
From RMB 100 to RMB 0.025 per Million Tokens: AI API Pricing Is Splitting Into a 4,000× Market
In July 2026, the large-model API market produced one of its most dramatic pricing contrasts to date.
During the same week, Kimi K3 set its output price at RMB 100 per million tokens.
At the other end of the market, Xiaomi’s MiMo-V2.5 series announced price reductions of up to 99%, with cache-hit input falling as low as RMB 0.025 per million tokens.
One side quoted RMB 100.
The other quoted two and a half fen.
The two headline figures were separated by a factor of 4,000.
Strictly speaking, this is not a like-for-like comparison: one figure represents output pricing, while the other is a discounted cache-hit input rate. But that does not weaken the signal. It makes the market transformation even clearer.
Large-model pricing is no longer moving along a single curve from expensive to cheap.
It is splitting into a wide spectrum shaped by:
Model capability;
Input and output type;
Cache utilization;
Context length;
Latency;
Task complexity;
Service priority;
Available compute;
Target customer profile.
This is no longer a conventional price war.
It is the creation of an entirely new pricing band.
Why Can Kimi K3 Charge RMB 100?
Moonshot AI released Kimi K3 on July 16.
According to the announced specifications, K3 contains:
2.8 trillion total parameters;
896 experts, with 16 activated for each inference;
A one-million-token context window;
Native support for long-horizon coding and knowledge work.
Its API pricing is equally distinctive:
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.
Among open-source models, this is an unusually confident price.
So where does that confidence come from?
The First Answer Is Architecture
K3’s pricing cannot be understood simply by looking at its total parameter count.
Its architecture is designed to separate model scale from active inference cost.
Although the model has 2.8 trillion parameters, only 16 of its 896 experts are activated during each inference step. The full model provides a broad knowledge and capability base, while only part of that capacity is used for each request.
K3 also introduces a KDA linear-attention mechanism, which reportedly improves decoding efficiency for million-token contexts by as much as 6.3 times.
Its AttnRes architecture is designed to improve training efficiency, reportedly delivering an approximately 25% gain with less than 2% additional compute.
The important point is not simply that the model is large.
It is that Moonshot AI is attempting to make a very large model commercially usable through selective activation, more efficient long-context processing, and improved inference architecture.
This weakens the traditional assumption that:
More parameters must always mean proportionally higher serving costs.
Artificial Analysis estimated K3’s average cost per task at approximately $0.94, compared with around $1.80 for Opus 4.8.
If those task-level economics hold in real production workloads, the RMB 100 output price may be less aggressive than it first appears.
The listed price is high, but the amount of output required to complete a task—and the model’s probability of finishing successfully—can matter more than the price of each individual token.
The Second Answer Is Cache Economics
The more important part of K3’s pricing strategy may be its Mooncake disaggregated inference architecture.
In many coding and enterprise workflows, the same large body of context is reused repeatedly:
Repository files;
System prompts;
Product documentation;
Database schemas;
Long conversation histories;
Agent instructions;
Tool definitions.
If that context must be processed from the beginning during every request, input costs rise rapidly.
Moonshot AI claims that cache-hit rates in coding scenarios can exceed 90%.
Under such conditions, the practical input cost can be far below the uncached list price.
A workflow may appear expensive when evaluated through the RMB 20 cache-miss rate, but if most repeated context is charged at RMB 2, the effective input bill changes significantly.
This creates a deliberate pricing structure:
Uncached input is expensive;
Reused context is much cheaper;
Output carries a premium;
Long-running, stable workflows gain the greatest benefit.
In other words, K3 is not necessarily designed for users who send unrelated one-off prompts.
It is designed for developers and enterprises that can build persistent, cache-friendly workflows around:
Large repositories;
Long documents;
Continuous agents;
Repeated tool use;
Multi-stage projects.
Its strategy can be summarized as:
Use a premium list price to filter demand, then use high cache efficiency to retain serious users.
K3 Is Selling Task Completion, Not Cheap Tokens
Kimi K3 is not trying to win the market by offering the lowest price per token.
It is trying to convince users to evaluate a different metric:
How much does it cost to complete the entire task?
A model with inexpensive output can still be costly if it:
Requires repeated prompting;
Loses context;
Fails midway through an agent workflow;
Produces incorrect code;
Needs several retries;
Requires human intervention.
A more expensive model may be economically preferable if it completes the task in fewer rounds and reduces engineering rework.
This is why K3’s premium output price may still appeal to users running:
Long-horizon coding agents;
Multi-file refactoring;
Complex research;
Large document processing;
End-to-end knowledge workflows.
Moonshot AI is effectively arguing that intelligence should be priced by delivered results, not only by generated token volume.
MiMo-V2.5 Is Fighting a Different Battle
At the other end of the market, Xiaomi’s MiMo-V2.5 series is using aggressive discounts to lower the cost of entry.
A cache-hit input price of RMB 0.025 per million tokens is not intended to signal exclusivity.
It signals scale.
This kind of pricing is particularly attractive for:
High-volume classification;
Data extraction;
Search preprocessing;
Embedding-related workflows;
Simple code generation;
Customer-service automation;
Repeated batch requests;
Cost-sensitive consumer applications.
In these scenarios, the model does not always need to solve the most difficult problem.
It needs to process enormous volumes of requests quickly, consistently, and cheaply.
MiMo’s pricing strategy therefore appears designed to:
Attract developers;
Increase API traffic;
Expand ecosystem adoption;
Move more workloads onto Xiaomi’s infrastructure;
Build volume before margins.
K3 is using premium pricing to defend the value of complex intelligence.
MiMo is using floor pricing to expand the market for routine intelligence.
One is optimizing revenue per difficult task.
The other is optimizing adoption and throughput.
They are not truly selling the same product.
The 4,000× Gap Is a Sign of Market Maturity
A wide pricing band is often interpreted as market disorder.
In reality, it can be a sign that the market is becoming more mature.
Early markets usually have simple pricing because products are still treated as interchangeable.
As the market develops, price differences begin to reflect real service differences.
AI APIs are now being divided according to:
Intelligence Level
A lightweight extraction model and a frontier coding agent do not generate the same economic value.
Input Type
Cache-hit input, uncached input, images, audio, and long-context documents carry different infrastructure costs.
Output Complexity
Short structured answers cost less to produce than long reasoning chains, code changes, or agent plans.
Latency and Priority
A batch task that can wait several hours should not be priced like a real-time production request.
Reliability
Guaranteed throughput and service-level commitments carry a premium.
Task Completion
A more expensive model may create lower total costs if it requires fewer retries and less human correction.
The market is therefore moving away from one universal question:
“Which model has the lowest token price?”
It is moving toward a more useful question:
“Which model offers the lowest total cost for this specific task?”
The Real Problem Is No Longer Model Selection—It Is Routing
When prices range from RMB 0.025 to RMB 100 per million tokens, no rational developer should send every task to the same model.
A mature AI workload contains many different types of requests:
Simple classification;
Document extraction;
Code completion;
Long-context analysis;
Agent planning;
Complex debugging;
Final high-risk review.
The economically optimal model is different for each task.
For example:
A low-cost model may handle tagging and extraction;
A mid-tier model may process routine development work;
A flagship model may perform architecture analysis;
A cache-optimized model may handle long-running repository agents;
A specialist model may process coding or multilingual requests.
The main challenge is therefore not finding one model that is cheapest or strongest.
It is building a system that can decide:
Which model should receive each request;
Whether cached context can be reused;
How much latency the task can tolerate;
When to escalate to a stronger model;
When to switch suppliers;
How to keep the completed-task cost under control.
This is where model aggregation and intelligent routing begin to matter.
Unexhub: Turning the Fragmented Price Band Into a Tradable Market
According to its public positioning, Unexhub is building a computing-power trading center for token aggregation and distribution.
The platform aims to bring together AI suppliers across different scales, including access to ecosystems such as:
Gemini;
ChatGPT;
Claude;
DeepSeek;
GLM;
Models supplied by individual developers and small or medium-sized companies.
In a market where different models have radically different prices, capabilities, and availability, aggregation can provide more than technical convenience.
It can create a unified layer for:
Price comparison;
Model selection;
Traffic routing;
Usage measurement;
Supplier distribution;
Financial settlement.
Instead of requiring developers to understand and maintain every provider’s pricing system, the platform can route requests according to the task’s actual requirements.
Millisecond-Level Routing: Do Not Pay RMB 100 for a RMB 0.025 Task
The clearest value of intelligent routing is avoiding capability waste.
A simple extraction request does not need to be sent to the most expensive frontier model.
Likewise, a complex multi-file coding task should not automatically be assigned to the cheapest model if repeated failures will eventually cost more.
Unexhub emphasizes millisecond-level routing, allowing requests to be matched according to factors such as:
Model capability;
Current price;
Cache availability;
Response latency;
Supplier stability;
Workload type;
Context length;
Traffic pressure.
The routing logic can be simple in principle:
Use economical models for standardized work;
Use stronger models for difficult reasoning;
Use cache-efficient routes for repeated long contexts;
Use priority capacity for urgent production tasks;
Use lower-priced capacity for flexible batch jobs.
In a market with a 4,000× headline price spread, routing is no longer an optional optimization.
It becomes a central part of cost control.
Developer Benefits: Pay Only for the Capability the Task Needs
Through Unexhub, developers can pay according to actual token consumption rather than purchasing fixed infrastructure or maintaining separate prepaid balances across multiple providers.
Depending on the model, supplier route, and market conditions, some services may be available at prices as low as approximately 30% of prevailing market rates.
For high-volume developers and enterprise teams, the value is not limited to one discounted request.
It can reduce spending across:
Batch processing;
Coding agents;
Long-context analysis;
Customer-service automation;
Document workflows;
Multi-model testing;
High-frequency production calls.
More importantly, developers gain the flexibility to combine different price tiers inside the same application.
They can reserve premium models for the tasks where premium capability truly produces value, while routing routine work through lower-cost services.
The discount therefore becomes more meaningful when paired with:
Unified access;
Dynamic routing;
Transparent usage;
Model fallback;
Centralized billing.
API Bidding: Let Supply and Demand Determine the Price
Unexhub also proposes an API bidding mechanism that allows prices to move dynamically with market supply and demand.
When suppliers have idle capacity:
Developers may receive lower prices;
Batch jobs can be processed more economically;
Suppliers can monetize unused GPUs;
Platform utilization improves.
When demand rises:
Prices may increase;
Suppliers receive incentives to release additional capacity;
Urgent workloads can obtain priority;
Scarce compute moves toward higher-value tasks.
This brings an additional dimension to the K3–MiMo price spectrum.
The final API price does not need to be determined only by the model’s official list rate.
It can also reflect:
Available capacity;
Time of day;
Route quality;
Supplier competition;
Required latency;
Service guarantees.
Tokens begin to behave not only as billing units, but as dynamically priced compute resources.
Supplier Benefits: One Integration, Wider Distribution
The platform’s value proposition is different for model and compute suppliers.
A supplier may have:
A competitive model;
Specialized domain capability;
Idle inference capacity;
Regional infrastructure;
Low-cost electricity.
But it may lack:
Developer traffic;
Enterprise sales channels;
Billing infrastructure;
Global distribution;
Settlement systems;
Multi-tenant API management.
Through a single integration, suppliers can potentially make their capacity available to a broader market.
Unexhub’s proposed real-time settlement model further shortens the path from providing capacity to receiving revenue:
Connect once, receive routed requests, and settle according to actual token usage.
For smaller suppliers, this can reduce customer-acquisition and cash-flow pressure.
For larger suppliers, it offers another channel for monetizing idle or off-peak capacity.
Early qualifying suppliers may also benefit from zero-commission incentives, lowering the cost of testing the platform, establishing transaction records, and building an initial market position.
From List Prices to Task Economics
The dramatic contrast between Kimi K3 and Xiaomi MiMo-V2.5 demonstrates that AI pricing can no longer be understood through a single list price.
A complete cost calculation must include:
Input price;
Output price;
Cache-hit rate;
Average output length;
Number of agent rounds;
Retry frequency;
Human intervention;
Latency requirements;
Task completion rate.
Consider two simplified workflows.
Workflow A: High-Volume Extraction
The task is standardized, the output is short, and errors are easy to detect.
A low-cost model may provide the best economics.
Workflow B: Multi-Hour Coding Agent
The model must understand a repository, modify several files, run tests, and repair failures.
A premium model with better long-horizon performance and high cache reuse may provide a lower completed-task cost, despite a much higher listed output price.
The correct model is determined by the workload—not by the most attractive number on a pricing page.
The Future Belongs to Model Portfolios, Not Single-Model Loyalty
The 4,000× headline price gap signals the end of the one-model-for-everything strategy.
Developers and enterprises will increasingly build model portfolios.
A typical production system may use:
Ultra-low-cost models for classification and extraction;
Balanced models for routine development and documents;
Cache-efficient models for persistent long-context agents;
Frontier models for difficult reasoning and final review;
Specialized models for coding, vision, audio, or local languages.
The competitive advantage will come from how these models are orchestrated.
The winning company may not be the one that always chooses the cheapest model or always chooses the strongest model.
It may be the one that routes every request to:
The lowest-cost model capable of completing the task reliably.
Final Thoughts: The Market Is No Longer Fighting Over One Price
Kimi K3 and Xiaomi MiMo-V2.5 are not simply two companies offering different discounts.
They represent two different theories of AI value.
K3 argues that difficult intelligence, long-context execution, and task completion deserve a premium.
MiMo argues that routine intelligence should become nearly free and operate at massive scale.
Both strategies can be correct.
The important change is that the market is now wide enough to support both.
A 4,000× headline price gap may look extreme, but it reflects a deeper reality:
Tokens are no longer a single product.
They represent different combinations of intelligence, latency, reliability, context, and computing cost.
In this environment, the greatest value may not belong to the model with the highest or lowest list price.
It may belong to the platform that can connect all of them, compare them, route them, and convert fragmented capacity into one usable market.
That is the opportunity Unexhub is pursuing through model aggregation, API bidding, millisecond-level routing, real-time supplier settlement, and early zero-commission incentives.
For developers, this means paying according to actual token usage and potentially accessing suitable routes at prices as low as approximately 30% of prevailing market levels.
For suppliers, it means connecting once, gaining distribution, and recovering revenue faster.
The next stage of the AI price war will not be won by cutting every model to the same price.
It will be won by making a market in which every task can find the right model, the right capacity, and the right price.
中文版本
从每百万 Token 100 元到 0.025 元:大模型 API 撕开了一条 4000 倍价格带
2026 年 7 月,AI 大模型市场呈现出一幅极具戏剧张力的价格图景。
同一周内,Kimi K3 将 API 输出价格定在了:
每百万 Token 100 元。
另一边,小米 MiMo-V2.5 系列则打出最高 99% 的降幅,缓存命中输入价格低至:
每百万 Token 0.025 元。
一边是 100 元。
一边是 2 分 5 厘。
两组最醒目的价格数字之间,相差整整 4000 倍。
严格来说,这并不是一次完全同口径的比较:100 元对应输出价格,而 0.025 元对应享受优惠后的缓存命中输入价格。
但这并不影响它释放出的信号,反而更加清晰地说明:
大模型价格已经不再沿着一条“从贵到便宜”的单一曲线变化。
市场正在按照以下因素,裂变出一条极宽的价格带:
模型能力;
输入与输出类型;
缓存利用率;
上下文长度;
响应延迟;
任务复杂度;
服务优先级;
实时算力供给;
目标用户群体。
这已经不再是一场传统意义上的价格战。
而是一个全新分层市场的形成。
Kimi K3 为什么敢把输出卖到 100 元?
7 月 16 日,月之暗面正式发布 Kimi K3。
根据公开参数,K3 拥有:
2.8 万亿总参数;
896 个专家,每次推理激活 16 个;
100 万 Token 上下文窗口;
面向长程编程与端到端知识工作的能力设计。
其 API 定价也非常鲜明:
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。
相比前代旗舰 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。
在全球开源模型中,这是一份相当有自信的价目表。
它的底气来自哪里?
第一层底气来自架构
理解 K3 的价格,不能只看 2.8 万亿参数这个数字。
K3 的架构,试图将总参数规模与每次推理的实际成本进行解耦。
模型虽然拥有 2.8 万亿参数,但每次推理只会激活 896 个专家中的 16 个。
完整参数为模型提供更宽广的知识与能力基础,而每次请求只调用其中一部分能力。
K3 同时采用 KDA 线性注意力机制。据公开信息,在百万 Token 级上下文中,其解码效率最高可提升 6.3 倍。
AttnRes 架构则希望以不到 2% 的额外算力,带来约 25% 的训练效率提升。
关键并不只是模型很大。
而是月之暗面正在通过稀疏激活、长上下文优化和推理架构改进,让超大模型具备可商业化调用的可能。
传统认知是:
参数越多,推理成本必然按比例增加。
K3 正试图打破这一关系。
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。
如果这种单任务经济性能够在真实生产环境中保持,那么每百万输出 Token 100 元的价格,可能并没有第一眼看起来那么昂贵。
挂牌单价很高,但一个任务究竟需要生成多少 Token、能否一次完成,同样影响最终成本。
第二层底气来自缓存经济
K3 定价策略中,更值得关注的可能是 Mooncake 分离式推理架构。
在许多编程和企业工作流中,大量上下文会被反复使用:
代码仓库;
系统提示词;
产品文档;
数据库结构;
长对话历史;
智能体规则;
工具定义。
如果每次请求都必须从头处理这些内容,输入成本会迅速增加。
月之暗面宣称,在编程场景中,缓存命中率可以超过 90%。
在这种条件下,开发者的实际输入成本可能远低于缓存未命中的挂牌价。
一个工作流按照每百万 Token 20 元的未命中价格计算,看起来可能十分昂贵;但如果大部分重复上下文都按照每百万 Token 2 元计费,最终账单便会明显改变。
因此,K3 的价格结构非常明确:
首次输入成本较高;
重复上下文价格明显降低;
输出能力带有溢价;
长周期、稳定运行的工作流获得最大收益。
换句话说,K3 并不一定适合大量彼此无关的一次性请求。
它更希望吸引能够围绕以下场景构建长期工作流的开发者与企业:
大型代码仓库;
超长文档;
持续运行的智能体;
高频工具调用;
多阶段项目。
其定价策略可以概括为:
用较高挂牌价筛选用户,再用高缓存命中率留住真正的高价值工作负载。
K3 卖的不是便宜 Token,而是任务完成能力
Kimi K3 并没有试图依靠最低 Token 单价赢得市场。
它试图让用户采用另一项指标:
完成整个任务究竟需要花多少钱?
一款输出价格很低的模型,如果存在以下问题,最终成本仍可能很高:
需要重复提示;
容易丢失上下文;
智能体任务中途失败;
生成错误代码;
需要多次重试;
需要人工频繁接管。
相反,一款单价较高的模型,如果能够用更少轮次完成任务,减少工程师返工,反而可能更加经济。
这也是 K3 的高输出价格仍然可能吸引以下用户的原因:
长程编程智能体;
多文件重构;
复杂研究任务;
大型文档处理;
端到端知识工作。
月之暗面真正希望传递的观点是:
智能能力不应只按生成了多少 Token 定价,也应按照最终交付了多少结果衡量。
MiMo-V2.5 打的是另一场战争
在市场的另一端,小米 MiMo-V2.5 系列通过极致降价,进一步降低模型调用门槛。
缓存命中输入低至每百万 Token 0.025 元,并不是为了制造稀缺感,而是为了追求规模。
这类价格尤其适合:
高频内容分类;
数据抽取;
搜索预处理;
向量相关工作流;
简单代码生成;
客服自动化;
批量重复请求;
对成本敏感的消费级应用。
在这些场景中,模型不一定需要解决最困难的问题。
它更需要快速、稳定、低成本地处理海量请求。
因此,MiMo 的定价逻辑更接近:
吸引开发者;
扩大 API 调用量;
建立生态覆盖;
将更多任务迁移至小米基础设施;
先扩大规模,再寻找利润。
K3 通过高价维护复杂智能的价值。
MiMo 通过地板价扩大日常智能的使用规模。
一个追求高难度任务的单次价值。
另一个追求标准化任务的调用规模。
它们看似面对同一批开发者,实际上销售的并不是完全相同的产品。
4000 倍价格带,可能是市场成熟的标志
极宽的价格差距,经常会被视为市场混乱。
但从另一个角度看,它也可能说明市场正在成熟。
市场发展早期,产品尚未充分分化,价格通常较为简单。
随着需求逐渐清晰,不同价格开始对应真实的服务差异。
大模型 API 正在按照以下维度分层。
智能水平
轻量抽取模型与前沿编程智能体创造的商业价值并不相同。
输入类型
缓存命中输入、普通输入、图像、音频和长文档,对基础设施的要求不同。
输出复杂度
短结构化回答,与长推理链、代码修改和智能体规划,成本并不相同。
延迟与优先级
可以等待数小时的批量任务,不应该与实时生产请求使用相同价格。
服务稳定性
保证吞吐、延迟和 SLA 的线路,通常需要额外溢价。
任务完成率
更昂贵的模型如果能够减少重试和人工修正,最终总成本可能更低。
因此,市场关注点正在从:
“哪个模型的 Token 单价最低?”
转向:
“针对这个具体任务,哪个模型的总完成成本最低?”
真正的难题不再是选模型,而是做路由
当价格从每百万 Token 0.025 元延伸到 100 元,理性的开发者不应该把所有请求都交给同一个模型。
一个成熟的 AI 应用,通常同时包含多种任务:
简单分类;
文档抽取;
代码补全;
长上下文分析;
智能体规划;
复杂调试;
高风险最终审核。
不同任务,对应的最优模型并不相同。
例如:
标签和抽取任务使用低成本模型;
常规开发任务使用中端模型;
架构分析交给旗舰模型;
长期代码仓库智能体选择高缓存效率模型;
编程或多语言任务使用专业模型。
真正的难题,已经不再是找到一个最便宜或最强的模型。
而是建立一套系统,持续判断:
每次请求应该交给哪个模型;
是否可以复用缓存;
任务能够容忍多少延迟;
什么时候需要升级至更强模型;
什么时候应该切换供应商;
如何控制每个任务的最终成本。
这正是模型聚合与智能路由开始体现价值的地方。
Unexhub:将割裂的价格带变成可交易市场
根据公开定位,Unexhub 正在建设一个 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,覆盖:
Gemini;
ChatGPT;
Claude;
DeepSeek;
GLM;
个人开发者与中小企业提供的模型。
在模型价格、能力和可用性高度分化的市场中,聚合平台所提供的已经不只是接入便利。
它可以形成一层统一的:
价格比较;
模型选择;
流量路由;
用量计量;
供应商分发;
资金结算。
开发者无需逐一研究并长期维护每一家模型厂商的计费规则,平台可以根据任务的真实要求完成线路分配。
毫秒级路由:不要用 100 元的模型,处理 0.025 元就能完成的任务
智能路由最直接的价值,是避免能力浪费。
一个简单的数据抽取任务,没有必要发送给最昂贵的前沿模型。
同样,一个复杂的多文件编程任务,也不应因为追求最低单价而交给屡次失败、最终反而增加总成本的模型。
Unexhub 强调毫秒级路由,请求可以按照以下因素进行匹配:
模型能力;
实时价格;
缓存可用性;
响应延迟;
供应商稳定性;
任务类型;
上下文长度;
当前流量压力。
基本逻辑可以概括为:
标准化任务走经济型模型;
高难度推理使用更强模型;
重复长上下文选择缓存效率更高的线路;
紧急生产任务使用优先算力;
弹性批量任务选择低价容量。
在一条达到 4000 倍的价格带中,路由不再只是可选优化。
它已经成为成本管理的核心环节。
开发者权益:只为任务真正需要的能力付费
通过 Unexhub,开发者可以按照真实 Token 用量付费,而不必自建固定基础设施,也无需在不同厂商处分别维护大量预付余额。
根据具体模型、供应线路和市场条件,部分服务价格最低可达到常规市场价格的约 3 折。
对高频开发者和企业团队而言,价值并不只体现在某一次请求获得折扣。
它可以持续降低:
批量数据处理成本;
编程智能体费用;
长上下文分析支出;
客服自动化预算;
文档工作流成本;
多模型测试费用;
高频生产调用支出。
更重要的是,开发者可以在同一个应用中组合不同价格等级。
真正需要高端能力的任务,才调用高价模型;大量常规工作,则通过成本更低的线路完成。
当价格优惠与以下能力结合时,其价值会更加明显:
多模型统一接入;
动态路由;
透明用量;
故障切换;
集中账单。
API 竞价:让供需决定最终价格
Unexhub 还提出 API 竞价机制,让价格根据市场供需动态变化。
当供应商存在闲置容量时:
开发者可以获得更低价格;
批量任务能够以更低成本运行;
供应商可以将闲置 GPU 转化为收入;
平台整体利用率得到提高。
当需求上升时:
价格可以适度上涨;
供应商获得增加供给的动力;
紧急任务能够购买优先级;
稀缺算力流向价值更高的场景。
这让 K3 与 MiMo 之间的价格带增加了新的变量。
开发者最终支付的价格,不必只由模型官方价目表决定,还可以受到以下因素影响:
实时可用容量;
使用时段;
路由质量;
供应商竞争;
延迟要求;
SLA 保障。
Token 由静态计费单位,逐渐演变为动态定价的算力资源。
供应商权益:一次接入,获得更广泛分发
对模型和算力供应商而言,平台提供的是另一种价值。
一家供应商可能拥有:
有竞争力的模型;
专业领域能力;
闲置推理算力;
区域基础设施;
较低电力成本。
但它可能缺少:
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。
通过一次接入,供应商可以将自身产能开放给更广泛的市场。
Unexhub 提出的实时结算机制,进一步缩短了从提供算力到获得收入的路径:
一次接入、接受平台路由,再按照实际 Token 调用完成收益结算。
对中小供应商而言,这能够降低获客和现金流压力。
对大型供应商而言,则增加了一条消化闲置或低谷算力的渠道。
符合条件的早期供应商还可享受零佣金激励,以更低成本测试平台、积累交易记录并提前建立市场位置。
从挂牌单价,走向任务经济学
Kimi K3 与小米 MiMo-V2.5 的强烈对比说明,大模型价格已经无法通过一个简单数字来理解。
完整的成本核算至少需要包括:
输入价格;
输出价格;
缓存命中率;
平均输出长度;
智能体交互轮次;
重试次数;
人工接管成本;
延迟要求;
任务完成率。
可以对比两类典型工作流。
工作流 A:高频数据抽取
任务高度标准化,输出较短,错误容易识别。
此时,低成本模型通常更加经济。
工作流 B:长时间编程智能体
模型需要理解代码仓库、修改多个文件、运行测试并处理失败。
此时,具备更强长程执行能力和更高缓存利用率的高价模型,即使输出挂牌价更高,也可能带来更低的单任务完成成本。
正确的模型取决于具体任务,而不是价目表上最吸引眼球的数字。
未来属于模型组合,而不是单一模型忠诚
4000 倍的价格差距,意味着“一种模型处理所有任务”的时代正在结束。
开发者和企业将越来越多地建立模型组合。
一套典型生产系统可能同时使用:
超低成本模型处理分类与抽取;
均衡型模型处理日常开发与文档;
高缓存效率模型运行持续长上下文智能体;
前沿旗舰模型负责困难推理与最终审核;
专业模型处理编程、视觉、音频或本地语言。
竞争力将来自这些模型如何被组织和调度。
最终赢家未必是始终使用最便宜模型的公司,也未必是所有任务都调用最强模型的公司。
而可能是能够将每个请求交给:
足以可靠完成任务的最低成本模型。
写在最后:市场争夺的已不再是同一个价格
Kimi K3 与小米 MiMo-V2.5,并不只是两家公司推出了不同力度的价格方案。
它们代表的是两种不同的 AI 价值观。
K3 认为,复杂智能、长上下文执行与任务完成能力,值得支付溢价。
MiMo 则认为,日常标准化智能应该接近免费,并通过巨大规模运行。
两种策略都可能成立。
真正重要的变化是,当前市场已经足够广阔,能够同时容纳它们。
4000 倍价格差距看似夸张,却揭示了一个更深层事实:
Token 已经不再是一种单一产品。
不同 Token 背后,是不同的智能能力、延迟、稳定性、上下文和算力成本组合。
在这样的市场中,最大价值未必属于挂牌价最高或最低的模型。
而可能属于能够将它们连接起来、进行比较、完成路由,并将分散算力转化为统一市场的平台。
这正是 Unexhub 希望通过模型聚合、API 竞价、毫秒级路由、供应商实时结算与早期零佣金激励抓住的机会。
对开发者而言,这意味着可以按实际 Token 用量付费,并在合适的模型和线路下,获得最低约为常规市场价格 3 折的服务。
对供应商而言,则意味着一次接入、获得分发,并更快完成资金回流。
下一阶段的大模型价格战,不会通过把所有模型压到同一个价格而结束。
它将通过建立一个市场来完成:
让每一种任务,都能找到合适的模型、合适的算力与合适的价格。