Token交易市场隐形霸主:OpenRouter 调度 450 万亿 Token,模型供应商沦为"被路由的商品"

On July 18, only two days after its release, Kimi K3 entered OpenRouter’s daily top ten, processing approximately 142 billion Tokens in a single day.

450 Trillion Tokens—and the Routing Power Rewriting AI Pricing

English–Chinese Bilingual Industry Analysis|English First, 中文完整版见后半部分


English Version

On July 18, only two days after its release, Kimi K3 entered OpenRouter’s daily top ten, processing approximately 142 billion Tokens in a single day.

A newly released model had moved from launch announcement to production-scale traffic in just 48 hours.

This was not an isolated event.

According to an OpenRouter analysis covering approximately 450 trillion Tokens processed between January and June 2026, DeepSeek’s share rose from around 9% to 18%. By early June, Chinese models had reportedly surpassed American models in Token usage on the platform.

The contrast with the recent past is striking.

At the end of 2024, open-weight models accounted for less than 1.2% of weekly Token consumption on OpenRouter.

Less than two years later, hundreds of trillions of Tokens were being allocated across models through a routing layer that could shift traffic with a single configuration change.

450 trillion Tokens are not merely platform traffic.

They represent the power to decide:

  • Which model receives real production workloads;

  • Which supplier earns API revenue;

  • Which pricing strategy gains market share;

  • Which model becomes visible to developers;

  • Which provider gradually disappears from the routing table.

The AI market is moving from an era of choosing models to an era of routing workloads.

And the company controlling the routing layer may have more influence over pricing than the company that trained the model.


From “Which Model Should We Use?” to “Where Should This Request Go?”

OpenRouter’s basic business model is not difficult to understand.

It does not train its own frontier model.

Instead, it provides a unified API gateway connecting developers to more than 400 models and approximately 70 providers.

A developer can switch among different models by changing an endpoint or modifying a model name.

Technically, the change may require only one line of configuration.

Commercially, however, this simple abstraction changes the balance of power across the entire model market.

Under the traditional provider-led model, developers had to establish separate relationships with each company.

To use Claude, they opened an Anthropic account.

To use GPT, they registered with OpenAI.

To access DeepSeek, Gemini, Qwen, or another model, they integrated a different API.

Each provider controlled its own:

  • Authentication;

  • Billing;

  • Documentation;

  • Rate limits;

  • Error handling;

  • Account balance;

  • Enterprise agreement;

  • Regional availability.

This fragmentation created switching costs.

Even when another model was cheaper, changing providers required engineering work, testing, compliance review, and billing migration.

Those costs gave model companies a degree of customer lock-in.

OpenRouter weakened that lock-in by turning the model into a configurable route.

Today, a workload can use DeepSeek.

Tomorrow, it can move to Qwen.

Next week, it can return to Claude.

The application does not need to be rebuilt from the ground up.

The model becomes one adjustable parameter inside a wider infrastructure layer.


The Most Important Model May Be the Router

Once several hundred models are accessible through one interface, the developer no longer needs to select one permanent winner.

The routing system can choose among providers according to:

  • Price;

  • Latency;

  • Availability;

  • Reliability;

  • Context length;

  • Model capability;

  • Regional constraints;

  • Privacy requirements;

  • Historical error rates.

This makes the routing layer more than a technical convenience.

It becomes a market-allocation mechanism.

A provider may publish an attractive benchmark result, but if its route is too expensive, too slow, or too unreliable, the routing layer may send traffic elsewhere.

A model supplier can therefore lose business without the end user explicitly rejecting the model.

The user may never even see it.

The route was simply not selected.

That is a deeper form of pricing pressure than an ordinary price comparison.

A high-priced model does not merely risk losing after the customer evaluates it.

It may never receive the request in the first place.


Price determines visibility

OpenRouter’s routing logic has been described as favoring lower-cost and more reliable providers, with price playing an important role in weighted allocation.

The exact routing method may vary according to user settings, provider availability, and platform updates, but the economic principle is clear:

The more expensive the route, the harder it becomes to win high-volume traffic unless it provides measurable additional value.

This means pricing no longer affects only revenue per request.

It affects whether a supplier receives requests at all.

A model may be technically excellent, but if another route is:

  • Cheaper;

  • Reliable enough;

  • Fast enough;

  • Compatible with the workload;

the cheaper supplier may receive most of the traffic.

The market is no longer asking only:

Is this model good?

It is asking:

Is this model sufficiently better to justify its routing cost?

That is a much more demanding test.


Suppliers Are Becoming Routable Commodities

When more than 400 models are placed inside the same routing system, model services become increasingly substitutable.

This does not mean every model is identical.

Models still differ in reasoning, coding, context, safety, speed, and reliability.

But once developers can replace one with another through a configuration change, those differences are evaluated inside a shared comparison framework.

The model is no longer a self-contained product ecosystem.

It becomes a unit of compute supply competing for the same request.

Consider an open-weight model such as Llama 3.3 70B.

The underlying model weights may be the same, yet different infrastructure providers can offer access at dramatically different prices.

One provider may charge around $0.10 per million Tokens, while another may charge more than $1.

The model brand has not changed.

What changes is:

  • Hardware efficiency;

  • Quantization;

  • Region;

  • Capacity;

  • Latency;

  • Provider margin;

  • Service quality.

This reveals an important shift.

The value is no longer determined solely by the model creator.

It is also determined by whoever can serve the model most efficiently.

The brand premium at the model layer begins to dissolve into competition at the infrastructure layer.


The same weights, different economics

When multiple providers host the same open model, developers begin comparing execution rather than origin.

They ask:

  • Which provider is cheaper?

  • Which one responds faster?

  • Which one fails less often?

  • Which one has available capacity?

  • Which one supports the required region?

  • Which one offers a better SLA?

This moves pricing power away from the company that released the model and toward the marketplace that compares providers.

A model supplier that once controlled its distribution channel may now compete against several implementations of the same model.

The product is no longer simply “Llama” or “DeepSeek.”

It is:

This particular route to this model, from this supplier, at this price and service level.

That is the language of a commodity market.


The Routing Layer Does Not Care Which Brand Wins

The revenue logic of a routing platform differs fundamentally from that of a model provider.

A model company wants its own model to receive more traffic.

A routing platform wants more total traffic to pass through its infrastructure.

It earns whether the request is served by:

  • Claude;

  • DeepSeek;

  • Gemini;

  • Qwen;

  • Llama;

  • Another model entirely.

OpenRouter’s platform usage reportedly grew from approximately $800,000 in monthly consumption in October 2024 to around $8 million by May 2025.

The platform has also processed weekly Token volumes in the tens of trillions.

Its fee model is based on activity flowing through the gateway.

That creates a neutral commercial incentive:

The routing platform does not need one model to dominate. It benefits when developers continuously compare, switch, and consume.

This neutrality is powerful.

The platform can prioritize whichever supply is most competitive without protecting a model brand.

For suppliers, this creates a new reality:

The route owns the customer relationship. The model provider competes for allocation.


From Pricing Authority to Traffic Dependency

In the old market structure, the supplier controlled:

  • The API endpoint;

  • The customer account;

  • The invoice;

  • The price;

  • The documentation;

  • The usage relationship.

In the routed market, much of that relationship moves upward to the gateway.

The developer sees one interface.

The routing layer sees the entire demand pool.

The supplier receives only the traffic allocated to it.

This transforms the supplier from a direct seller into a capacity provider.

Its revenue depends on:

  • Route ranking;

  • Market price;

  • Availability;

  • Performance;

  • Routing policy;

  • Platform demand.

Even a powerful model can become commercially invisible if the gateway does not send it sufficient traffic.

An API URL with no routed demand is not a business.

It is unused capacity.


Kimi K3 Shows How Fast Routing Can Create a Market

Kimi K3’s rapid rise illustrates the positive side of this system.

Under the traditional distribution model, a newly released model might need months to:

  • Build developer awareness;

  • Publish SDKs;

  • Negotiate integrations;

  • Complete procurement reviews;

  • Attract enterprise customers;

  • Generate meaningful API volume.

Through an established routing platform, that process can be compressed dramatically.

A new model enters the catalog.

Developers test it.

Routers compare price and performance.

Production traffic arrives.

Within 48 hours, the model may be handling more than one hundred billion Tokens per day.

This reduces the distribution advantage historically enjoyed by the largest providers.

A new supplier no longer needs to build every part of the market-access stack independently.

The gateway gives it immediate access to existing demand.

But the same mechanism that accelerates entry also accelerates replacement.

A model can rise in two days.

It can also lose traffic just as quickly when a cheaper or stronger alternative appears.


The New Competition Is for Routing Priority

In a routed market, model providers compete for more than users.

They compete for a position in the allocation algorithm.

That position may depend on:

  • Lower price;

  • Better uptime;

  • Faster responses;

  • Higher success rates;

  • Larger context;

  • More reliable tool use;

  • Greater regional coverage;

  • Stronger compliance support.

This creates a continuous ranking system.

The supplier cannot rely on a one-time launch campaign.

It must remain competitive every day.

A provider that increases prices may lose routing weight.

A provider that suffers repeated errors may be deprioritized.

A newly optimized supplier may suddenly receive more traffic.

The pricing page is no longer the end of the commercial process.

It is one input into an always-running allocation engine.


Unexhub: Turning Routing Into Trading

If OpenRouter turns models into routable services, Unexhub aims to push the logic one stage further—toward a Token marketplace.

Unexhub positions itself as a global Token aggregation, distribution, and compute-trading center.

According to its current platform narrative, it aims to connect model resources associated with ecosystems such as:

  • Gemini;

  • ChatGPT;

  • Claude;

  • DeepSeek;

  • GLM;

  • Independent developers;

  • Small and medium-sized AI suppliers.

Its proposed model is built around three core components:

  1. Aggregation;

  2. Faster usage-based settlement;

  3. API bidding and dynamic routing.

The distinction is important.

OpenRouter primarily answers:

Where should the request be routed?

Unexhub’s market vision also attempts to answer:

At what price should the request be traded?

The first is allocation.

The second is price discovery.

When routing and market pricing are combined, suppliers may lose not only control over traffic distribution, but also exclusive control over price formation.


Aggregation creates the supply pool

The first requirement for any marketplace is sufficient supply.

Unexhub’s aggregation model aims to bring multiple model providers into one technical and commercial layer.

For developers, this can reduce the need to maintain:

  • Separate APIs;

  • Multiple account balances;

  • Different authentication systems;

  • Independent invoices;

  • Repeated provider integrations.

For suppliers, it can provide access to a broader pool of demand without requiring each company to build a complete sales and billing operation.

Aggregation turns isolated model capacity into a common supply pool.

Without that pool, routing remains a collection of direct integrations.

With it, capacity can be compared and allocated more dynamically.


Faster settlement improves supplier participation

Routing traffic is not enough if suppliers wait months to receive revenue.

Smaller model and compute providers face immediate costs:

  • GPU rental;

  • Electricity;

  • Bandwidth;

  • Engineering;

  • Monitoring;

  • Capacity maintenance.

Customer revenue may arrive much later.

Unexhub’s proposed usage-based settlement mechanism aims to shorten the interval between API consumption and supplier payment.

The practical value may include:

  • Faster cash recovery;

  • Lower receivables pressure;

  • Better utilization of idle capacity;

  • Reduced billing infrastructure costs;

  • Easier access to developer demand.

It is more realistic to say that faster settlement can significantly reduce financial pressure, rather than reduce it to zero.

But for a smaller supplier, a shorter cash-conversion cycle can determine whether it can continue serving traffic.

Routing decides who receives demand.

Settlement decides who can afford to keep supplying it.


API bidding creates a market price

Under the traditional model, each supplier publishes an official price.

Under a bidding system, suppliers can compete to serve demand under changing market conditions.

When capacity is abundant, prices may fall.

When demand rises, available supply may command a premium.

Different service conditions can also create different prices:

  • Low-priority batch processing;

  • Real-time interactive requests;

  • High-context workloads;

  • Regulated enterprise traffic;

  • Reserved low-latency capacity.

This gives Tokens more than a provider price.

They gain a transaction price determined by:

  • Supply;

  • Demand;

  • Time;

  • Service quality;

  • Urgency;

  • Routing requirements.

The provider can still set a minimum acceptable rate.

But it no longer determines the entire market alone.


Routing Plus Trading Changes the Pricing Formula

OpenRouter-style routing compares routes.

Unexhub-style trading aims to let those routes compete economically.

When the two concepts are combined, the pricing formula becomes more complex.

The final cost of a request may depend on:

  • Which model is capable of completing it;

  • Which suppliers currently have capacity;

  • Which region is acceptable;

  • How quickly the result is needed;

  • Whether the workload can be delayed;

  • Which provider offers the best bid;

  • Whether the route satisfies compliance requirements.

This means the official model price becomes only one reference point.

The actual market price is produced at the moment of allocation.

The model company may have trained the intelligence.

The routing and trading layer decides how that intelligence is monetized.


What This Means for Developers

For developers, the shift from direct model selection to routing and trading creates several advantages.

Lower integration costs

One API can provide access to several model ecosystems.

Developers spend less time maintaining separate integrations.

More flexible model choice

The application can switch among models without a full rewrite.

Better cost control

Routine tasks can use low-cost models, while complex tasks escalate to stronger ones.

Greater resilience

If one provider becomes unavailable, traffic can move to another route.

More transparent comparison

Price, latency, reliability, and quality can be evaluated inside one framework.

The developer no longer needs to choose between permanently committing to one premium model and permanently committing to one cheap model.

The system can use both.


What “As Low as 30% of Market Price” Means

According to Unexhub’s current positioning, eligible model routes may be available at prices as low as approximately 30% of prevailing market or official rates.

The word eligible matters.

Actual pricing may depend on:

  • Model availability;

  • Route conditions;

  • Region;

  • Service level;

  • Market supply;

  • Usage volume;

  • Platform rules.

The value is not limited to one discounted API request.

For a high-volume product, lower effective Token costs can mean:

  • More calls within the same budget;

  • Greater room for long-context processing;

  • Lower model-evaluation costs;

  • More affordable agent workflows;

  • Additional capacity during traffic peaks;

  • Lower financial risk during product validation.

Combined with dynamic routing, the benefit becomes larger.

The platform can use lower-cost resources for routine execution while preserving premium models for the requests that genuinely need them.

The result is not simply a cheaper model.

It is a lower total cost per completed workflow.


What This Means for Suppliers

For model suppliers, the new market creates both opportunity and pressure.

The opportunity

A smaller supplier can reach developers without building:

  • A large sales team;

  • A full billing system;

  • An independent distribution network;

  • Multiple regional integrations.

It can monetize idle compute and receive demand through the shared market.

The pressure

The same supplier must compete transparently on:

  • Price;

  • Availability;

  • Latency;

  • Reliability;

  • Model quality.

Brand and private pricing can no longer hide weak economics.

Aggregation opens the market.

It also makes competition more direct.


Early suppliers may gain a distribution advantage

Unexhub’s platform narrative also emphasizes simplified supplier onboarding and early-stage incentives.

For suppliers entering early, potential advantages may include:

  • Lower initial distribution costs;

  • Earlier access to developer traffic;

  • More usage history;

  • Better understanding of demand;

  • Faster settlement;

  • A stronger position before supply becomes crowded.

However, early entry creates value only if the market develops:

  • Stable demand;

  • Sufficient liquidity;

  • Transparent measurement;

  • Reliable routing;

  • Auditable settlement;

  • Clear compliance rules.

Low commission or promotional access can encourage participation.

It cannot substitute for a functioning market.


A Necessary Clarification About Unexhub

Unexhub’s detailed operating model—including aggregation, rapid settlement, API bidding, millisecond-level routing, supplier coverage, and promotional pricing—should be understood according to the platform’s own positioning and latest official information.

Independent authoritative reporting verifying every specific operational claim remains limited.

Search results involving similarly named financial or trading products may also refer to unrelated platforms and should not be treated as evidence about Unexhub.

However, the broader industry direction is real.

Token aggregation, compute scheduling, model routing, and usage-based settlement are moving from concept toward infrastructure experimentation.

Specific discounts, commissions, routing guarantees, and settlement schedules should therefore be confirmed through Unexhub’s current official terms.


Routing Does Not Make Every Model Identical

The rise of routing and trading does not mean model quality no longer matters.

Different models remain unequal in:

  • Reasoning;

  • Coding;

  • Safety;

  • Context capacity;

  • Multilingual performance;

  • Tool use;

  • Output reliability;

  • Compliance support.

The cheapest route may create higher total costs if it produces:

  • More failures;

  • More retries;

  • Longer outputs;

  • Additional human correction;

  • Greater operational risk.

The correct metric is not simply:

Price per million Tokens.

It is:

Cost per successfully completed task.

A mature routing market should therefore optimize for more than the lowest quote.

It should balance:

  • Price;

  • Quality;

  • Latency;

  • Reliability;

  • Risk.

The best route is not always the cheapest model.

It is the lowest-cost route that can complete the task correctly.


450 Trillion Tokens Represent Control Over Market Discovery

The real power of 450 trillion routed Tokens is not simply their size.

It is the information created by that flow.

The routing layer can observe:

  • Which models developers test;

  • Which models they abandon;

  • How sensitive demand is to price;

  • Which providers fail under load;

  • Which tasks require stronger capability;

  • Where regional capacity is scarce;

  • How quickly traffic moves after a price change.

This information gives the gateway an advantage over both buyers and suppliers.

It can see the market forming in real time.

The model provider sees its own traffic.

The routing layer sees everyone’s traffic.

That makes the routing layer not only a distributor, but also a price-discovery engine.


The Era of One Model Conquering Everything Is Ending

OpenRouter CEO Alex Atallah has argued that the era of selecting one model to conquer the world is over.

The deeper conclusion may be even more consequential:

The era in which one model provider controls the entire customer relationship is also ending.

Enterprises will increasingly use:

  • One model for classification;

  • Another for reasoning;

  • Another for coding;

  • Another for regulated data;

  • Another for low-cost batch processing.

The model is becoming modular.

The route becomes strategic.

The provider that wins one request may lose the next.

The platform deciding between them becomes the new center of power.


Final Conclusion

Kimi K3’s rise into OpenRouter’s top ten within 48 hours shows how quickly a routing layer can create production-scale demand.

OpenRouter’s analysis of approximately 450 trillion Tokens shows something larger:

The most important power in the next AI market may not be the power to train models.

It may be the power to allocate traffic among them.

Once developers can switch models by changing one parameter:

  • Brand premiums weaken;

  • Switching costs fall;

  • Price competition accelerates;

  • New models can scale faster;

  • Existing suppliers can be replaced overnight.

OpenRouter turns models into routable services.

Unexhub aims to take the next step by turning routable Token capacity into a dynamically priced market.

According to Unexhub’s current positioning, the platform seeks to provide:

  • Unified access to Gemini, ChatGPT, Claude, DeepSeek, GLM, and other model resources;

  • Distribution channels for independent and smaller AI suppliers;

  • Pay-per-Token usage;

  • Eligible routes priced as low as approximately 30% of market or official rates;

  • Millisecond-level request routing;

  • Faster usage-based supplier settlement;

  • API bidding based on changing supply and demand.

These advantages will matter only if they are supported by:

  • Stable capacity;

  • Transparent Token measurement;

  • Verifiable service quality;

  • Auditable settlement;

  • Credible compliance and governance.

The old AI market asked:

Which model is the best?

The routed market asks:

Which model should receive this request?

The trading market goes one step further:

At what price should that request be served?

When those questions are answered by the routing and market layer, model providers no longer control the entire pricing process.

The true meaning of 450 trillion Tokens is not that a platform processed a large amount of traffic.

It is that the routing layer gained the power to decide who receives that traffic—and, increasingly, what that traffic is worth.


中文版

450 万亿 Token 的调度权,正在重写 AI 的定价逻辑

7 月 18 日,发布仅两天后,Kimi K3 便进入 OpenRouter 日榜前十,单日调用量约为 1420 亿 Token。

一个新模型从正式发布到获得生产级流量,只用了 48 小时。

这并不是孤例。

根据 OpenRouter 对 2026 年 1 月至 6 月约 450 万亿 Token 的分析,DeepSeek 的平台份额从约 9% 上升至 18%。到 6 月初,中国模型的 Token 使用占比据称已经超过美国模型。

与不久前相比,这种变化十分惊人。

2024 年底,OpenRouter 上开放权重模型每周消耗的 Token 占比还不足 1.2%。

不到两年后,数百万亿 Token 已经通过统一路由层,在不同模型和供应商之间被重新分配。

450 万亿 Token,不只是一个平台的流量。

它代表着一种决定权:

  • 哪个模型能够获得真实生产请求;

  • 哪家供应商能够获得 API 收入;

  • 哪种价格策略能够赢得市场;

  • 哪个模型能够被开发者看见;

  • 哪个供应商会逐渐从路由表中消失。

AI 市场正在从“选择一个模型”,走向“调度每一条请求”。

而掌握路由层的平台,可能比训练模型的公司更能影响最终价格。


从“选哪个模型”,走向“这条请求该去哪里”

OpenRouter 的商业模式并不复杂。

它不训练自己的前沿模型,而是提供统一 API 网关,通过一个接口连接 400 多个模型和约 70 家供应商。

开发者想切换模型,通常只需更改接口地址,或者修改一行模型名称。

从技术操作看,这只是一次很小的调整。

但从商业关系看,这种抽象正在重新分配整个模型市场的权力。

传统模式下,开发者必须分别与每家供应商建立关系。

想用 Claude,需要注册 Anthropic。

想用 GPT,需要开通 OpenAI。

想调用 DeepSeek、Gemini、Qwen 或其他模型,又要重新接入另一套 API。

每家供应商都拥有自己的:

  • 身份认证;

  • 计费体系;

  • 开发文档;

  • 限速规则;

  • 错误处理;

  • 账户余额;

  • 企业合同;

  • 区域服务范围。

这些碎片化体系形成了切换成本。

即使另一家模型更便宜,企业也需要重新开发、测试、审核合规并迁移账单。

这种成本让模型厂商拥有一定程度的用户锁定能力。

OpenRouter 拆掉了这堵墙。

模型从一整套独立生态,变成路由表中的一个配置参数。

今天可以使用 DeepSeek。

明天可以切换 Qwen。

下周也可以重新换回 Claude。

应用不需要推倒重建。

底层模型只是可以随时调整的一条路线。


最重要的模型,可能是路由器

当数百个模型都能通过一个接口访问时,开发者不再需要永久选择一个赢家。

路由系统可以综合以下因素,在不同供应商之间分配请求:

  • 价格;

  • 时延;

  • 可用性;

  • 稳定性;

  • 上下文长度;

  • 模型能力;

  • 区域限制;

  • 隐私要求;

  • 历史错误率。

路由层因此不再只是技术便利工具。

它正在变成市场资源的分配机制。

一家供应商可以拥有漂亮的评测成绩,但如果调用路线太贵、太慢或不够稳定,路由系统就可能把流量交给别人。

模型供应商甚至不需要被用户主动拒绝,就可能失去业务。

用户可能从未看到它。

因为路由层根本没有选择这条路线。

这比普通的价格比较更残酷。

高价模型面对的不只是用户评估后选择其他产品。

而是它可能根本没有获得被评估的机会。


价格决定的不只是收入,还有可见度

OpenRouter 的路由机制被描述为更偏向价格较低、稳定性较高的供应商,价格会影响流量的加权分配。

具体算法可能随着用户设置、供应商状态和平台规则更新而发生变化,但其经济逻辑十分清晰:

一条路线越贵,就越难获得大规模流量,除非它能证明自己提供了可量化的额外价值。

这意味着,价格不再只决定每条请求能赚多少钱。

它还会决定供应商能不能拿到请求。

一个模型可能技术能力出色。

但如果另一个供应商能够做到:

  • 价格更低;

  • 稳定性足够;

  • 响应速度足够;

  • 能完成当前任务;

低价路线就可能获得大部分调用量。

市场不再只问:

这个模型好不好?

而是进一步追问:

它是否好到足以证明更高的路由成本?

这个标准比单纯的模型评测更加严格。


供应商正在变成“可被路由的商品”

当 400 多个模型被放进同一套路由系统,模型服务就会变得越来越容易替换。

这并不意味着所有模型都完全相同。

它们依然在推理、编程、上下文、安全和响应速度等方面存在差异。

但当开发者只需修改一个参数,就可以把一种模型替换为另一种模型时,这些差异就会被放入同一个比较框架。

模型不再是一整套无法拆分的产品生态。

而是争夺同一笔请求的算力供给单位。

以 Llama 3.3 70B 这样的开放权重模型为例。

底层模型权重完全相同,但不同基础设施供应商在 OpenRouter 上提供的价格可能从每百万 Token 约 0.10 美元,到超过 1 美元不等。

模型品牌没有变化。

真正变化的是:

  • 硬件效率;

  • 量化方案;

  • 服务地区;

  • 可用容量;

  • 响应时延;

  • 供应商利润;

  • 服务质量。

这揭示了一个重要转向。

模型的价值不再只由模型开发者决定。

还取决于谁能以更高效率提供推理服务。

模型层的品牌溢价,正在被基础设施层的价格竞争逐渐拆解。


同一份权重,不同的商业价值

当多个供应商托管同一款开放模型时,开发者比较的重点会从“模型来自哪里”,转向“谁执行得更好”。

他们会直接比较:

  • 谁的价格更低?

  • 谁的响应更快?

  • 谁的失败率更低?

  • 谁有足够可用容量?

  • 谁支持需要的地区?

  • 谁能提供更可靠的 SLA?

定价权因此开始从发布模型的公司,转移到比较和分配供应商的市场平台。

一家模型供应商过去可能完全控制自己的分销渠道。

现在,它可能需要与同一模型的多个托管版本竞争。

产品不再只是“Llama”或“DeepSeek”。

而是:

某家供应商,在某个地区,以某个价格和服务水平提供的具体路线。

这已经是大宗商品市场的语言。


路由层并不在乎哪个品牌赢

路由平台与模型供应商的收入逻辑存在根本差异。

模型公司希望自己的模型获得更多调用。

路由平台则希望更多 Token 经过自己的网关。

无论请求最终由以下哪种模型处理:

  • Claude;

  • DeepSeek;

  • Gemini;

  • Qwen;

  • Llama;

  • 其他模型;

只要流量继续增长,平台就能从中获得收益。

OpenRouter 的平台月度消耗金额,据称从 2024 年 10 月约 80 万美元,增长到 2025 年 5 月约 800 万美元。

平台每周处理的 Token 规模也已经达到数十万亿级别。

其收入来自经过统一网关的调用活动,而不是押注某一个模型品牌。

这形成了一种中立的商业激励:

路由平台不需要某一家模型永远领先。它需要开发者不断比较、不断切换、不断调用。

这种中立对平台非常有利。

它可以把流量分配给当下最具竞争力的供给,而不需要保护某个模型品牌。

对供应商而言,新的现实是:

客户关系掌握在路由层手中,模型供应商只能竞争流量分配。


从定价主体,变成依赖流量分配的供给方

在旧市场结构中,供应商掌握:

  • API 入口;

  • 用户账号;

  • 账单;

  • 定价;

  • 开发文档;

  • 调用关系。

在路由市场中,这些关系中的很大一部分会向上转移到统一网关。

开发者只看到一个接口。

路由层掌握整个需求池。

模型供应商只能获得平台分配给自己的请求。

这会让供应商从直接销售者,逐渐变成算力供给方。

其收入开始依赖:

  • 路由排名;

  • 市场价格;

  • 可用容量;

  • 模型表现;

  • 路由规则;

  • 平台需求。

即使模型能力很强,如果统一网关没有为它分配足够流量,它在商业上也可能完全不可见。

没有调用量的 API,只是一串地址。

不是一门生意。


Kimi K3 展示了路由层造市场的速度

Kimi K3 的快速增长,展示了这套系统积极的一面。

在传统分销模式下,一个新模型可能需要几个月才能完成:

  • 建立开发者认知;

  • 发布 SDK;

  • 谈判接入;

  • 完成采购审核;

  • 获得企业用户;

  • 积累真实调用量。

通过成熟的路由平台,这个过程可以被大幅压缩。

模型进入平台目录。

开发者立即测试。

路由器比较价格和性能。

生产流量随即进入。

发布 48 小时后,一个新模型就可能承接每天上千亿 Token。

这削弱了头部厂商过去拥有的分销优势。

新供应商不再需要独立建设完整的市场准入体系,就有机会接触现成需求。

但加速进入市场的机制,也会加速模型被替换。

一个模型可以在两天内迅速崛起。

当更便宜或更强的替代品出现时,它也可能同样迅速失去流量。


新竞争,不只是争夺用户,而是争夺路由优先级

在路由市场中,模型供应商竞争的不只是最终用户。

还包括自己在调度算法中的位置。

这一位置可能取决于:

  • 更低价格;

  • 更高在线率;

  • 更快响应;

  • 更高任务成功率;

  • 更长上下文;

  • 更稳定的工具调用;

  • 更广泛的区域覆盖;

  • 更成熟的合规支持。

这形成了一套持续运行的排名机制。

供应商无法依赖一次成功发布会,长期保住市场位置。

它必须每天保持竞争力。

价格上涨,路由权重可能下降。

错误率升高,平台可能降低优先级。

新的优化供应商上线,又可能突然获得更多请求。

官网价目表不再是商业流程的终点。

它只是持续运行的分配引擎中的一个参数。


Unexhub:把“路由”进一步变成“交易”

如果说 OpenRouter 将模型变成了可路由的服务,那么 Unexhub 希望把这套逻辑进一步推进到 Token 交易市场。

Unexhub 将自身定位为全球 Token 聚合、分发与算力交易中心。

按照平台当前定位,其希望连接的模型和供应资源包括:

  • Gemini;

  • ChatGPT;

  • Claude;

  • DeepSeek;

  • GLM;

  • 个人开发者;

  • 中小型 AI 供应商。

其规划的核心机制包括:

  1. 聚合;

  2. 更快的按量结算;

  3. API 竞价与动态路由。

二者之间的区别非常重要。

OpenRouter 主要回答:

这条请求应该被路由到哪里?

Unexhub 的市场构想还希望回答:

这条请求应该以什么价格成交?

前者解决调度。

后者参与价格发现。

当路由与市场定价结合后,模型供应商失去的可能不只是流量分配权。

还可能包括对价格形成过程的独占控制。


聚合,建立统一供给池

任何交易市场首先需要足够丰富的供给。

Unexhub 的聚合模式,希望把多个模型供应商放入统一的技术和商业体系。

对开发者而言,可以减少维护以下系统的工作:

  • 多套 API;

  • 多个账户余额;

  • 不同身份认证;

  • 独立账单;

  • 重复供应商接入。

对供应商而言,则有机会接触更广泛的需求,而不必每家公司都独立建设完整的销售、计费和收款体系。

聚合让原本孤立的模型容量,进入共同供给池。

没有供给池,所谓路由只是多套直接接口的集合。

形成供给池后,容量才能被更动态地比较和分配。


更快结算,提高供应商的参与能力

仅仅获得流量并不够。

如果供应商需要几个月才能收到收入,中小团队仍然可能无法维持服务。

模型和算力供应商的成本是即时发生的:

  • GPU 租赁;

  • 电费;

  • 带宽;

  • 工程人员;

  • 监控系统;

  • 容量维护。

客户费用却可能很晚才到账。

Unexhub 规划中的按实际使用量结算机制,希望缩短 API 消费与供应商获得收益之间的时间。

潜在价值包括:

  • 更快回收现金;

  • 减少应收账款压力;

  • 提升闲置算力利用率;

  • 降低自建计费体系的成本;

  • 更容易接触开发者需求。

更准确地说,快速结算能够显著缓解供应商的财务压力,而不是绝对将资金压力降为零。

但对中小供应商而言,现金转换周期缩短,可能直接决定它能否继续承接平台流量。

路由决定谁能获得需求。

结算决定谁有能力继续提供供给。


API 竞价,形成市场成交价

传统模式下,每一家供应商公布自己的官方价格。

进入竞价机制后,供应商可以根据实时市场状态竞争需求。

当市场容量充足时,价格可能下降。

当需求增长时,可用供给可能获得溢价。

不同服务条件,也会形成不同价格:

  • 低优先级批量处理;

  • 实时互动请求;

  • 超长上下文任务;

  • 受监管企业流量;

  • 预留低时延容量。

这样,Token 就不再只有厂商价格。

它还拥有由以下因素共同决定的市场成交价:

  • 供给;

  • 需求;

  • 时间;

  • 服务质量;

  • 任务紧急程度;

  • 路由要求。

供应商依然可以设定自己愿意接受的最低价格。

但它不再能够独自决定整个市场的结算水平。


路由加交易,重写最终价格公式

OpenRouter 式路由负责比较不同路线。

Unexhub 式市场,希望让这些路线在经济上进一步竞争。

当两种逻辑结合时,一条请求的最终成本会由更多变量决定:

  • 哪个模型有能力完成任务;

  • 哪些供应商当前拥有容量;

  • 哪些地区可以使用;

  • 结果需要多快返回;

  • 任务能否延迟;

  • 哪家供应商报价更合适;

  • 路线是否满足合规要求。

这意味着,模型官方价格只会成为一个参考变量。

实际成交价会在请求被分配的时刻形成。

模型公司创造了智能能力。

路由与交易平台则决定这种能力如何被变现。


对开发者意味着什么?

从直接选择模型,走向路由和交易,会为开发者带来多项价值。

接入成本更低

一个 API 即可访问多个模型生态,减少重复集成。

模型选择更灵活

无需重写整个应用,就可以在不同模型之间切换。

成本控制更精细

常规任务使用低成本模型,复杂任务再升级至高能力模型。

供应韧性更强

某家供应商中断时,可以自动切换其他路线。

比较更加透明

价格、时延、稳定性和能力可以放在同一体系中评估。

开发者不再需要在“永远绑定高价模型”和“永远绑定低价模型”之间二选一。

同一套业务完全可以同时使用两者。


“最低三折”真正意味着什么?

按照 Unexhub 当前平台定位,在符合具体模型、路由和服务条件时,部分资源价格最低可接近市场或官方价格的三折。

这里最重要的是“符合条件”。

实际价格可能受到以下因素影响:

  • 模型可用性;

  • 路由条件;

  • 服务地区;

  • 服务等级;

  • 市场供给;

  • 调用规模;

  • 平台规则。

这项权益的意义,不只是某一次 API 调用更便宜。

对于高调用量产品,更低的有效 Token 成本可能意味着:

  • 相同预算支持更多请求;

  • 长上下文处理空间更大;

  • 模型测试成本更低;

  • 智能体工作流更容易长期运行;

  • 流量高峰期拥有更多容量;

  • 产品验证阶段的财务风险更小。

当它与动态路由结合时,价值会被进一步放大。

常规执行可以使用低成本资源,真正复杂的请求再交给高端模型。

最终降低的不是某个模型的单价。

而是成功完成整条工作流的总成本。


对供应商意味着什么?

对模型供应商而言,新市场既带来机会,也带来压力。

机会

中小供应商无需独立建设以下体系,也有机会接触开发者:

  • 庞大销售团队;

  • 完整计费系统;

  • 独立分销网络;

  • 多地区接入。

闲置算力也可以进入市场获得变现机会。

压力

供应商需要更加透明地竞争:

  • 价格;

  • 可用性;

  • 时延;

  • 稳定性;

  • 模型质量。

品牌与不透明定价很难继续掩盖缺乏竞争力的成本结构。

聚合平台让市场准入更公平。

也会让竞争更加直接。


早期供应商可能获得分销优势

Unexhub 的平台定位还包括简化供应商接入,并为早期参与者提供一定激励。

早期供应商可能获得:

  • 更低的初期分销成本;

  • 更早接触开发者流量;

  • 更多历史调用数据;

  • 更快理解市场需求;

  • 更短收益结算周期;

  • 在供给拥挤前占据位置。

但早期加入是否真正有价值,仍然取决于市场能否建立:

  • 稳定需求;

  • 足够流动性;

  • 透明计量;

  • 可靠路由;

  • 可审计结算;

  • 清晰合规规则。

低佣金或推广权益可以吸引供应商加入。

但无法代替一个真正有效的市场。


关于 Unexhub,需要明确的信息边界

Unexhub 所描述的聚合、快速结算、API 竞价、毫秒级路由、供应商覆盖和价格权益,应以平台官方定位与最新公告为准。

目前,针对每一项具体运营数据的独立权威报道仍较有限。

搜索中出现的同名或相似金融交易产品,也可能与 Unexhub 无关,不能作为其业务证据。

不过,Token 聚合、算力调度、模型路由和按量结算这一更广泛的产业方向,正在从构想进入基础设施验证阶段。

涉及具体折扣、佣金、路由保障和结算周期时,应以 Unexhub 当前正式规则为准。


路由并不会让所有模型变得完全一样

路由与交易市场的兴起,并不意味着模型质量失去意义。

不同模型在以下方面仍然存在明显差异:

  • 推理;

  • 编程;

  • 安全;

  • 上下文容量;

  • 多语言表现;

  • 工具调用;

  • 输出稳定性;

  • 合规支持。

最便宜的路线,如果产生以下问题,最终总成本反而可能更高:

  • 更多失败;

  • 更多重试;

  • 更长输出;

  • 更多人工修正;

  • 更大运营风险。

因此,真正应该比较的不是:

每百万 Token 的价格。

而是:

成功完成一项任务的实际总成本。

成熟的路由市场不能只追求最低报价。

还需要平衡:

  • 价格;

  • 质量;

  • 时延;

  • 稳定性;

  • 风险。

最优路线不一定是绝对最便宜的模型。

而是在能够正确完成任务的前提下,总成本最低的路线。


450 万亿 Token,代表的是市场发现权

450 万亿 Token 真正重要的,不只是数量巨大。

而是这些流量产生了庞大的市场信息。

路由层能够观察:

  • 开发者正在测试哪些模型;

  • 哪些模型正在被放弃;

  • 市场对价格有多敏感;

  • 哪些供应商高峰期容易失败;

  • 哪些任务必须使用强模型;

  • 哪些地区容量不足;

  • 价格调整后流量多快迁移。

这些信息让路由平台拥有买卖双方都不具备的全局优势。

模型供应商只能看到自己的流量。

路由层可以看到所有人的流量。

这使它不只是一个分销渠道。

还是一套实时运行的价格发现系统。


靠一个模型征服世界的时代正在结束

OpenRouter CEO Alex Atallah 曾表示,依靠选择一个模型征服世界的时代已经结束。

更进一步的结论也许是:

一家模型供应商独占全部客户关系的时代,也正在结束。

企业未来会同时使用:

  • 一个模型做分类;

  • 一个模型做推理;

  • 一个模型做编程;

  • 一个模型处理受监管数据;

  • 一个模型完成低成本批量任务。

模型能力正在模块化。

路由层则变成战略核心。

供应商可能赢得这一笔请求,又输掉下一笔。

在它们之间作出选择的平台,正在成为新的权力中心。


最后的结论

Kimi K3 在 48 小时内进入 OpenRouter 前十,说明路由层可以多快地为新模型创造生产级需求。

OpenRouter 对约 450 万亿 Token 的分析,则揭示了更深层的变化:

下一阶段 AI 市场最重要的能力,未必只是训练模型。

还包括把流量分配给谁。

当开发者只需修改一个参数就可以切换模型时:

  • 品牌溢价开始松动;

  • 切换成本快速下降;

  • 价格竞争持续加剧;

  • 新模型能够更快获得规模;

  • 旧供应商也可能一夜之间被替换。

OpenRouter 把模型变成可路由的服务。

Unexhub 则希望进一步把可路由的 Token 容量变成动态定价的交易市场。

按照 Unexhub 当前平台定位,其希望提供:

  • 对 Gemini、ChatGPT、Claude、DeepSeek、GLM 等模型资源的统一接入;

  • 面向个人和中小 AI 供应商的分发渠道;

  • 按实际 Token 使用量付费;

  • 符合条件的路线最低接近市场或官方价格三折;

  • 毫秒级请求路由;

  • 更快的按量收益结算;

  • 根据供需变化进行 API 竞价。

这些权益能否真正形成长期价值,最终取决于平台能否提供:

  • 稳定容量;

  • 透明 Token 计量;

  • 可验证服务质量;

  • 可审计结算;

  • 可信合规与治理。

旧 AI 市场问的是:

哪个模型最强?

路由市场问的是:

这条请求应该交给哪个模型?

交易市场则进一步追问:

这条请求应该以什么价格成交?

当这些问题由路由层与交易层共同回答时,模型供应商就不再独占完整的定价过程。

450 万亿 Token 真正代表的,不是一个平台处理了多少流量。

而是路由层已经获得决定流量流向的权力,并开始决定这些流量究竟值多少钱。