
From RMB 100 to RMB 0.025 per Million Tokens: China’s AI Model Market Is Splitting Into Two Pricing Worlds
In July 2026, the pricing logic of China’s large-model market split into two sharply different paths.
On one side, Kimi K3 priced output at RMB 100 per million tokens.
On the other, Xiaomi’s MiMo-V2.5 reduced cache-hit input pricing to as little as RMB 0.025 per million tokens.
The two headline figures differ by a factor of 4,000.
Strictly speaking, they do not represent identical billing items: K3’s RMB 100 rate applies to output, while MiMo’s RMB 0.025 rate applies to cache-hit input. Yet the contrast still captures a fundamental change in the market.
This is no longer a conventional price war in which every provider races toward the same bottom.
Different model companies are now building different bridges across the same river:
One prices according to the value of a completed task;
Another prices according to the marginal cost of inference;
One protects premium intelligence;
Another attempts to make standardized intelligence nearly free.
The market is no longer asking which model is simply cheaper.
It is beginning to ask:
What type of token is being purchased, and what business result can it deliver?
Why Can Kimi K3 Charge RMB 100?
Moonshot AI released Kimi K3 on July 16.
According to the published specifications, the model features:
2.8 trillion total parameters;
896 experts, with 16 activated during each inference;
A one-million-token context window;
Native visual understanding;
A focus on long-horizon coding, complex reasoning, and end-to-end knowledge work.
The following day, K3 reportedly scored 1,679 on Frontend Code Arena, ahead of Claude Fable 5 at 1,631 and GPT-5.6 Sol at 1,618.
Its API pricing attracted just as much attention:
Cache-hit input: RMB 2 per million tokens;
Cache-miss input: RMB 20 per million tokens;
Output: RMB 100 per million tokens.
Compared with the RMB 27 output price of the previous K2.6 flagship, K3’s listed output rate increased by more than 3.5 times.
This was a rare moment in which a Chinese open model adopted a price structure close to that of a leading American model.
Converted into dollars, K3’s RMB 20 input and RMB 100 output rates are broadly comparable to Claude Sonnet 5’s standard pricing tier of approximately $3 input and $15 output per million tokens.
For years, Chinese models had largely competed by offering similar capability at a lower price.
K3 is attempting something different:
It is asking the market to pay a premium for the value of the intelligence itself.
K3’s First Advantage: Architecture Separates Scale From Active Cost
K3’s pricing cannot be understood by looking only at its 2.8-trillion-parameter scale.
Although the full model is extremely large, only 16 of its 896 experts are activated for each inference. The total parameter count provides a broad capability base, while actual computation uses only a small portion of that capacity for each request.
This mixture-of-experts structure is designed to separate model scale from serving cost.
K3 also uses the KDA linear-attention mechanism, which reportedly improves decoding speed by as much as 6.3 times in million-token contexts.
Its AttnRes design is said to produce an approximately 25% improvement in training efficiency while adding less than 2% extra compute.
The key idea is straightforward:
A model can become much larger without requiring every parameter to participate in every request.
This weakens the traditional assumption that inference cost must rise proportionally with total model size.
Artificial Analysis has estimated K3’s average cost per task at approximately $0.94, compared with around $1.80 for Opus 4.8.
If that task-level efficiency remains stable in real workloads, the RMB 100 output rate may not be as expensive as the headline suggests.
A model can charge more per output token and still offer a lower completed-task cost if it:
Requires fewer retries;
Completes more tasks successfully;
Produces shorter effective workflows;
Reduces human intervention;
Maintains context across long projects.
K3 is not trying to sell the cheapest token.
It is trying to sell a more valuable completed result.
K3’s Second Advantage: Mooncake and the Economics of Caching
The more important element of K3’s pricing model may be its Mooncake disaggregated inference architecture.
In many coding and enterprise workflows, users repeatedly submit the same large body of context:
A code repository;
System instructions;
Product documentation;
Database schemas;
Tool definitions;
Long conversation histories;
Agent rules.
If the model has to process all of that content from the beginning during every request, input costs can grow rapidly.
Moonshot AI has stated that cache-hit rates in coding scenarios can exceed 90%.
At a 90% cache-hit rate, the blended input price changes dramatically.
Using K3’s listed rates:
90% of input is billed at RMB 2;
10% is billed at RMB 20.
The resulting blended input cost is approximately:
RMB 3.8 per million tokens.
That is significantly lower than the RMB 20 uncached rate displayed on the pricing page.
For developers repeatedly working with the same repository, documentation set, or persistent agent context, the effective cost can therefore be much closer to the cache price than the full list price.
K3’s pricing strategy becomes clearer when viewed this way:
The premium list price filters casual demand;
High cache reuse rewards persistent workflows;
Expensive output reflects the value of complex generation;
Long-running users receive much better effective economics.
The RMB 100 output rate is the visible price.
The RMB 2 cache-hit input rate is the mechanism designed to retain serious users.
K3 is effectively saying:
If you build your workflow properly, the real bill may be far lower than the rate card suggests.
Xiaomi MiMo: Why Can It Cut the Price to Two and a Half Fen?
Xiaomi is taking the opposite path.
On May 27, the company announced permanent price reductions for the MiMo-V2.5 series, with cuts of up to 99%.
The cache-hit input price for MiMo-V2.5-Pro fell to:
RMB 0.025 per million tokens.
The standard MiMo-V2.5 model went even lower:
RMB 0.02 per million tokens.
The broader pricing structure—RMB 0.025, RMB 3, and RMB 6—closely aligns with the benchmark established by DeepSeek V4 Pro.
Xiaomi was not attempting to invent an entirely new premium category.
It was matching one of the most aggressive cost baselines in the domestic market.
Its message to developers was completely different from K3’s:
Tokens do not always need to be expensive. With sufficient engineering optimization, routine intelligence can become almost negligible in cost.
MiMo’s Price Cut Is Supported by Inference Engineering
Xiaomi’s ability to reduce prices depends on more than a marketing subsidy.
A major part of the improvement comes from deep KV-cache optimization.
With SGLang HiCache providing broader support for sliding-window attention, data movement between different KV-cache storage layers was reportedly reduced to nearly one-seventh of its previous level.
At the same time, the number of cacheable tokens increased to almost five times the earlier capacity.
These improvements matter because inference cost is not determined only by mathematical operations.
It also depends heavily on:
Memory movement;
Cache storage;
Repeated context processing;
Communication between storage layers;
GPU utilization;
Scheduling efficiency.
When frequently reused context can remain available in cache, the system performs less redundant computation and transfers less data.
Xiaomi’s technical lead Luo Fuli previously noted that the reduction of up to 99% in cache-hit input cost was rooted in deep optimization of the inference framework and KV cache.
In other words, MiMo’s low price is not simply saying:
“We are willing to lose money to acquire users.”
It is also saying:
“We have reduced the cost of serving repeated input enough to reset the market price.”
Xiaomi Also Removed Long-Context Pricing Penalties
Another aggressive part of Xiaomi’s strategy is the removal of traditional pricing differentiation based on context-window length.
Many providers charge more once a request exceeds a certain context threshold.
Xiaomi instead increased the number of usable tokens in its Token Plan packages by approximately five to eight times within the same payment tier.
This creates a simple user proposition:
Use longer contexts;
Process larger batches;
Reuse more cached material;
Pay less per unit of useful work.
K3 encourages developers to pay for high-value output and optimize around caching.
MiMo encourages developers to expand usage because the marginal cost has become extremely low.
One protects premium economics.
The other uses low marginal cost to stimulate volume.
Two Pricing Paths, Two Different Definitions of Value
K3 and MiMo are not simply choosing different numbers.
They are applying different economic theories to AI.
K3 Uses Task-Value Pricing
K3 is designed for workloads such as:
Long-horizon coding;
Complex reasoning;
Multi-step agents;
End-to-end knowledge work;
Large-repository analysis;
Persistent context workflows.
Its RMB 100 output rate is intended to represent more than the cost of generating text.
It represents the ability to complete a difficult block of work.
A Morgan Stanley commentary summarized this logic sharply:
K3 is driving API pricing rather than merely responding to compute demand.
The model is attempting to tell the market:
High-quality tokens are not interchangeable with ordinary tokens, and premium intelligence deserves a premium price.
MiMo Uses Marginal-Cost Pricing
MiMo follows a different philosophy.
As model capabilities become more standardized across many everyday tasks, pricing begins to converge toward the cost of producing the next token.
Its RMB 0.025 cache-hit input rate is close to saying:
“For another cached request, we will charge little more than the electricity, depreciation, memory movement, and serving cost.”
MiMo is not trying to prove that every token is valuable.
It is trying to prove that a massive volume of useful tokens can be produced extremely cheaply.
The Two Positions Are Not Contradictory
K3 is saying:
“I can complete work that cheaper models may not finish, so I deserve a premium.”
MiMo is saying:
“For work that many models can already complete, I can make the cost almost disappear.”
Both claims can be true.
They serve different workloads, different users, and different budget structures.
A company may use MiMo for:
Classification;
Extraction;
Basic scripting;
Batch processing;
Routine customer service.
The same company may use K3 for:
Complex code refactoring;
Long-running agents;
Architecture analysis;
High-value knowledge work;
Tasks where failure is expensive.
The two models are not crossing the same bridge.
But they are crossing the same river of global token demand.
The 4,000× Gap Is Not Market Fragmentation
A 4,000× gap may appear irrational.
But it may actually indicate that the market is finally learning to price different forms of intelligence differently.
The value of a token varies according to:
Model capability;
Task difficulty;
Context reuse;
Output quality;
Latency;
Reliability;
Domain specialization;
Probability of task completion.
In extreme cases, the cross-scenario value gap can become enormous.
A token used in pharmaceutical research, legal due diligence, or high-value engineering may be worth orders of magnitude more than a token used for casual conversation.
Some industry estimates place specialized drug-discovery workloads at as much as $1,000 per million tokens, while low-value conversational traffic can fall toward $0.01.
The gap can reach tens of thousands of times.
This does not mean the market is broken.
It means that “a token” is no longer a sufficient description of the product.
The market must also ask:
Which model produced it?
What task did it complete?
How quickly was it delivered?
How reliable was the service?
What would failure have cost?
RMB 100 tokens and RMB 0.025 tokens can coexist because they represent different combinations of capability and economics.
The market is finally learning to attach different price tags to different intelligence.
The Token Marketplace Will Amplify the Price Split
This differentiation will become even more visible in a token-trading market.
According to its public positioning, Unexhub is building a computing-power trading center for token aggregation and distribution.
The platform aims to connect suppliers across multiple levels, including access to model ecosystems such as:
Gemini;
ChatGPT;
Claude;
DeepSeek;
GLM;
Models offered by individual developers and small or medium-sized providers.
Its proposed API bidding model allows prices to fluctuate with supply and demand.
During off-peak periods:
Developers may receive lower prices;
Suppliers can monetize idle capacity;
Flexible batch tasks can use economy routes.
During high-demand periods:
Prices can rise;
Suppliers can earn a scarcity premium;
Urgent tasks can purchase priority capacity.
The wider the official price range becomes, the more important this marketplace function becomes.
When K3’s premium tokens and MiMo’s ultra-low-cost tokens are offered in the same environment, developers can make task-specific choices in real time:
Use MiMo for a simple script;
Use a balanced model for routine development;
Use K3 for a complex coding agent;
Use a specialist model for a high-value vertical task.
The gap creates choice.
The marketplace turns that choice into an operational system.
Millisecond-Level Routing Turns Price Differences Into Savings
Unexhub also emphasizes millisecond-level routing across multiple models and suppliers.
The practical objective is to prevent two types of waste:
Paying for more intelligence than a task needs;
Choosing a cheap model that repeatedly fails and becomes more expensive overall.
A routing system can consider:
Task complexity;
Context length;
Cache-hit potential;
Model capability;
Current price;
Response latency;
Route availability;
Supplier reliability;
User-defined budgets.
A simple task can be assigned to a low-cost route.
A difficult task can be escalated to a stronger model.
A repeated long-context request can use a cache-optimized provider.
An urgent production task can use priority capacity.
In a market with a 4,000× headline price range, intelligent routing is not merely a technical feature.
It is the difference between an affordable AI application and an uncontrolled bill.
Developer Benefits: Pay Only for the Intelligence You Actually Need
Through Unexhub, developers can pay according to actual token usage instead of maintaining separate prepaid balances and technical integrations with every model provider.
Depending on the model, route, supplier, and market conditions, some services may be available at prices as low as approximately 30% of prevailing market rates.
For developers and enterprise teams, this creates several practical benefits:
Lower costs for standardized workloads;
More affordable multi-model testing;
Greater flexibility for long-context projects;
Reduced spending on batch processing;
Easier escalation to premium models;
Less dependence on a single provider;
Centralized usage and billing.
The most valuable part of the discount is not simply obtaining one cheap API call.
It is being able to build a layered model strategy:
Low-cost models handle volume;
Premium models handle difficulty;
Routing determines when each is used;
The user pays only for the capability required by the task.
This makes the phrase “prices as low as 30% of the market rate” more meaningful.
The saving comes not only from a lower route price, but also from avoiding unnecessary premium calls.
Supplier Benefits: Distribution, Settlement, and Early Zero Commission
For model and compute suppliers, Unexhub’s value lies on the other side of the marketplace.
A smaller supplier may possess:
A competitive model;
Specialized domain performance;
Idle inference capacity;
Regional infrastructure;
A cost advantage.
But it may lack:
Developer traffic;
Enterprise sales channels;
Billing infrastructure;
Global distribution;
Settlement capabilities;
Multi-tenant API management.
Unexhub proposes a one-step integration process that allows suppliers to connect their capacity to a wider market.
Requests are routed according to demand, and suppliers receive revenue based on actual token usage.
The platform also proposes real-time or near-real-time settlement, shortening the delay between providing capacity and receiving payment.
For smaller suppliers, faster settlement can:
Reduce receivables pressure;
Improve cash flow;
Support faster expansion;
Lower the financial burden of serving more traffic.
Qualifying early suppliers may also receive zero-commission incentives.
The benefit is not only saving a fee.
It gives suppliers a lower-cost opportunity to:
Complete integration;
Test demand;
Establish transaction records;
Validate settlement;
Build an early market position.
The Future Is Not One Price—It Is a Portfolio of Prices
The K3–MiMo contrast suggests that the market will not converge on one universal token price.
Instead, developers and enterprises will operate portfolios of models and prices.
A mature AI stack may contain:
Ultra-low-cost models for extraction and classification;
Balanced models for routine coding and documents;
Cache-efficient models for persistent long-context workflows;
Premium models for difficult agents and complex reasoning;
Specialized models for medicine, law, finance, audio, or vision.
The competitive advantage will come from orchestration.
The winning company will not necessarily be the one that always chooses the lowest-priced model.
Nor will it be the one that sends every request to the strongest model.
It will be the one that assigns each task to:
The lowest-cost model that can complete it reliably.
Final Thoughts: Different Bridges, the Same River
Kimi K3 and Xiaomi MiMo-V2.5 are traveling along two very different pricing paths.
K3 believes valuable intelligence should command a premium.
MiMo believes standardized intelligence should approach marginal cost.
K3 prices according to what the task is worth.
MiMo prices according to what the next token costs to produce.
They are not fighting on the same bridge.
But beneath them flows the same river: a rapidly expanding global demand for tokens.
The hotter the token market becomes, the faster that river flows.
Its direction will increasingly be influenced by the platforms that control:
Price discovery;
Model routing;
Supplier access;
Usage settlement;
Developer distribution.
Unexhub’s ambition is to become one of those platforms by combining model aggregation, API bidding, millisecond-level routing, real-time supplier settlement, and early zero-commission incentives.
For developers, that could mean paying by actual usage and obtaining suitable model routes at prices as low as approximately 30% of prevailing market levels.
For suppliers, it could mean connecting once, reaching more customers, and recovering revenue faster.
The coexistence of RMB 100 tokens and RMB 0.025 tokens does not mean the market has fractured.
It means the market has finally begun to understand that different tokens should carry different price tags.
The next stage of competition will not be about forcing every model onto the same price bridge.
It will be about building a market where every task can find the right bridge—and cross the river at the right cost.
中文版本
从每百万 Token 100 元到 0.025 元:国产大模型走上了两条完全不同的定价道路
2026 年 7 月,国产大模型的定价逻辑,被拉向了两条截然不同的路径。
一边,Kimi K3 将 API 输出价格锚定在:
每百万 Token 100 元。
另一边,小米 MiMo-V2.5 则将缓存命中输入价格压低至:
每百万 Token 0.025 元。
两组最醒目的数字,相差整整 4000 倍。
严格来说,这并不是完全相同口径下的比较:K3 的 100 元对应输出价格,而 MiMo 的 0.025 元对应缓存命中输入价格。
但这一差距依然清晰地揭示了市场正在发生的变化。
这已经不再是一场所有厂商共同冲向最低价的传统价格战。
不同模型厂商,开始在同一条 Token 河流上,各自修建不同的桥:
一方按照任务价值定价;
另一方按照推理边际成本定价;
一方维护高端智能的溢价;
另一方试图让标准化智能接近免费。
市场开始关心的,不再只是哪个模型更便宜。
而是:
用户购买的究竟是哪一种 Token,它最终能够交付什么结果?
Kimi K3 为什么敢把输出卖到 100 元?
7 月 16 日,月之暗面发布 Kimi K3。
根据公开参数,K3 拥有:
2.8 万亿总参数;
896 个专家,每次推理仅激活 16 个;
100 万 Token 上下文窗口;
原生视觉理解能力;
面向长程编程、复杂推理和端到端知识工作的产品定位。
发布次日,K3 在 Frontend Code Arena 上取得 1679 分,据称超过 Claude Fable 5 的 1631 分和 GPT-5.6 Sol 的 1618 分。
与跑分一同引发讨论的,是它的 API 价目表:
缓存命中输入:2 元/百万 Token;
缓存未命中输入:20 元/百万 Token;
输出:100 元/百万 Token。
相比前代 K2.6 每百万 Token 27 元的输出价格,K3 的挂牌价上涨超过 3.5 倍。
这是国产开源模型少见的一次高价定标。
换算成美元后,K3 的输入与输出价格,已经大致进入 Claude Sonnet 5 每百万 Token 3 美元输入、15 美元输出的标准价格区间。
过去,国产模型大多通过“相近能力、更低价格”参与竞争。
K3 则试图走出另一条路:
不再只靠便宜取胜,而是要求市场为智能能力本身支付溢价。
K3 的第一层底气:架构让参数规模与实际成本解耦
理解 K3 的价格,不能只看 2.8 万亿参数这个数字。
虽然完整模型规模极大,但每次推理只激活 896 个专家中的 16 个。
总参数为模型提供更加广阔的知识与能力基础,单次请求却只调用其中一小部分。
这种混合专家架构,试图将模型规模与实际推理成本分离。
K3 同时采用 KDA 线性注意力机制。据公开资料,在百万 Token 级上下文中,其解码速度最高可提升 6.3 倍。
AttnRes 则希望通过不到 2% 的额外算力,带来约 25% 的训练效率增益。
核心逻辑是:
模型可以变得更大,但不需要让全部参数参与每一次推理。
这削弱了“参数越多,服务成本必然按比例增加”的传统关系。
根据 Artificial Analysis 的测算,K3 的单任务平均成本约为 0.94 美元,低于 Opus 4.8 的约 1.80 美元。
如果这种任务级效率能够在真实生产工作流中保持,那么每百万输出 Token 100 元的价格,并没有标题看起来那么昂贵。
一款模型即使输出单价更高,只要它能够:
减少重试;
提高任务完成率;
缩短实际执行链路;
减少人工接管;
在长周期任务中维持上下文;
最终的单任务成本依然可能更低。
K3 不是在卖最便宜的 Token。
它试图出售更有价值的任务结果。
K3 的第二层底气:Mooncake 与缓存经济
K3 定价策略中更加关键的一环,是 Mooncake 分离式推理架构。
在大量编程和企业工作流中,用户会反复提交同一批长上下文:
代码仓库;
系统提示词;
产品文档;
数据库结构;
工具定义;
长对话历史;
智能体规则。
如果每次请求都要从头处理这些内容,输入费用会迅速增加。
月之暗面披露,编程场景中的缓存命中率可以超过 90%。
按照 90% 缓存命中率计算,K3 的混合输入成本会发生明显变化:
90% 的输入按照 2 元计费;
10% 的输入按照 20 元计费。
最终混合输入成本约为:
3.8 元/百万 Token。
这远低于价目表上的 20 元未命中价格。
对反复处理同一代码仓库、文档集或持续 Agent 上下文的开发者来说,实际成本会更接近缓存价格,而不是完整挂牌价。
由此,K3 的定价策略变得十分清晰:
高挂牌价筛选一次性和低价值需求;
高缓存利用率奖励持续工作流;
输出价格反映复杂生成能力;
长期、高频用户获得更好的实际经济性。
100 元,是最醒目的标价。
2 元的缓存命中价格,才是其留住深度用户的机制。
K3 实际上是在告诉开发者:
工作流设计得越好,最终账单与挂牌价之间的差距就越大。
小米 MiMo:为什么敢砍到 2 分 5 厘?
小米走上了完全相反的道路。
5 月 27 日,小米宣布 MiMo-V2.5 系列 API 永久降价,最高降幅达到 99%。
其中,MiMo-V2.5-Pro 的缓存命中输入价格降至:
0.025 元/百万 Token。
标准版 MiMo-V2.5 更低至:
0.02 元/百万 Token。
0.025 元、3 元、6 元这组价格,与 DeepSeek V4 Pro 建立的低价基准高度接近。
小米并不是试图创造一个新的高端溢价区间,而是在直接对齐国内市场最激进的成本底线。
它向开发者传递的信号与 K3 完全不同:
Token 并不一定昂贵。只要工程优化足够深入,日常智能可以便宜到接近忽略不计。
MiMo 的低价来自推理框架优化
小米能够大幅降价,并不完全依赖市场补贴。
其中一项关键能力,是对 KV Cache 的深度优化。
基于 SGLang HiCache 对滑动窗口注意力的全面支持,KV Cache 在多级存储之间的数据传输量,据称降低至优化前的近七分之一。
可缓存 Token 数量则提升至接近原来的五倍。
这些优化十分重要,因为推理成本不只来自模型计算,还来自:
内存搬运;
缓存存储;
重复上下文处理;
不同存储层之间的通信;
GPU 利用率;
调度效率。
当重复上下文可以被更有效地保留在缓存中,系统就不必反复计算相同内容,也能减少大量数据传输。
小米技术团队负责人罗福莉此前在公开分享中提到,缓存命中输入成本最高下降 99%,根本原因正是推理框架对 KV Cache 的深度优化。
因此,MiMo 的低价不只是说:
“我们愿意暂时亏钱换市场。”
它还在表达:
“我们已经把重复输入的服务成本降到足以重新定义价格。”
小米还取消了长上下文价格惩罚
小米策略中更激进的一点,是不再按照上下文窗口长度进行传统分档定价。
许多模型会在上下文超过一定长度后,进入更高价格区间。
小米则在相同付费档位下,将 Token Plan 套餐的可用 Token 数量提升约 5 至 8 倍。
对用户而言,权益十分直接:
可以使用更长的上下文;
处理更大的批量任务;
复用更多缓存内容;
降低单位任务成本。
K3 鼓励开发者为高价值输出付费,再通过缓存优化实际账单。
MiMo 则鼓励开发者大胆扩大用量,因为边际成本已经被压得极低。
一个维护高端智能的价格。
另一个通过低成本刺激规模增长。
两条道路,代表两种完全不同的价值逻辑
K3 与 MiMo 的区别,不只是价目表上的数字不同。
它们采用的是两种不同的 AI 经济学。
K3 采用“任务价值定价”
K3 面向的是:
长程编程;
复杂推理;
多步骤智能体;
端到端知识工作;
大型代码仓库分析;
持续上下文任务。
它的 100 元输出价格,不只是生成文字所需的成本。
它试图代表完成一项高难度工作的能力。
摩根士丹利的一项评论概括得十分直接:
K3 正在主动推动 API 价格,而不只是被动响应算力需求。
它希望告诉市场:
高质量 Token 与普通 Token 并非同一种商品,复杂智能应该获得溢价。
MiMo 采用“边际成本定价”
MiMo 的逻辑则完全不同。
当大量日常任务的模型能力开始趋于同质化,价格会逐渐向生产下一个 Token 的边际成本靠拢。
0.025 元的缓存命中输入价格,几乎是在表达:
“你再调用一次,我只收取接近电力、折旧、内存搬运和服务运行的成本。”
MiMo 不是要证明每个 Token 都很值钱。
它要证明的是:有用的 Token 可以以极低成本被大规模生产。
两种观点并不矛盾
K3 在说:
“我能完成便宜模型可能做不完的工作,所以我应该更贵。”
MiMo 在说:
“对于大家都能完成的工作,我可以把成本降得极低。”
两句话可以同时成立。
它们服务的是不同的任务、用户和预算结构。
一家企业可以用 MiMo 处理:
分类;
抽取;
简单脚本;
批量任务;
常规客服。
同时用 K3 处理:
复杂代码重构;
长时间智能体;
架构分析;
高价值知识工作;
失败成本较高的任务。
它们各过各的桥,却跨越同一条全球 Token 需求的河流。
4000 倍价差,不代表市场割裂
4000 倍的价格差距,看起来似乎并不理性。
但它也可能意味着,市场终于开始学会给不同形态的智能能力定出不同价格。
Token 的价值会受到以下因素影响:
模型能力;
任务难度;
上下文复用率;
输出质量;
延迟;
稳定性;
行业专业度;
任务完成概率。
在极端行业场景中,跨任务价值差距甚至可能达到数万倍。
例如,药物研发、法律尽调和高价值工程任务中的 Token,其商业价值可能远高于普通闲聊 Token。
部分行业估算中,药物研发类任务的价格可以达到每百万 Token 1000 美元,而低价值闲聊流量可能低至 0.01 美元。
价格差距可能达到十万倍。
这并不意味着市场失灵。
它说明“Token”本身已经不足以完整描述商品。
市场还需要了解:
它由哪个模型生产?
完成了什么任务?
以多快速度交付?
服务有多稳定?
如果失败,代价有多大?
100 元档 Token 与 2 分 5 厘档 Token 可以共存,因为它们代表的是完全不同的能力与成本组合。
市场终于开始为不同的智能能力贴上不同价签。
Token 交易市场会进一步放大价格分化
这种价格分层,在 Token 交易市场中会变得更加明显。
根据公开定位,Unexhub 正在建设 Token 聚合与分发的算力交易中心。
平台希望汇聚不同量级的 AI 供应商,包括:
Gemini;
ChatGPT;
Claude;
DeepSeek;
GLM;
个人开发者与中小企业提供的模型。
其 API 竞价机制,希望让价格随市场供需动态波动。
在低谷时段:
开发者可以获得更低价格;
供应商能够出售闲置算力;
可延后批量任务选择经济线路。
在需求高峰:
价格可以上浮;
供应商获得稀缺性溢价;
紧急任务购买优先算力。
官方价格带越宽,这类市场的作用就越明显。
当 K3 的高价值 Token 与 MiMo 的超低价 Token 被放在同一个市场中,开发者可以按照任务实时选择:
简单脚本使用 MiMo;
日常开发使用均衡型模型;
复杂智能体调用 K3;
高价值行业任务选择专业模型。
价差带来了选择。
交易市场则将选择变成一套可运行的系统。
毫秒级路由,把价格差转化为实际节省
Unexhub 同时强调跨模型、跨供应商的毫秒级路由。
其直接目标,是避免两种浪费:
为简单任务购买过量智能能力;
为追求低单价选择频繁失败、最终总成本更高的模型。
路由系统可以综合考虑:
任务复杂度;
上下文长度;
缓存命中潜力;
模型能力;
实时价格;
响应延迟;
线路可用性;
供应商稳定性;
用户预算。
简单任务自动进入低成本线路。
复杂任务升级至更强模型。
重复长上下文请求选择缓存效率更高的供应商。
紧急生产任务使用优先算力。
在一条达到 4000 倍的价格带中,智能路由不只是技术功能。
它直接决定一个 AI 应用的成本能否被长期控制。
开发者权益:只为任务真正需要的能力付费
通过 Unexhub,开发者可以按照实际 Token 用量付费,不必分别在多个模型平台维护预付余额与技术接口。
根据具体模型、线路、供应商和市场条件,部分服务价格最低可达到常规市场水平的约 3 折。
对开发者和企业团队来说,这能够带来多项实际权益:
降低标准化任务成本;
更低成本测试多个模型;
提高长上下文项目灵活性;
减少批量处理支出;
更方便地升级至高端模型;
降低对单一供应商的依赖;
统一用量与账单管理。
优惠的价值,并不只是获得一次低价调用。
更重要的是能够建立分层模型策略:
低价模型承担调用规模;
高端模型解决困难任务;
路由系统决定何时使用哪一个;
用户只为任务真正需要的能力付费。
此时,“最低可至市场价约 3 折”才不再只是宣传数字。
节省不仅来自更低的线路报价,也来自减少不必要的高价模型调用。
供应商权益:分发、结算与早期零佣金
对模型和算力供应商而言,Unexhub 提供的是市场另一侧的价值。
中小供应商可能拥有:
有竞争力的模型;
专业领域表现;
闲置推理算力;
区域基础设施;
成本优势。
但它们可能缺少:
开发者流量;
企业销售渠道;
计费基础设施;
全球分发能力;
结算系统;
多租户 API 管理。
Unexhub 提出一键接入,使供应商有机会将自身产能开放至更广泛的市场。
平台根据需求路由请求,再按照真实 Token 用量完成收益结算。
即时或准即时结算机制,则进一步缩短了从提供算力到收到资金的时间。
对中小供应商而言,更快结算可以:
降低应收账款压力;
改善现金流;
支持更快扩容;
减轻承接更多流量时的资金负担。
符合条件的早期供应商还可以享受零佣金激励。
其价值不只是节省抽成,而是可以用更低成本:
完成平台接入;
测试真实需求;
建立交易记录;
验证结算流程;
提前建立市场位置。
未来不是一种价格,而是一个价格组合
K3 与 MiMo 的对比表明,未来市场不会最终收敛为一个统一 Token 价格。
开发者和企业更可能运营一套模型与价格组合。
成熟的 AI 技术栈可能同时包含:
超低成本模型处理抽取与分类;
均衡型模型承担常规编程与文档任务;
高缓存效率模型运行持续长上下文工作流;
高端模型负责困难智能体与复杂推理;
专业模型服务医疗、法律、金融、音频或视觉任务。
真正的竞争力来自如何组织这些模型。
赢家未必是永远选择最低价模型的企业。
也不会是让所有任务都调用最强模型的企业。
而是能够将每一个任务分配给:
足以可靠完成它的最低成本模型。
写在最后:各过各的桥,桥下是同一条河
Kimi K3 与小米 MiMo-V2.5,正在沿着两条完全不同的定价道路前进。
K3 认为,高价值智能应当获得溢价。
MiMo 认为,标准化智能应该接近边际成本。
K3 按照任务价值定价。
MiMo 按照下一个 Token 的生产成本定价。
它们并没有在同一座桥上竞争。
但桥下流淌的,是同一条不断扩张的全球 Token 需求之河。
Token 市场越热,这条河的流速就越快。
未来水流的方向,将越来越受到以下平台能力影响:
价格发现;
模型路由;
供应商接入;
用量结算;
开发者分发。
Unexhub 希望通过模型聚合、API 竞价、毫秒级路由、供应商即时结算和早期零佣金激励,成为其中的一座关键枢纽。
对开发者来说,这意味着按照实际用量付费,并在合适的模型与线路下,获得最低约为常规市场价格 3 折的服务。
对供应商来说,则意味着一次接入、获得更多分发,并加快资金回流。
100 元档 Token 与 2 分 5 厘档 Token 同时存在,并不意味着市场已经割裂。
它意味着市场终于开始理解:
不同 Token,本来就应该拥有不同的价格。
下一阶段的竞争,不会是强迫所有模型走上同一座价格桥。
而是建立一个市场,让每一种任务都能找到合适的桥,并以合适的成本到达对岸。