不止看 Token 单价!AI 算力交易中心的缓存命中、Batch 批次、长上下文三层计价,让 Sol 实际成本直降 70%

决定 AI 成本水位的,不只是 Token 单价,还有缓存命中率。


70% Isn’t a Discount — It’s What Happens When You Calculate the Bill Correctly

GPT-5.6 Sol is available at half price on OpenRouter: input pricing falls from $5 to $2.50 per million tokens, while output pricing drops from $30 to $15.

The headline is obvious:

50% off.

But the real AI cost breakthrough is not hidden in those two words.

It is hidden in how the bill is calculated.

On August 17, DeepSeek introduced peak/off-peak pricing that made this point particularly clear. For V4 Pro, cached input during peak periods was priced at RMB 0.30 per million tokens, while uncached input was RMB 9.00.

That means:

The same model can have a 30× price difference depending on whether the input is cached.

Thirty times.

That is no longer simply a discount.

It is a different definition of what AI inference actually costs.


Layer One: Cache Hits — Turning Repeated Computation Into Floor Pricing

One of the most expensive stages of large-model inference is prefill.

The model needs to process the input context and construct the KV cache before generating an answer.

If the same input context is repeatedly processed from scratch, the computing cost is effectively paid again and again.

But when a subsequent request shares the same prefix as an earlier request, the model can reuse the previously calculated KV cache.

This creates an important cost distinction:

Cached input is fundamentally different from uncached input.

DeepSeek has made that distinction explicit in its pricing structure.

For example, DeepSeek V4 Flash can price cached input during off-peak periods at around RMB 0.05 per million tokens, compared with RMB 1.50 for uncached input.

That is another 30× difference.

Alibaba Cloud’s Qwen 3.7-Flash similarly offers substantially lower pricing for cached input, with cached tokens priced at only a small fraction of the standard uncached rate.

The lesson is simple:

Your cache-hit rate can determine your AI cost floor.

This is particularly important for agent workloads.

AI agents repeatedly send:

  • System instructions

  • Tool definitions

  • Conversation history

  • Context summaries

  • Repeated workflow information

A large portion of that content may remain unchanged between calls.

If stable components such as system prompts, tool definitions, and historical summaries are structured as cacheable prefixes, increasing the cache-hit rate can significantly reduce the bill.

In real-world workloads, teams have reported substantial cost reductions through prefix-aware scheduling and semantic caching, without sacrificing output quality.

The important point is:

Caching does not make the model less capable.

It simply prevents you from paying repeatedly for computation that can be reused.


Layer Two: Batch Processing — Give Non-Urgent Work a 50% Discount

The second optimization is much simpler:

If the task does not need an immediate response, batch it.

Batch APIs are designed for exactly this type of workload.

Many AI tasks do not require millisecond-level responses:

  • Code refactoring

  • Large-scale testing

  • Data annotation

  • Document processing

  • Model evaluation

  • Offline agent experiments

These workloads can often be placed into a batch queue and processed later.

In exchange for accepting higher latency, providers can offer substantially lower token prices.

Alibaba Cloud, for example, has stated that supported batch workloads can be charged at 50% of standard real-time inference pricing.

Google Gemini also offers batch processing at discounted rates, with processing windows that can extend to 24 hours.

OpenRouter’s GPT-5.6 Sol batch pricing similarly provides another layer of savings, with rates around:

  • $1.25 per million input tokens

  • $7.50 per million output tokens

The model has not changed.

The output quality has not changed.

Only the execution mode has changed.

That distinction matters.

A company does not always need to optimize the model.

Sometimes it simply needs to optimize when and how the model is called.


Cache + Batch: The Savings Can Stack

The more interesting part is that these optimization layers can work together.

A request can potentially be:

cache-optimized + batch-processed

at the same time.

First, reuse existing computation through caching.

Then, move non-urgent workloads into a discounted batch queue.

The result is a fundamentally different cost structure from ordinary real-time, uncached inference.

This is why sophisticated AI cost optimization should not focus on one discount.

It should focus on the entire inference pipeline.

Instead of asking:

“Which model is cheapest?”

the better question is:

“What is the cheapest way to execute this workload while maintaining the required quality and latency?”

That is a much more powerful way to think about AI economics.


Layer Three: Long Context — Cross the Threshold and the Price Jumps

This is one of the most overlooked cost variables.

GPT-5.6 Sol’s standard $5/$30 pricing applies only within the specified standard input range.

Once requests cross the relevant context threshold, pricing can move into a higher tier.

In the pricing structure described here, requests beyond 32K input tokens move to approximately:

  • $10 per million input tokens

  • $45 per million output tokens

The implication is important:

Long-context inference carries a real economic premium.

The same model can effectively have different economics depending on how much context you ask it to process.

Why?

Because longer contexts require:

  • More prefill computation

  • Larger KV caches

  • Greater memory utilization

  • More inference resources

Long context is therefore not simply a technical feature.

It is also a cost variable.

The practical strategy is to route workloads according to their actual requirements.

Short conversations should use the standard path.

Long-document analysis should use long-context capacity only when necessary.

Sending every request through a long-context route is effectively paying a premium for capacity you may not need.


Three Layers Combined: Where Does the 70% Come From?

Now put the three mechanisms together.

1. Cache optimization

For repeated prefixes, cached input can reduce input costs dramatically—depending on the provider and pricing tier, the difference can reach 80%-96.8%.

2. Batch processing

For non-urgent workloads, batch execution can reduce input and output pricing by roughly 50% compared with standard real-time inference.

3. Context-aware routing

Short requests stay on standard pricing.

Only workloads that actually require long-context capacity pay the corresponding premium.

The result?

Under the right workload structure, the effective bill can potentially fall by 70% or more without changing the underlying model.

This is the key distinction:

70% is not necessarily a headline discount offered by a model provider.

It can be the result of stacking several independent optimization mechanisms.

One practical workload review described a combination of:

  • Cache optimization reducing repeated-call costs

  • Batch processing moving a significant portion of offline traffic into discounted execution

  • Better routing reducing unnecessary retries and expensive execution

Together, these changes produced an effective cost level of roughly 70% of the original baseline.

In other words:

The 70% figure is not negotiated. It is calculated.


The Compute Trading Center: Turning Optimization Into Infrastructure

This is where Unexhub’s model becomes particularly relevant.

Unexhub is not positioned simply as a cheaper API reseller.

Its broader goal is to turn AI cost optimization into default infrastructure at the compute trading layer.

Today, enterprises often need to solve these problems themselves.

They have to:

  • Understand each provider’s caching rules

  • Design prefix structures

  • Build batch schedulers

  • Monitor Token consumption

  • Detect context thresholds

  • Create routing logic

  • Optimize supplier selection

  • Track pricing changes

In other words:

The customer becomes its own AI cost-optimization department.

Unexhub aims to move much of this complexity into the platform layer.

At the Token aggregation and distribution level, optimization mechanisms such as:

  • Cache-aware routing

  • Batch scheduling

  • Context-aware model selection

  • Supplier competition

  • Dynamic API pricing

can become part of the infrastructure rather than another engineering project for the customer.

The objective is simple:

Developers should not need to become AI pricing experts just to obtain efficient inference costs.


Beyond Cost Optimization: Real-Time Settlement and Supplier Competition

Cost optimization is only one side of the compute market.

The supply side matters just as much.

Unexhub aims to connect AI suppliers of different sizes, from major model ecosystems to independent developers and small and medium-sized AI companies.

The ecosystem direction includes models and providers across categories such as:

  • Gemini

  • GPT / ChatGPT

  • Claude

  • DeepSeek

  • GLM

alongside specialized models developed by smaller teams.

The larger the supplier network becomes, the more room there is for:

  • Price competition

  • Workload matching

  • Capacity utilization

  • Model specialization

  • Dynamic routing

Unexhub also plans a real-time settlement mechanism, allowing suppliers to receive usage-based revenue according to actual Token consumption.

For eligible early suppliers, the platform offers 0% commission incentives, reducing the cost of entering the marketplace and allowing more of the eligible inference revenue to remain with the supplier.

This creates a two-sided economic advantage:

Developers gain more competitive supply.

Suppliers gain more distribution opportunities.


API Bidding: Let the Market Help Set the Price

Traditional AI APIs generally follow a simple model:

Provider sets price → customer accepts price.

A compute trading center introduces another possibility:

Supply competes for demand → market conditions influence price.

When a supplier has excess capacity, it may compete more aggressively.

When demand rises, scarce inference capacity can command higher prices.

This transforms AI inference from a static price catalog into a dynamic marketplace.

For developers, that means more choice.

For suppliers, it means more opportunities to monetize unused capacity.

And for the market as a whole, it creates a mechanism for price discovery.


The Real Developer Benefit: Not Just “Cheaper,” but More Control

This is why the most important Unexhub benefit is not simply another temporary discount.

The larger value is pricing flexibility.

Developers can potentially:

Pay by Actual Token Usage

Use what you need.

Pay for what you consume.

No need to purchase unnecessary capacity simply because a package is available.

Access Multiple Suppliers

Instead of depending on one provider, developers can access a broader supplier ecosystem through a unified marketplace.

Use Intelligent Routing

Requests can be routed according to:

  • Price

  • Latency

  • Availability

  • Model capability

  • Workload requirements

Pursue Lower Market Pricing

Through supplier competition, dynamic bidding, and intelligent routing, eligible scenarios may reach pricing as low as approximately 30% of prevailing market rates.

That is a fundamentally different proposition from waiting for the next vendor promotion.

The discount is not necessarily a single coupon.

It is the result of market competition plus infrastructure optimization.


Conclusion: 70% Is Not a Discount. It Is a Calculation.

The GPT-5.6 Sol half-price story is news.

But the real inflection point is not:

“AI became cheaper.”

It is:

“The way we calculate AI cost is changing.”

For years, developers looked at one number:

Price per million tokens.

Increasingly, that is not enough.

A more complete AI cost model needs to consider:

  1. Cached input pricing

  2. Uncached input pricing

  3. Batch pricing

  4. Context-length premiums

  5. Routing efficiency

  6. Retry waste

  7. Supplier competition

  8. Real-time market pricing

The Token unit price determines the baseline.

But utilization determines the final bill.

That is why a 70% cost reduction is not necessarily something you negotiate with a vendor.

It can be something you calculate through better infrastructure design.

Cache what should be cached.

Batch what does not need to be immediate.

Keep short requests away from expensive long-context routes.

Route workloads toward competitive supply.

And let the market compete for your demand.

That is the logic behind Unexhub’s compute trading center.

Its ambition is to transform AI cost optimization from an enterprise burden into platform infrastructure through Token aggregation and distribution, intelligent routing, dynamic API bidding, supplier competition, and real-time settlement.

For developers, the practical value is straightforward:

Pay by actual Token consumption, access a broader AI supplier ecosystem, use millisecond-level intelligent routing, and pursue market-based pricing that can reach approximately 30% of prevailing rates in eligible scenarios.

For suppliers, the value is equally concrete:

Connect once, access a broader pool of enterprise and developer demand, participate in competitive bidding, receive usage-based revenue through real-time settlement, and benefit from 0% commission incentives if eligible as an early supplier.

The headline may be:

“GPT-5.6 Sol is 50% cheaper.”

But the deeper lesson is:

The biggest AI cost savings may not come from waiting for another 50% discount. They may come from learning how to calculate, route, and trade AI compute more intelligently.

70% is not a discount.

It is what happens when you finally calculate the bill correctly.

70% 不是降价,是“算对了账”

GPT-5.6 Sol 在 OpenRouter 上半价。

输入价格从 5 美元 / 百万 Token 降到 2.5 美元,输出价格从 30 美元降到 15 美元。

媒体为这场“腰斩”式降价欢呼。

但真正值得关注的成本杀手,并不只是“半价”两个字。

真正决定企业最终账单的,是:

你到底是怎么调用模型的。

8 月 17 日,DeepSeek 推出峰谷定价之后,这一点变得更加明显。

以 V4 Pro 为例,缓存命中与缓存未命中的输入价格存在巨大差异:

  • 高峰时段缓存命中输入:0.30 元 / 百万 Token

  • 缓存未命中输入:9.00 元 / 百万 Token

同一个模型。

同样的 Token。

仅仅因为是否命中缓存:

价格可以相差 30 倍。

30 倍。

这已经不只是“降价”。

而是在重新定义:

到底什么才是真正的 AI 成本。


第一层:缓存命中——把重复计算变成地板价

大模型推理中,一个重要的成本环节叫 Prefill。

简单来说,就是模型需要先读取输入上下文,并完成相应计算,建立 KV Cache,然后才能继续生成输出。

如果每一次调用都把完全相同的上下文重新计算一遍:

同一份计算成本,就会被反复支付。

而如果后续请求与之前的请求拥有相同或高度重复的前缀,模型就可以复用已经计算好的 KV Cache。

于是:

缓存命中与缓存未命中,就产生了完全不同的成本结构。

DeepSeek 已经通过定价机制明确体现了这一点。

例如,DeepSeek V4 Flash 在空闲时段的缓存命中输入价格可以低至:

0.05 元 / 百万 Token

而缓存未命中输入价格约为:

1.5 元 / 百万 Token

两者相差:

30 倍。

阿里云 Qwen 3.7-Flash 等模型同样提供明显低于标准输入价格的缓存命中价格。

这说明一个非常重要的问题:

决定 AI 成本水位的,不只是 Token 单价,还有缓存命中率。


为什么智能体尤其需要关注缓存?

因为 AI Agent 的工作方式天然会产生大量重复上下文。

一次 Agent 工作流中,可能反复发送:

  • System Prompt

  • 工具定义

  • 历史对话

  • 上下文摘要

  • 工作流规则

  • 项目背景信息

其中大量内容其实并没有发生变化。

如果这些稳定内容能够被组织成可缓存前缀,那么每一次调用就不需要重新为相同计算付费。

例如:

系统提示词不变。

工具定义不变。

项目背景不变。

历史摘要只增加少量新内容。

那么真正需要重新计算的,可能只是新增部分。

因此,缓存命中率越高:

最终账单越低。

一些实际工作负载测试也已经显示,通过前缀聚类、缓存调度和语义缓存等方式,可以在不降低模型输出质量的情况下显著减少成本。

核心逻辑其实非常简单:

缓存不会让模型变笨。

它只是让你不必为已经计算过的内容重复付费。


第二层:批处理——给“不着急”的任务打五折

第二层优化更加直接:

不着急,就批处理。

Batch API 的核心逻辑就是:

接受更长的响应时间。

换取更低的 Token 价格。

很多 AI 工作负载实际上并不需要秒级返回。

例如:

  • 代码重构

  • 大规模测试

  • 数据标注

  • 文档批量处理

  • 模型评测

  • 离线 Agent 实验

  • 大规模数据分析

这些任务完全可以:

提交。

排队。

等待处理。

而不是要求系统实时响应。

阿里云已经明确表示,在支持批量调用的模型中,批量任务的输入和输出 Token 单价可以按照实时推理价格的 50% 收取。

Google Gemini 同样提供 Batch API,并允许部分批处理任务采用更长的处理时间窗口。

OpenRouter 上 GPT-5.6 Sol 的批量价格也进一步下降至约:

  • 输入:1.25 美元 / 百万 Token

  • 输出:7.50 美元 / 百万 Token

注意:

模型没有变。

输出质量没有因为 Batch 而自动下降。

变化的只是:

调用方式。

所以,有时候真正应该优化的并不是模型。

而是:

什么时候调用模型,以及用什么方式调用模型。


缓存 + Batch:成本优化可以叠加

真正有意思的地方在于:

这些优化机制不是互斥的。

一个请求可以同时:

命中缓存 + 使用 Batch。

第一步:

通过缓存复用已经完成的计算。

第二步:

对于不需要实时返回的任务,进入 Batch 通道。

于是:

原本的标准实时、未缓存调用:

可能同时承担两种成本。

而经过优化之后:

重复计算被减少。

实时调用溢价被降低。

这就是为什么 AI 成本优化不能只盯着某一次“打折”。

真正应该优化的是:

整个推理链路。

与其问:

“哪个模型最便宜?”

不如问:

“在满足质量和延迟要求的前提下,这项任务最便宜的执行方式是什么?”

这才是 AI 算力成本管理真正需要解决的问题。


第三层:长上下文——超过阈值,价格可能直接跳档

这是最容易被忽视的一层。

很多开发者看到 GPT-5.6 Sol 的:

5 美元 / 30 美元

就认为所有请求都按照这个价格计算。

实际上,输入上下文长度本身就是重要的成本变量。

按照本文讨论的价格结构:

32K 以内的输入适用标准价格。

超过相应阈值后,价格可能进入更高档位,例如:

  • 10 美元 / 百万输入 Token

  • 45 美元 / 百万输出 Token

也就是说:

同一个模型,处理短对话和处理超长文档,可能对应完全不同的成本。

原因并不复杂。

更长的上下文意味着:

  • 更多 Prefill 计算

  • 更大的 KV Cache

  • 更高的显存占用

  • 更高的推理资源需求

因此:

长上下文不是免费的技术能力。

它本身就是一个成本变量。

真正合理的策略不是:

“所有请求都走最长上下文。”

而应该:

短任务走标准通道。

长文档分析才使用长上下文能力。

如果一份只有几千 Token 的请求,也被放进昂贵的长上下文路径,本质上就是:

主动缴纳不必要的溢价。


三层叠加:70% 到底是怎么来的?

现在,把三个层次放在一起。

第一层:缓存优化

对于重复前缀,缓存命中可以显著降低输入成本。

根据不同供应商和价格档位,缓存命中与未命中的价格差异甚至可以达到 80%-96.8%。

第二层:Batch 优化

对于非实时任务:

Batch 可以让输入和输出价格相比标准实时推理下降约 50%。

第三层:上下文分级

短请求:

不走昂贵的长上下文通道。

长任务:

只有真正需要时才使用长上下文能力。

三个机制叠加之后:

在合适的工作负载结构下,同一个模型、同样的输出质量,最终账单有机会降低 70% 甚至更多。

这里需要特别理解一个概念:

70% 并不一定来自某一家模型厂商直接宣布“70% off”。

它可能来自:

多种独立优化机制叠加后的综合结果。

例如,一组实际工作负载复盘中,通过:

  • 缓存优化降低重复调用成本

  • Batch 将部分离线任务转移到折扣通道

  • 优化路由减少不必要的重试与浪费

最终让综合成本下降到原基准的大约 70%。

也就是说:

“7 折”不是谈出来的,是算出来的。


算力交易中心:把三层优化变成平台基础设施

这正是 Unexhub 希望进一步解决的问题。

Unexhub 并不是简单地做一个:

“更便宜的 API 二道贩子”。

它希望把 AI 成本优化能力直接放到算力交易中心的基础设施层。

今天,一家企业如果想做好 AI 成本管理,往往需要自己研究:

  • 每一家模型厂商的缓存规则

  • Prompt 前缀如何设计

  • Batch API 如何接入

  • Token 长度如何监控

  • 上下文阈值在哪里

  • 什么任务适合离线处理

  • 哪家供应商当前价格更低

  • 哪条路由当前更高效

换句话说:

企业自己变成了 AI 成本优化部门。

而 Unexhub 希望把这部分复杂度下沉到平台。

在 Token 聚合与分发层,将:

  • 缓存感知路由

  • Batch 调度

  • 上下文分级

  • 多供应商竞争

  • API 动态竞价

等能力逐步融入统一的路由与计费体系。

让开发者不需要成为“AI 算价专家”,也能够获得更高效的 Token 使用成本。


不只是降低成本:实时结算让供应商也获得价值

算力市场不是只有需求方。

供应商同样重要。

Unexhub 希望连接不同规模的 AI 供应商:

从大型模型生态,到个人开发者、中小 AI 企业,再到垂直领域模型团队。

生态方向包括:

  • Gemini

  • GPT / ChatGPT

  • Claude

  • DeepSeek

  • GLM

以及更多由个人和中小企业开发的专业 AI 模型。

供应商越丰富:

价格竞争越充分。

资源匹配越灵活。

开发者选择越多。

闲置算力利用率也越有可能得到提升。

与此同时,Unexhub 通过实时结算机制,让供应商按照实际 Token 调用获得对应收入。

对于符合条件的早期供应商,还提供:

0 佣金激励。

这意味着:

早期加入生态的供应商可以降低平台成本,让更多符合条件的调用收入留在供应商侧。

最终形成:

开发者获得更具竞争力的供应。

供应商获得更多分发机会。


API 竞价:让市场参与价格形成

传统 AI API 的逻辑非常简单:

厂商定价 → 用户接受。

算力交易中心则引入另一种可能:

供应竞争需求 → 市场影响价格。

当某个供应商拥有大量闲置算力时:

它可以通过更具竞争力的价格获取订单。

当市场需求突然增加时:

稀缺推理资源则可能获得更高价格。

因此,AI 推理逐渐从:

固定价目表

转向:

动态市场。

对于开发者:

意味着更多选择。

对于供应商:

意味着更多变现闲置算力的机会。

对于整个市场:

意味着真正开始形成价格发现机制。


对开发者而言:真正的优惠不是“便宜一次”,而是拥有更多选择

因此,Unexhub 最值得关注的价值,并不只是一次性的折扣。

更重要的是:

价格选择权。

按 Token 实际付费

使用多少。

支付多少。

不必为了套餐容量购买自己并不需要的资源。

多供应商选择

通过统一生态访问更多模型和供应商,降低对单一模型供应方的依赖。

智能路由

根据:

  • 价格

  • 延迟

  • 供应能力

  • 模型性能

  • 工作负载

寻找更合适的调用路径。

追求更低的市场价格

通过供应商竞争、API 动态竞价和智能路由,在符合条件的场景下,开发者最低可争取到:

约为市场价格 3 折的 AI 调用成本。

这里的意义并不是:

“平台再发一张优惠券。”

而是:

通过市场竞争和基础设施优化,让更低价格成为持续存在的可能。


对供应商而言:从“等待客户”到“主动竞争订单”

AI 供应商真正需要的,并不只是模型本身。

还需要:

分发。

一个优秀模型如果缺乏稳定用户,就很难形成持续收入。

Unexhub 希望提供新的 AI 分发渠道。

供应商完成一次接入后,可以触达:

  • 开发者

  • 企业

  • AI 应用团队

  • 更广泛的市场需求

并按照实际 Token 调用获得收益。

对于符合条件的早期供应商:

0 佣金激励进一步降低生态接入成本。

再配合实时结算:

供应商无需单纯等待传统周期结算。

而可以根据真实调用获得相应收入。

因此,供应商获得的不只是:

一个 API 销售渠道。

而是一套:

流量 + 交易 + 竞价 + 结算

的 AI 商业化基础设施。


结语:70% 不是降价,是“算对了账”

GPT-5.6 Sol 半价当然是新闻。

但真正的行业拐点,并不是:

“AI 变便宜了。”

而是:

“AI 的成本计算方式正在改变。”

过去,开发者看 API 成本,只看一个数字:

每百万 Token 多少钱。

现在,这个数字已经远远不够。

真正完整的 AI 成本模型,需要同时考虑:

  1. 缓存命中价格

  2. 缓存未命中价格

  3. Batch 价格

  4. 长上下文溢价

  5. 路由效率

  6. 重试浪费

  7. 供应商竞争

  8. 实时市场价格

Token 单价只是成本基线。

真正决定最终账单的,是:

资源利用率。

所以,70% 的成本降幅不一定需要去找供应商谈判。

它可以通过:

更合理的计算方式。

更高效的路由方式。

更科学的任务调度。

一点一点“算”出来。

该缓存的缓存。

该 Batch 的 Batch。

短任务不走昂贵的长上下文通道。

能够利用低价供应的请求,就不必坚持走高价路径。

让供应商之间真正竞争订单。

这正是 Unexhub 算力交易中心希望解决的问题:

把企业过去必须自己完成的“算账”,变成平台默认提供的基础设施。

通过 Token 聚合与分发、供应商竞争、API 动态竞价、毫秒级智能路由以及实时结算,Unexhub 正在探索构建更加开放、高效的全球 AI 算力交易市场。

对于开发者而言,实际权益并不只是“打折”。

而是:

按 Token 实际付费、多供应商选择、智能路由,以及在符合条件的场景下争取低至市场价格约 3 折的 AI 调用成本。

对于供应商而言:

一键接入、更广泛的开发者和企业需求、实时获得调用分成,以及符合条件的早期 0 佣金激励。

所以,真正值得记住的不是:

“GPT-5.6 Sol 打五折了。”

而是:

未来 AI 最大的成本优势,可能不是等待下一次 50% 的促销。

而是学会如何缓存、如何批处理、如何路由、如何竞价,以及如何把每一个 Token 都算得更精准。

70% 不是降价。

是算对了账。