Between 0.02 Yuan and 1.5 Yuan: A 75× Gap Reveals Two Different Economies of Compute
Bilingual Edition · English First, Chinese Version Below ↓
If you are reading this in English, keep scrolling—the complete Chinese edition follows after the English article.
In August 2026, the Token market is becoming a real-world laboratory for AI pricing.
DeepSeek V4 Flash can reach a quoted RMB 0.02 per million input tokens for cache-hit requests, while Qwen3.8-Max has been quoted at around RMB 1.5 per million tokens for implicit cache-hit input in the pricing scenario discussed here.
That creates a striking 75× price difference.
But this is not simply a question of which model is cheaper.
It represents a collision between two fundamentally different approaches to the economics of AI inference.
1. DeepSeek's Calculus: Trade Ultra-Low Prices for Massive Volume
DeepSeek's pricing philosophy can be summarized in one sentence:
Push the marginal cost as low as possible, then use volume to make the economics work.
Under the pricing figures discussed here, V4 Flash uses a differentiated structure:
Cache-hit input: RMB 0.02 / million tokens
Cache-miss input: RMB 1 / million tokens
Output: RMB 2 / million tokens
The higher-tier V4 Pro follows a different pricing level, with quoted rates of approximately RMB 0.025, RMB 3, and RMB 6 respectively.
But the more interesting part is not the absolute price.
It is the introduction of peak/off-peak pricing.
During weekday peak periods—roughly 9:00–12:00 and 14:00–18:00—pricing can increase, while lower rates are available during nights, weekends, and holidays.
This turns pricing into a form of compute scheduling.
Instead of simply saying, “AI costs more now,” the system uses price signals to encourage users to shift workloads toward periods when computational resources are less constrained.
At a cache-hit price of RMB 0.02, even a 2× peak multiplier would only bring the figure to RMB 0.04.
The message is obvious:
DeepSeek is trying to monetize time as well as tokens.
And the strategy appears to have generated enormous demand.
The OpenRouter usage data referenced in the original analysis places the DeepSeek V4 family among the world's highest-volume models, with approximately 6.6 trillion tokens of weekly usage in the cited ranking.
That is why the “price butcher” label is more than a marketing slogan.
It is supported by actual demand.
But there is another side to the equation.
On August 6, DeepSeek indicated that it planned significant API pricing adjustments. If prices rise, the market will get a clearer answer to a fundamental question:
How much of DeepSeek's enormous demand is driven by model capability—and how much is driven by price?
2. Qwen3.8-Max's Calculus: Use Regional Pricing to Anchor Value
Alibaba's pricing strategy is fundamentally different.
Qwen3.8-Max has been positioned with separate domestic and international pricing.
The quoted domestic rates are:
RMB 12 / million input tokens
RMB 36 / million output tokens
The quoted international rates are:
US$2 / million input tokens
US$6 / million output tokens
For implicit cache hits, the domestic figure discussed here is approximately RMB 1.5 per million tokens.
There is an interesting detail hidden inside this structure.
RMB 1.5 represents 12.5% of the standard RMB 12 input price.
The corresponding overseas cache price of approximately US$0.25 also represents 12.5% of the US$2 standard input price.
That consistency suggests that caching is being positioned not merely as an infrastructure feature, but as a deliberate pricing mechanism.
The idea is straightforward:
When the system does not need to fully recompute repeated context, the customer receives a substantial discount.
Alibaba is therefore using two layers of pricing simultaneously:
Regional differentiation to adapt to different markets.
Caching discounts to reduce the cost of repeated computation.
At the same time, Qwen3.8-Max is competing aggressively on the capability side.
Based on the performance figures cited in the original analysis, its international output price is only a fraction of Claude Opus 5's quoted price, while its Agentic Index score was cited at 55.4, narrowly ahead of Claude Opus 5 at 55.3.
Whether every benchmark remains consistent over time is less important than the economic signal:
High-end AI capability is becoming available at dramatically different price points.
3. The Real Difference: Resource Scheduling vs. Market Segmentation
Put the two pricing systems side by side, and the difference becomes clearer.
DeepSeek: Pricing as a computational resource scheduler
Peak/off-peak pricing, cache-hit discounts, and multiple model tiers all serve one larger purpose:
maximize the utilization efficiency of expensive GPU resources.
Prices become signals.
When resources are scarce, prices rise.
When resources are abundant, prices fall.
Users are encouraged to move flexible workloads toward cheaper periods.
Qwen: Pricing as a market segmentation mechanism
Qwen's model is different.
Regional prices, cache discounts, and different commercial plans are designed to maximize the model's competitiveness across different customer groups and geographic markets.
The objective is not simply:
“How do we schedule GPUs?”
It is closer to:
“How do we maximize the commercial value of the same intelligence across different markets?”
One approach focuses more heavily on resource efficiency.
The other focuses more heavily on market value.
That is why the 75× price difference is not necessarily a pricing anomaly.
It may actually be the natural result of two different economic strategies.
4. Why a 75× Price Gap Matters
When one model can charge RMB 0.02 for a cache-hit million tokens while another charges around RMB 1.5, the natural question is:
Why would anyone pay the higher price?
The answer is that Token price is only one component of total value.
An enterprise may care about:
reasoning quality
coding reliability
latency
context length
tool use
availability
SLA requirements
geographic coverage
data governance
support
integration cost
A model that costs 75× more on a particular pricing dimension may still be economically rational if it reduces failure rates, engineering overhead, or business risk.
Conversely, for a high-volume classification, extraction, summarization, or batch-processing workload, paying a premium for capabilities that are never used may make little economic sense.
This creates two very different customer profiles.
DeepSeek-style economics favor workloads where cost per token dominates the decision.
Qwen-style economics can appeal to customers seeking a stronger balance between high-end capability and cost.
The real competition is therefore not:
cheap vs. expensive.
It is:
Which price-performance curve is optimal for this workload?
5. Unexhub: Turning the Price Gap into a Market Opportunity
This is where the role of Unexhub becomes interesting.
In the traditional model, developers choose among vendor-specific prices.
You register with one provider, recharge your account, integrate its API, and accept its pricing structure.
Then you repeat the same process for another model.
The price difference exists, but the user has to discover and exploit it manually.
Unexhub's proposition is to put that fragmented supply into a unified marketplace.
Its positioning is:
A Token aggregation, distribution, and compute trading center.
Instead of treating Token pricing as something fixed by individual vendors, Unexhub aims to make price differences part of a broader market-based allocation mechanism.
6. Instant Settlement: Removing Friction on the Supply Side
For AI suppliers, one of the biggest problems in traditional API distribution is cash flow.
A supplier may need to provide resources first and wait for settlement later.
Unexhub emphasizes an instant-settlement model, allowing suppliers to receive their corresponding share based on actual usage.
That can reduce several common barriers:
less capital tied up
less need to stockpile API capacity
shorter settlement cycles
faster conversion of compute into revenue
For smaller AI suppliers in particular, this can be meaningful.
Instead of competing with large providers on capital reserves, they can compete on:
price + quality + availability.
7. API Bidding: Let Supply and Demand Determine the Price
The more ambitious part of the model is API bidding.
Instead of every supplier maintaining one rigid price, market conditions can influence the effective price.
When supply is abundant, developers may find lower-cost capacity.
When demand rises, suppliers can potentially earn higher margins.
This creates a feedback loop:
more supply → more competition → better prices
and potentially:
higher demand → tighter capacity → higher supplier returns.
The goal is to transform Token procurement from a fixed-price purchase into something closer to market-based compute allocation.
8. The 3× Opportunity for Developers
Unexhub also emphasizes a direct economic benefit for developers.
According to the platform positioning described here, developers can:
Pay according to actual Token consumption
Rather than committing to a single provider's ecosystem, users can access multiple AI resources through a unified layer and pay based on usage.
Potentially access prices as low as 30% of prevailing market prices
Subject to the applicable model, supplier, availability, and promotional conditions, Unexhub says developers can access pricing as low as approximately 30% of market prices.
For a low-volume user, the saving may be modest.
But for an enterprise processing millions or billions of tokens, even a relatively small reduction in per-token cost can accumulate into a meaningful annual saving.
The value therefore becomes much clearer for high-frequency users:
The more tokens you consume, the more important routing and procurement efficiency become.
Millisecond-level routing
Unexhub also emphasizes millisecond-level routing, allowing workloads to be directed across different model and supplier resources.
The purpose is not merely speed.
It is to make the best available compute resource easier to reach.
One-click supplier access
For suppliers, Unexhub offers another side of the marketplace equation: a unified distribution channel.
Suppliers can connect their AI resources and reach developers and enterprise customers without having to build an entirely new distribution network from scratch.
Real-time revenue settlement
Usage-based settlement means supplier revenue can be calculated from actual API traffic, helping turn compute capacity into a more predictable and measurable revenue stream.
9. From Vendor Pricing to Market Pricing
The deeper significance of the 75× gap is not that one company is “too expensive” or another is “too cheap.”
It is that AI inference no longer has a single natural price.
The final cost depends on:
which model you use
which provider supplies it
whether caching is triggered
whether demand is at peak
how much capacity is available
how the workload is routed
what commercial terms apply
Once those variables begin interacting, Token starts behaving less like a simple billing unit and more like a market-priced compute resource.
That is the economic space Unexhub is trying to occupy.
Its broader vision is to aggregate AI suppliers of different sizes—including major model ecosystems such as Gemini, ChatGPT, Claude, DeepSeek, and GLM, alongside resources supplied by individuals and smaller businesses—and allow them to compete for developer and enterprise demand.
The more suppliers participate, the more meaningful market-based pricing can potentially become.
10. The 75× Gap Is Only the Beginning
DeepSeek's RMB 0.02 cache-hit price and Qwen3.8-Max's quoted RMB 1.5 implicit cache-hit price reveal something larger than a simple price difference.
They reveal that AI compute has different economic layers.
One provider may optimize GPU utilization.
Another may optimize regional market value.
A third may optimize enterprise reliability.
A fourth may optimize distribution.
And a marketplace can potentially connect all four.
This is why Token economics is becoming increasingly important.
The question for developers is no longer simply:
“Which model is best?”
It is becoming:
“Which model, at which price, from which supplier, at which time, is best for this workload?”
That is a much more sophisticated procurement problem.
And if platforms such as Unexhub can successfully turn that complexity into automated aggregation, competitive pricing, intelligent routing, and usage-based settlement, the Token market may gradually evolve from vendor pricing into market pricing.
The 75× gap is therefore not just a number.
It is a signal that AI compute is becoming a tradable economic resource.
中文版
0.02 元与 1.5 元之间:75 倍价差背后的两种算力经济学
双语文章 · English Version 已在上方,中文完整版从这里开始 ↓
2026 年 8 月,Token 交易市场正在经历一次真正意义上的定价机制实验。
按照本文引用的价格数据,DeepSeek V4 Flash 在缓存命中的情况下,输入价格最低可以达到 0.02 元/百万 Token;而 Qwen3.8-Max 的隐式缓存命中输入价格约为 1.5 元/百万 Token。
两者之间,形成了惊人的:
75 倍价格差。
但这并不只是“谁更便宜”的问题。
它实际上代表着两种完全不同的AI 算力经济学,正在同一个市场中正面碰撞。
一、DeepSeek 的算术:用极致低价换取最大调用量
如果用一句话概括 DeepSeek 的定价逻辑,那就是:
把边际成本压到足够低,再用规模化调用把商业模型跑起来。
按照本文引用的价格体系,V4 Flash 的价格被拆得非常细:
缓存命中输入:0.02 元/百万 Token
缓存未命中输入:1 元/百万 Token
输出:2 元/百万 Token
更高一级的 V4 Pro,则对应约 0.025 元、3 元和 6 元的价格体系。
但真正值得关注的,不只是绝对价格。
而是 DeepSeek 引入的:
峰谷分时定价。
在工作日 9:00—12:00、14:00—18:00 等高峰时间段,价格可能提高;而夜间、周末和节假日则维持较低价格。
这意味着:
价格本身成为了一种算力调度工具。
它不是简单告诉用户:
“现在 GPU 很贵,所以你要多付钱。”
而是通过价格信号,引导一部分非紧急任务主动迁移到算力压力较低的时间段。
即便 0.02 元的缓存命中价格在高峰期翻倍,也不过达到 0.04 元/百万 Token。
这背后的信号非常清晰:
DeepSeek 正在尝试把“时间”也变成 Token 经济的一部分。
二、低价策略带来的结果:规模
这套策略到底有没有效果?
从本文引用的 OpenRouter 调用数据来看,DeepSeek V4 系列已经拥有非常庞大的调用规模,相关周榜数据显示其调用量达到约 6.6 万亿 Token。
所以“价格屠夫”并不只是一个营销标签。
它背后确实存在真实的市场需求。
但问题也开始出现。
8 月 6 日,DeepSeek 已释放调整 API 产品线价格的信号。
这意味着市场接下来真正要观察的是:
DeepSeek 的巨大调用量,到底有多少来自模型能力,又有多少来自极低价格?
如果价格上涨之后调用量仍然保持强劲,那么它证明的是模型价值。
如果调用量明显下降,那么市场则会看到低价策略对需求的巨大影响。
三、Qwen3.8-Max 的算术:通过区域差异重新锚定价值
阿里的定价逻辑完全不同。
Qwen3.8-Max 采用了国内与海外两套价格体系。
按照本文引用的数据:
国内价格
输入:12 元/百万 Token
输出:36 元/百万 Token
海外价格
输入:2 美元/百万 Token
输出:6 美元/百万 Token
而国内隐式缓存命中的价格约为:
1.5 元/百万 Token。
这里有一个很容易被忽略的细节。
1.5 元恰好是标准输入价格 12 元的:
12.5%。
海外对应的缓存价格约为 0.25 美元,同样是标准输入价格 2 美元的:
12.5%。
这种比例的一致性,很难简单理解为随机定价。
它更像是一套经过设计的:
标准价格 + 缓存折扣体系。
也就是说:
当系统无需完整重复计算相同上下文时,用户可以获得大幅度的 Token 成本优惠。
阿里实际上同时使用了两层定价逻辑:
第一层:区域差异化。
针对国内和海外市场分别建立价格体系。
第二层:缓存折扣。
通过降低重复计算成本,提高大规模调用场景下的价格竞争力。
四、Qwen 的优势:高能力与低成本之间寻找平衡
与此同时,Qwen3.8-Max 并不是单纯依靠低价竞争。
按照本文引用的性能数据,其 Agentic Index 得分为 55.4,略高于 Claude Opus 5 的 55.3。
同时,其国际输出价格也只有 Claude Opus 5 的一小部分。
这里真正重要的并不是某一个 Benchmark 永远保持第一。
而是市场正在出现一种新的现象:
过去只有旗舰模型才能提供的高端能力,现在正在出现越来越多价格完全不同的替代方案。
这意味着模型选择开始从单纯的:
“谁最强?”
转变成:
“谁的性能/价格比最适合我的任务?”
五、75 倍价差的本质:资源调度 vs. 市场细分
把 DeepSeek 和 Qwen 的定价放在一起看,两套逻辑就非常明显了。
DeepSeek:把价格当作算力调度工具
峰谷价格、缓存命中、不同模型档位,本质上都在解决一个问题:
如何最大化昂贵 GPU 资源的利用效率?
当算力紧张时,价格上涨。
当算力充足时,价格下降。
用户则通过价格信号,把一部分任务迁移到低峰时间。
因此,DeepSeek 更像是在:
管理资源。
Qwen:把价格当作市场细分工具
Qwen 则更加注重:
国内与海外市场差异
不同客户群体
缓存折扣
企业级使用场景
模型能力与商业价值
它解决的问题更接近:
如何让同一个模型,在不同市场中实现更高的商业价值?
因此可以简单概括为:
DeepSeek 更强调算力资源效率。
Qwen 更强调市场价值效率。
六、为什么 75 倍价差反而很重要?
当一个模型缓存命中只需要 0.02 元,而另一个模型需要 1.5 元时,第一个问题自然会出现:
为什么有人愿意支付 75 倍的价格?
答案是:
Token 价格并不是 AI 使用成本的全部。
企业真正需要考虑的还有:
推理能力
编程能力
输出稳定性
上下文长度
Agent 能力
延迟
可用性
SLA
数据治理
技术支持
集成成本
如果一个更贵的模型可以显著降低错误率、人工审核成本或者工程维护成本,那么更高的 Token 单价依然可能是合理的。
反过来,如果只是做批量分类、摘要、信息提取等大量简单任务,那么为了用不到的能力支付高额 Token 成本,同样可能是不经济的。
因此,真正的竞争并不是:
便宜 vs. 昂贵。
而是:
哪一条性能—价格曲线,最适合当前任务?
七、Unexhub:把 75 倍价差变成市场空间
这也是 Unexhub 值得关注的地方。
传统模式下,开发者面对的是一个个独立的模型供应商。
你注册一家。
充值一家。
接入一家。
然后再注册另一家。
价格差异确实存在,但:
发现价格差异、比较价格、切换供应商,都需要开发者自己完成。
Unexhub 想做的事情,是把这些分散的 AI 算力供应聚合到一个市场中。
其定位是:
Token 聚合、分发与算力交易中心。
换句话说,它试图把原本固定的“厂商报价”,进一步转化成可以被市场比较、竞争和重新分配的:
市场价格。
八、即时结算:让中小供应商降低资金压力
传统 API 分销最大的痛点之一,是资金占用。
供应商可能需要:
先提供资源 → 再产生调用 → 最后等待结算。
账期越长,资金压力越大。
Unexhub 强调即时结算模式,则希望让供应商根据实际调用获得对应收益。
这意味着:
减少资金占用
降低囤货压力
缩短结算周期
加快算力变现
对于中小供应商来说,这一点尤其重要。
因为他们不一定拥有大型 AI 厂商的资金实力,却可以通过:
价格 + 能力 + 可用性
参与市场竞争。
九、API 竞价:让供需真正决定 Token 价格
更值得关注的是:
API 竞价机制。
传统 API 定价相对固定。
而市场化机制则允许价格根据:
供给 + 需求 + 时间 + 算力资源
发生变化。
当供应充足时,开发者可能获得更低价格。
当需求暴涨时,供应商则可能获得更高收益。
于是市场形成一个新的循环:
供应增加 → 竞争增强 → 价格下降
以及:
需求增加 → 算力紧张 → 价格上升。
这意味着 Token 不再只是一个固定的账单数字。
它开始拥有:
市场价格。
十、Unexhub 为开发者提供的实际价值
Unexhub 的价值也不只是“聚合更多模型”。
平台希望通过供应商竞争、动态路由和市场化价格,让开发者真正获得更灵活的采购方式。
1. 按 Token 实际付费
开发者可以根据实际 Token 消耗进行结算,而不是简单绑定某一家模型厂商的固定采购体系。
对于高频 API 用户而言,这意味着成本可以更加透明,也更容易进行精细化核算。
2. 最低可享受市场价格约 3 折
按照平台当前的产品定位,在符合具体模型、供应商、库存和活动条件的情况下,开发者最低有机会享受约为**市场价格 30%**的调用成本。
这对于普通用户可能只是单次调用的价格差异。
但对于每天消耗数百万、数千万甚至数十亿 Token 的企业而言:
每一个 Token 节省一点点,最终都可能变成非常可观的年度成本优化。
因此,Unexhub 的低价价值对于高强度 API 用户尤其明显。
3. 毫秒级路由
通过统一的路由层,开发者无需频繁修改底层接口,就可以更加灵活地访问不同模型和供应商资源。
真正的价值并不仅仅是“快”。
而是:
让更合适的算力资源更容易被找到。
4. 供应商一键接入
对于 AI 供应商而言,平台提供统一的分发渠道。
供应商无需单独建设一套完整的销售体系,就能够接触开发者和企业客户。
5. 实时结算调用分成
供应商根据实际 API 调用产生收益,并通过平台进行结算。
这使 AI 算力从:
“资源库存”
进一步变成:
“持续产生现金流的数字资源”。
十一、从“厂商定价”走向“市场定价”
所以,75 倍价差真正重要的地方,并不是:
谁太贵,谁太便宜。
而是:
AI 推理已经越来越难拥有一个统一的“标准价格”。
最终成本会受到很多因素影响:
模型
供应商
缓存命中率
使用时间
峰谷价格
算力供给
API 路由
批量采购
市场竞争
当这些变量开始同时发生作用时,Token 就不再只是一个简单的计量单位。
它开始更像:
一种可以被定价、比较、路由和交易的算力资源。
这正是 Unexhub 希望占据的市场位置。
平台希望汇聚不同规模的 AI 供应商,包括 Gemini、ChatGPT、Claude、DeepSeek、GLM 等大型模型资源,同时吸纳个人和中小企业提供的优质 AI 模型及算力资源。
供应商越多,市场竞争越充分。
市场竞争越充分,开发者获得差异化价格和供应选择的机会也就越多。
十二、75 倍价差,只是开始
DeepSeek V4 Flash 的 0.02 元缓存价格与 Qwen3.8-Max 约 1.5 元的隐式缓存价格,表面上看只是一个价格数字。
但它背后真正揭示的是:
AI 算力正在形成不同的经济层次。
有人负责:
GPU 利用率。
有人负责:
区域市场价值。
有人负责:
模型能力。
有人负责:
供应和分发。
而一个真正成熟的算力交易市场,需要把这些环节连接起来。
因此,未来开发者面对的问题可能不再只是:
“哪个模型最好?”
而会变成:
“对于我的任务,现在什么模型、什么供应商、什么时间、什么价格最合适?”
这才是真正意义上的 Token 经济学。
如果 Unexhub 能够把这种复杂性进一步转化为:
供应聚合 → 市场竞价 → 智能路由 → 实时结算 → 按量付费
那么 Token 市场就可能从今天的:
“厂商报价体系”
逐步走向:
“市场价格体系”。
而 0.02 元与 1.5 元之间的 75 倍价差,或许并不是 AI 算力市场混乱的证明。
恰恰相反。
它可能是一个真正的 Token 算力交易市场正在形成的信号。