catalog. They can also reflect how much capacity is available and how strongly the market needs it at that moment.
Millisecond-Level Routing Makes Market Pricing Usable
A bidding market is only valuable if developers do not need to manage it manually.
No engineering team wants to compare dozens of providers before every API call.
Unexhub therefore combines competitive supply with automated routing.
The platform can evaluate available providers according to:
Current price;
Model capability;
Response latency;
Throughput;
Historical reliability;
Regional availability;
Compliance requirements;
Enterprise service-level targets.
It can then match a request with an appropriate supplier in milliseconds.
The developer continues to use one access layer and pays according to actual Token consumption.
Supplier comparison, traffic allocation, metering, and settlement happen in the background.
For high-volume users, this provides more than convenience.
It allows routine tasks to use more economical capacity while premium models remain available for complex reasoning and business-critical work.
Selected services may be available at prices as low as approximately 30% of typical market rates, depending on the model, supplier, demand period, and routing conditions.
The value of that discount becomes much more meaningful at scale.
Saving a fraction of a cent on one call may not matter.
Saving it across millions of classifications, document extractions, customer conversations, or agent steps can materially change a product’s gross margin.
From a Cheap Token to a Lower Cost per Completed Task
The AI industry often compares providers using price per million Tokens.
That metric is useful, but incomplete.
Enterprises do not purchase Tokens for their own sake. They purchase results:
A customer issue resolved;
A document reviewed;
A contract analyzed;
A report generated;
A coding task completed;
A voice recording transcribed;
A sales lead qualified.
A cheaper model may generate longer outputs, fail more often, or require more retries.
A more expensive model may complete the same task in one attempt.
The more useful metric is therefore:
Cost per successful completed task.
DeepSeek’s lower list price improves the starting economics.
Caching reduces the cost of repeated context.
OpenRouter helps select and manage provider routes.
Unexhub adds competitive allocation and market-responsive pricing.
When these layers are combined, enterprises gain several opportunities to reduce unit-task cost:
Choose an appropriately sized model;
Cache reusable prompts and responses;
Route routine traffic to more economical suppliers;
Schedule batch jobs during off-peak periods;
Avoid paying full price for repeated context;
Reduce failed requests and unnecessary retries;
Pay according to actual usage instead of unused packages.
This is the deeper meaning of Token cost optimization.
It is not one permanent discount.
It is a system that continuously searches for a more efficient way to deliver the same business outcome.
The Token Spending Index Shows That the Market Is Responding
The impact of lower model prices and stronger routing competition is beginning to appear in market indicators.
The Silicon Data LLM Token Spending Index, which tracks the average price paid per million LLM Tokens, reportedly rose sharply after its launch but later declined from its May peak.
Some financial institutions have also projected that average Token prices could fall further over the following twelve months.
At the same time, routing-platform data indicates growing developer interest in lower-cost and open-weight models.
DeepSeek models have gained substantial traffic across aggregation platforms, while the share of open-weight models in AI gateway usage has increased significantly.
This reflects a straightforward market behavior:
When a model is capable, accessible, and economical, developers move traffic toward it.
Brand loyalty matters less when an application can switch providers through one compatible interface.
The routing layer gives developers the ability to vote with every request.
Three Layers, One Cost Revolution
DeepSeek, OpenRouter, and Unexhub represent three different layers of the same transformation.
DeepSeek Changes the Supply Cost
Architecture and inference optimization allow the model provider to lower its public price.
OpenRouter Changes Access and Reuse
A unified gateway, provider routing, and caching reduce the engineering and repetition costs surrounding model usage.
Unexhub Changes Distribution and Settlement
Competitive supply, dynamic routing, faster settlement, and market-responsive pricing aim to turn Token capacity into a more liquid marketplace resource.
One makes the model cheaper.
One makes repeated use cheaper.
One attempts to make the market itself more efficient.
Together, they point toward a future in which AI cost is determined not only by the model selected, but also by:
Which supplier serves the request;
Whether the context is cached;
When the workload is executed;
How much capacity is available;
How the request is routed;
How quickly supplier revenue is settled.
Final Takeaway
DeepSeek’s permanent 75% reduction opened a visible gap in the established Token-pricing structure.
OpenRouter’s caching mechanisms push the cost of repeated prompts and identical requests much closer to zero.
Unexhub points toward the next stage: Token capacity aggregated across suppliers, competitively priced, routed in milliseconds, and settled according to actual usage.
For developers and enterprise users, this can translate into:
Unified access to major domestic and international models;
Pay-as-you-go Token billing;
Less capital trapped in unused prepaid packages;
Lower costs for repetitive workloads through caching;
Dynamic selection between price, speed, and capability;
Selected model services available at approximately 30% of typical market rates.
For suppliers, the proposed benefits include:
One-step access to broader distribution;
Automated traffic allocation;
Faster usage-based settlement;
Early-stage commission rates as low as 0%;
Improved utilization of idle inference capacity;
A lower-cost path to enterprise and international demand.
The real cost revolution is therefore no longer only about identifying the cheapest model.
It is about answering three more practical questions:
When should the capacity be purchased?
Which supplier should provide it?
How can the same business result be completed with fewer paid Tokens?
Token is moving beyond its original role as a technical measurement unit.
Across text, images, speech, documents, and agent workflows, it is becoming a unit of pricing, routing, settlement, and resource allocation.
That is the deeper truth behind the next phase of the Token marketplace.
中文版|English Version Above
本文为英中双语文章。以上为英文全文,以下为完整中文版本。
DeepSeek 逆市降价 75%:真正的降本革命,正在从“便宜 Token”走向“可交易算力”
2026 年上半年,全球 AI 产业似乎只剩下一条主线:成本上涨。
HBM 价格据称在半年内暴涨超过 500%,高端 GPU 持续供不应求。亚马逊、微软以及国内多家主流云厂商陆续调整 API 价格,部分品类的涨幅甚至达到数倍。
市场似乎形成了一个简单共识:
随着 AI 需求持续增长,模型服务只会越来越贵。
DeepSeek 却没有按照这套剧本发展。
它没有把不断上涨的基础设施成本直接转嫁给开发者,而是宣布将旗舰模型 V4-Pro 的 API 价格永久下调 75%。
这次调整的意义,不只是某个模型突然变便宜。
它撕开了一道新的竞争口子:未来决定 AI 任务成本的,不再只有模型目录价格,还包括底层架构、缓存命中率、供应商路由,以及更加市场化的算力分发机制。
永久降价 75%,缓存命中输入降至原来的四分之一
5 月 22 日,DeepSeek 宣布永久下调 V4-Pro API 价格。
原定于 5 月 31 日结束的 2.5 折限时优惠,被直接转化为长期挂牌价格。
调整后的价格包括:
缓存命中输入:从每百万 Token 0.1 元降至 0.025 元;
缓存未命中输入:从每百万 Token 12 元降至 3 元;
输出:从每百万 Token 24 元降至 6 元。
每百万 Token 0.025 元的缓存命中输入价格,按汇率折算约为 0.0036 美元。
对于系统提示词、知识库内容和工作流指令高度重复的应用而言,这已经是极具竞争力的价格水平。
DeepSeek 的降价并不是孤立事件。
随后,小米 MiMo-V2.5 API 系列宣布永久降价,部分使用场景的降幅接近 99%;腾讯云也在 6 月初跟进,部分产品最高降幅达到 97.5%。
一场由 DeepSeek 引发的 Token 价格调整,迅速蔓延到更多模型与云服务平台。
这也揭示了一个容易被忽略的事实:
即使硬件价格持续上涨,只要模型架构、推理效率和资源利用率提升得更快,最终交付给开发者的模型调用价格仍然可以下降。
DeepSeek 为什么敢在硬件涨价时降价?
DeepSeek 的降价并不只是短期补贴带来的市场策略。
更重要的基础,是推理成本结构发生了变化。
V4 系列据称采用自研稀疏注意力机制与混合专家架构。在处理百万级 Token 长上下文时,其算力消耗明显低于上一代,KV Cache 占用也得到显著压缩。
与此同时,V4 已经完成与华为昇腾硬件的深度适配,使模型在部署时拥有更多基础设施选择,减少对单一海外高端加速器体系的依赖。
多项成本优势由此叠加:
每次请求只激活部分专家参数;
长上下文注意力计算量下降;
KV Cache 占用减少;
国产硬件适配能力提升;
推理集群利用率提高;
调用规模扩大后,固定成本进一步摊薄。
这也是永久降价与补贴价格战之间的根本区别。
补贴可以让 API 暂时便宜几个月。
架构效率提升,则可能让模型长期保持更低价格。
当模型厂商开始竞争“每一美元能够产出多少智能”,Token 价格就不再只是商业部门决定的数字,也会直接反映底层工程效率。
缓存:把重复内容变成最便宜的部分
永久降价 75%,只是看得见的明牌。
对企业而言,缓存可能才是影响成本更深的一张牌。
生产环境中的大量 AI 请求,本身具有很高的重复性。
客服系统会反复发送相同的服务政策;代码智能体会重复读取仓库规范;文档助手会多次加载相同的模板、品牌要求和合规规则。
没有缓存时,这些内容每次都需要重新处理,并按照标准输入价格计费。
有了缓存之后,重复内容只需要完成一次高成本计算,后续读取价格便可以大幅下降。
OpenRouter 于 4 月 30 日推出 Response Caching,也就是响应缓存。
当一项完全相同的 API 请求再次命中缓存时,平台可以直接返回此前保存的结果,不必重新执行模型推理,也不会产生正常的 Token 生成费用。
缓存还能明显降低延迟。
原本需要数秒才能完成的推理,在命中响应缓存后,可能只需要数十到数百毫秒;缓存查找本身甚至可以压缩到几毫秒。
OpenRouter 还支持 Prompt Caching,也就是提示缓存。
在这种模式下,重复提示内容并非完全免费,但读取价格可以远低于标准输入价格。
例如,某模型的正常输入价格为每百万 Token 3 美元,缓存读取可能只需要 0.30 美元,成本正好下降 90%。
DeepSeek、阿里通义千问等模型生态,也为支持的工作负载提供了较低的缓存读取价格。
当一段上下文需要重复使用数万次、数百万次时,这种差异会直接改变企业 AI 应用的成本结构。
响应缓存与提示缓存,解决的是两类问题
虽然两种缓存机制都能降低成本,但它们适合的场景并不相同。
响应缓存:整个请求完全相同
当提示词、参数与上下文全部相同时,平台可以直接返回此前保存的完整答案,无需再次调用模型。
适合场景包括:
高频常见问题;
重复分类请求;
固定商品介绍;
标准政策说明;
确定性较强的信息查询;
开发和测试环境中的重复调用。
在理想情况下,缓存命中后的模型推理成本可以接近于零。
提示缓存:大部分上下文相同
当系统提示词、知识库材料、代码库说明或企业手册保持不变,只有用户问题发生变化时,可以重复使用已缓存的上下文。
适合场景包括:
企业知识库助手;
长上下文编程智能体;
合同审查;
批量文档分析;
客服系统;
多步骤 Agent 工作流。
提示缓存不会让整项请求完全免费,但可以显著降低重复处理长上下文所产生的费用。
两种机制结合后,企业无需再为模型已经读取过的相同内容反复支付全价。
“缓存命中低至 1/50”,是否代表总成本降低 98%?
不一定。
极低的缓存命中价格,只适用于成功复用的那部分输入内容。
新问题、未缓存材料、模型输出、工具调用和应用基础设施仍然会产生费用。
因此,实际节省多少,取决于业务的缓存命中率。
如果请求内容每次都完全不同,缓存价值可能相对有限。
如果一项任务中有 80% 甚至 90% 的内容长期保持不变,成本降幅就会更加明显。
假设一家企业的智能体每次都需要读取一份 10 万 Token 的操作手册,而员工的新问题只有 1,000 Token。
没有缓存时,每一次请求都需要重新处理约 101,000 Token。
使用提示缓存后,10 万 Token 的操作手册可以按照大幅折扣后的读取价格重复使用,只有新增问题和模型输出按照常规价格计算。
当调用次数达到数万甚至数百万次时,这部分节省可能远大于模型挂牌价下降 75% 所带来的收益。
因此,更准确的表述应该是:
缓存不会让所有 AI 请求都趋近于免费,但会让重复上下文变得非常便宜。
对于高重复率业务,当低价模型、高缓存命中、批量执行和智能路由同时发挥作用时,端到端成本下降 90% 以上是可能出现的结果。
OpenRouter:让现有低价更容易被调用
DeepSeek 降低的是模型供应价格。
OpenRouter 则让开发者能够通过更加统一的入口,调用、比较、路由并缓存不同供应商的模型服务。
它所解决的主要问题包括:
使用一套兼容接口调用多个模型;
在一个平台中查看不同供应商用量;
根据价格和可用性完成路由;
在供应商失败时切换节点;
使用提示缓存和响应缓存;
减少多套 SDK 与分散账单的维护成本。
不过,底层模型价格通常仍然由供应商提前公布。
开发者可以选择更便宜的供应节点,但价格本身仍然是一张供应商价目表,而不是持续由市场供需形成的实时成交价。
因此,OpenRouter 优化的是:
怎样更高效地购买已经存在的模型服务?
Unexhub 试图解决的,则是另一个问题:
怎样让模型算力本身成为可以动态定价与快速结算的市场资源?
Unexhub:让 Token 从计价单位走向市场资源
Unexhub 正在建设 Token 聚合与分发的算力交易中心,用于连接主流模型供应商、专业 AI 服务商、企业用户和开发者。
平台计划聚合 Gemini、ChatGPT、Claude、DeepSeek、GLM 等海内外主流模型生态,同时也为个人开发者、中小型 AI 企业,以及拥有垂直模型或闲置推理能力的供应商提供分发渠道。
它的方向并不止于“用一个接口调用多个模型”。
Unexhub 希望建立一层更加市场化的基础设施,让 Token 算力可以被:
从多家供应商统一聚合;
根据市场条件参与竞价;
按照价格与性能动态路由;
根据真实消耗精确计量;
在请求完成后更快结算收益。
在这样的体系中,Token 不再只是模型账单上的一个数字。
它还会成为衡量算力供给、分配调用流量、计算供应商收益和完成交易结算的基础单位。
更快结算:改善供应商资金周转效率
传统 API 分销往往存在较长账期。
中小型供应商今天已经完成了大量模型调用,却可能需要等待数周才能收到对应款项。
但在这段时间里,GPU、云服务、带宽、存储和运维费用仍然需要持续支付。
对缺少大规模资金储备的中小供应商而言,较长结算周期会形成明显的现金流压力。
Unexhub 希望缩短这一周期。
当请求成功完成、Token 用量得到确认后,平台便可以计算供应商应获得的收益,并更快推进结算。
这可以带来多项实际权益:
更快回收基础设施支出;
提高现金流可见性;
缓解营运资金压力;
获得更多用于扩充算力的周转资金;
提高供应商释放闲置资源的意愿。
早期入驻的供应商,还有机会获得平台最低 0 佣金的阶段性权益。
对于中小型模型服务商而言,这意味着可以用更低的渠道成本进入更广泛的 Token 分销网络,而无需独立建设完整的获客、跨境计费、流量路由和资金清算体系。
这项机制并不能让供应商的所有资金压力瞬间消失。
它真正改善的,是从完成推理服务到获得实际收入之间的周转效率。
API 竞价:让价格随市场供需浮动
Unexhub 更具想象空间的一项设计,是 API 竞价机制。
不同模型和供应商不必全天维持完全固定的价格,而可以根据当前可用算力和调用需求调整报价。
这套逻辑类似电力现货市场。
需求旺盛、供给紧张时,电价会上升;需求下降、资源过剩时,价格则会回落。
AI 推理同样具有明显的周期。
工作日白天往往是企业调用高峰,深夜和周末则可能存在大量闲置算力。
市场化定价可以同时为两端创造价值。
对开发者而言:
非紧急批量任务可以安排在低价时段;
请求可以自动匹配成本更低的供应商;
按实际 Token 消耗付费,减少预充值闲置;
可以根据任务灵活平衡价格、速度与质量;
部分模型服务最低可至市场常规价格的约 3 折。
对供应商而言:
需求高峰时可以获得更高收益;
低谷时段可以通过优惠报价变现闲置算力;
平台能够根据真实需求分配流量;
收益按照实际完成的 Token 调用计算;
中小供应商可以凭价格、性能与专业能力参与竞争。
这种机制并不意味着未来所有 Token 都会变成同一个价格。
模型能力、上下文窗口、延迟、稳定性、节点地区、安全和合规要求,仍然会形成不同价格层级。
真正改变的是价格发现方式。
价格不再只由供应商长期固定挂牌,也可以受到当前市场需求与可用资源数量的共同影响。
毫秒级路由,让市场竞价真正可用
竞价市场只有在用户不需要手动参与每次比较时,才具有真正价值。
没有开发团队愿意在发送每一个 API 请求之前,先逐家查看几十个供应商的报价。
因此,Unexhub 将供应商竞价与自动路由结合起来。
平台可以综合评估:
当前价格;
模型能力;
响应延迟;
处理吞吐;
历史稳定性;
地区可用性;
合规需求;
企业 SLA 要求。
随后,在毫秒级时间内,将请求匹配到更合适的供应商。
开发者仍然通过一个入口发送请求,并按照实际 Token 用量付费。
供应商比较、流量分配、用量计量和收益结算,都由平台在后台完成。
对于高频调用企业而言,这项能力不只是更加方便。
它可以让日常分类、提取、批量生成和预处理任务优先使用更具成本优势的资源,而把高性能模型保留给复杂推理与关键业务。
部分服务最低约 3 折的价格权益,也会随着调用规模增长体现出更高价值。
一次请求节省几厘钱可能并不起眼。
但当企业每月需要处理数百万次文档提取、客服对话、内容分类和 Agent 步骤时,这些节省足以直接改变产品毛利率。
从“哪个 Token 更便宜”,转向“哪个任务成本更低”
AI 行业习惯使用“每百万 Token 价格”比较模型。
这个指标重要,却并不完整。
企业购买 Token,并不是为了拥有 Token 本身,而是为了获得业务结果:
解决一个客户问题;
审核一份文件;
分析一份合同;
生成一份报告;
完成一项编程任务;
转录一段录音;
筛选一个销售线索。
价格较低的模型,可能输出更长、失败率更高,或者需要多次重试。
价格更高的模型,也可能一次就完成任务。
因此,更有价值的指标应该是:
每项成功任务的综合成本。
DeepSeek 的低价降低了起始成本。
缓存减少了重复上下文的费用。
OpenRouter 帮助选择并管理不同供应商路径。
Unexhub 则进一步引入竞价分发与市场化定价。
当这些机制结合起来,企业可以通过多种方式降低单位任务成本:
根据任务难度选择合适模型;
缓存可重复使用的提示与响应;
将日常流量路由到更经济的供应商;
在低谷时段执行非紧急任务;
避免为重复上下文反复支付全价;
减少失败请求和无效重试;
按真实用量付费,而不是购买闲置套餐。
这才是 Token 降本更深层的含义。
它不是一次永久不变的折扣,而是一套持续寻找更高效交付路径的机制。
Token 支出指数:市场正在重新选择
模型降价与路由竞争的影响,已经开始反映在部分市场指标中。
衡量每百万 LLM Token 平均支付价格的相关支出指数,曾在设立后快速上涨,随后又从 5 月高点出现明显回落。
部分金融机构也预测,未来一年 Token 平均价格仍存在进一步下降空间。
与此同时,模型聚合和 AI 网关数据显示,开发者正在把更多流量转向低成本模型与开放权重模型。
DeepSeek 系列在部分聚合平台上获得较高调用量,开放权重模型在 AI Gateway 流量中的占比也明显上升。
市场行为给出了非常直接的答案:
只要模型能力足够、接入方便、价格合理,开发者就会把流量迁移过去。
当不同模型可以通过同一套兼容接口切换时,品牌忠诚度的重要性会下降。
开发者会用每一次请求,为更适合自己的模型和供应商投票。
三个层面,同一场成本革命
DeepSeek、OpenRouter 与 Unexhub,分别代表了同一场变化中的三个层面。
DeepSeek 改变供给成本
通过模型架构和推理优化,让供应商能够主动降低公开价格。
OpenRouter 改变调用与复用方式
通过统一网关、多供应商路由和缓存机制,减少模型接入与重复处理成本。
Unexhub 改变分发与结算机制
通过供应商竞价、动态路由、更快结算与市场化定价,让 Token 算力逐渐成为流动性更高的市场资源。
一个让模型更便宜。
一个让重复调用更便宜。
一个试图让整个市场运行得更高效。
三条路径最终汇聚到同一个方向:
未来 AI 的使用成本,不仅取决于选择了哪个模型,还取决于:
请求最终由哪家供应商执行;
上下文是否命中缓存;
任务在什么时段运行;
当前市场有多少可用算力;
平台如何完成路由;
供应商收益能否快速结算。
说到底
DeepSeek 永久降价 75%,在原有 Token 定价体系上撕开了一道明显缺口。
OpenRouter 的缓存机制,让重复提示词和相同请求的边际成本逐渐接近于零。
Unexhub 则指向下一阶段:多家供应商的 Token 算力被统一聚合、参与竞价、通过毫秒级路由完成分发,并根据真实调用量结算收益。
对于开发者与企业用户而言,这意味着:
统一接入海内外主流模型;
按实际 Token 用量付费;
减少资金沉淀在未使用的预付套餐中;
通过缓存降低重复任务成本;
在价格、速度和模型能力之间动态选择;
部分模型服务最低可至市场常规价格的约 3 折。
对供应商而言,则意味着:
一键进入更广泛的分发渠道;
根据自身能力获得自动化流量分配;
按真实调用量获得收益;
更快完成收入结算;
早期最低 0 佣金的阶段性激励;
将闲置或未充分利用的推理能力转化为收入。
因此,真正的成本革命,已经不再只是寻找“哪一家模型最便宜”。
而是回答三个更实际的问题:
算力在什么时间购买更划算?
当前请求应该交给哪家供应商?
如何用更少的付费 Token,完成同样的业务结果?
Token 正在跨越文本、图像、语音、文档与 Agent 工作流,逐渐从技术计量单位,演变为价格、路由、结算和资源分配的基础单位。
这才是下一阶段 Token 交易市场真正的降本逻辑。