token 交易市场 8 月价格战全景:GPT-5.6 Luna 降价 80%,DeepSeek 登顶后却宣布整体涨价

On July 30, 2026, OpenAI announced an 80% price reduction for the GPT-5.6 Luna API.


One Price Cut, One Price Increase: Two Forces Sitting at the Same Table

On July 30, 2026, OpenAI announced an 80% price reduction for the GPT-5.6 Luna API.

The new pricing:

  • Input tokens: reduced from $1 to $0.2 per million tokens;

  • Output tokens: reduced from $6 to $1.2 per million tokens.

At the same time:

  • GPT-5.6 Terra received a 20% price reduction;

  • The flagship GPT-5.6 Sol remained unchanged.

Exactly one week later, on August 6, DeepSeek released a developer-side announcement:

“DeepSeek API pricing will undergo an overall adjustment in the near future, with a relatively significant increase expected.”

One company pushed its chips onto the table.

Another began collecting them back.

At first glance, these moves appear contradictory.

But they are actually two sides of the same market mechanism:

The AI Token economy is searching for its equilibrium price.


Why OpenAI Cut Prices: AI Reduced Its Own Costs

OpenAI’s explanation for the price reduction was particularly interesting.

It was not simply:

“Sharing savings with users.”

The logic was:

AI helped reduce the cost of running AI.

GPT-5.6 Sol participated in optimizing OpenAI’s own production infrastructure.

According to OpenAI, Sol analyzed production workloads, identified inefficient resource allocation, rewrote production kernels using Triton and Gluon, and conducted hundreds of architectural experiments for speculative decoding systems.

These improvements reportedly achieved:

  • Around 20% lower end-to-end service costs;

  • More than 15% improvement in token generation efficiency.

In a sense:

AI optimized the infrastructure that runs AI.

And that infrastructure optimization created room for lower prices.


However, the market interpretation goes deeper.

The price reduction also reflects increasing competitive pressure.

Over the past several months:

  • DeepSeek V4;

  • Kimi K3;

  • Other high-performance Chinese AI models

have continued gaining global developer attention through aggressive cost-performance advantages.

Recent market data showed that Chinese models accounted for approximately 63.5% of global AI model usage share in late July 2026.

Meanwhile, OpenRouter’s latest weekly ranking showed:

  1. DeepSeek V4 Flash;

  2. Xiaomi MiMo-V2.5;

  3. Tencent Hy3;

  4. DeepSeek V4 Pro;

  5. Zhipu GLM-5.2.

All top five models came from Chinese AI companies.

OpenAI’s price reduction was not only a technological optimization.

It was also a response to a changing market.


Why DeepSeek Increased Prices: Success Created Its Own Pressure

DeepSeek’s situation was completely different.

On August 4, OpenRouter’s weekly statistics showed:

DeepSeek V4 Flash reached 7.22 trillion tokens in weekly usage, ranking first globally.

Even more surprising:

OpenCode, an overseas developer platform, reported that on August 1 alone, V4 Flash processed approximately:

8 trillion tokens in one day.

For comparison:

OpenRouter aggregates more than 400 AI models and processes approximately 6.6 trillion tokens per day across its entire platform.

One model processed more daily traffic than an entire ecosystem of hundreds of models.

The pricing behind this massive demand was extremely aggressive:

DeepSeek V4 Flash:

  • Cached input: approximately 1 RMB / million tokens;

  • Output: approximately 2 RMB / million tokens.

Compared with premium models from Anthropic, the price difference reached approximately:

85× to 105×.

This level of cost efficiency created enormous demand.

But extraordinary demand creates infrastructure pressure.

On August 4, DeepSeek V4 Flash API experienced capacity issues due to:

“unprecedented traffic volume.”

Even with a highly efficient MoE architecture, when usage reaches trillion-token scale, the bottleneck is no longer algorithmic efficiency.

It becomes physical infrastructure.

More servers are required.

And servers cost money.

DeepSeek did not necessarily want to raise prices.

Its own success forced the adjustment.


The Market Is Searching for Balance — Token Exchanges Are Being Built

Together, these two events reveal a much deeper trend:

Tokens are evolving from a pricing unit controlled by AI companies into a tradable digital commodity.

Traditional AI API pricing is:

  • Static;

  • Linear;

  • Model-centered.

Users simply choose:

OpenAI:
lower prices.

DeepSeek:
lower costs.

Claude:
premium performance.

But the market is moving toward a different structure.

The question is no longer:

“Which company sets the cheapest price?”

The question becomes:

“What is the real market value of one unit of intelligence?”


Industry trends already show this direction.

Token-based computing markets are emerging globally.

In China, Token-based computing platforms have begun exploring:

  • Compute scheduling;

  • Model access;

  • Token settlement;

  • Resource trading.

Government and industry initiatives are also promoting flexible pricing models based on:

  • Card-hours;

  • Core-hours;

  • Token consumption.

Token is gradually becoming a new form of:

AI computing currency.


Unexhub: Building Infrastructure for a Token Economy

Under this trend, platforms like Unexhub are moving beyond traditional API gateways.

The goal is not simply connecting models.

The goal is creating:

a marketplace where AI computing resources can be aggregated, priced, traded, and optimized.

The infrastructure focuses on three major mechanisms.


1. Instant Settlement: Improving Supplier Liquidity

Traditional AI infrastructure providers often face financial pressure.

They provide computing resources first.

Revenue arrives later.

This creates challenges for smaller AI suppliers.

A Token marketplace model introduces usage-based settlement:

  • Suppliers receive revenue according to actual Token consumption;

  • Cash-flow pressure is reduced;

  • More AI providers can participate.

Instant settlement improves liquidity across the AI supply chain.


2. Supplier Incentives: Expanding the AI Supply Network

A healthy AI marketplace requires diverse suppliers.

Early ecosystem participants can receive platform incentives, such as reduced service fees.

This is not simply a discount strategy.

It is a liquidity-building mechanism.

The platform that attracts more high-quality AI resources gains stronger market influence.


3. Dynamic API Bidding: Letting the Market Discover Prices

Traditional AI pricing is controlled by model providers.

Future Token markets may work differently.

Prices can move according to:

  • Supply;

  • Demand;

  • Real-time usage;

  • Available computing capacity.

When supply is abundant:

Developers receive lower prices.

When demand increases:

Suppliers receive better returns.

Token pricing becomes dynamic.

AI computing resources become market-driven assets.


When DeepSeek pricing moves upward from extremely low levels, and OpenAI pricing moves downward from premium levels, both sides are approaching the same destination:

The market equilibrium price of Tokens.

A Token exchange does not replace AI providers.

It enables transparent price discovery.


Engineer Perspective: Price Wars Are Not the Final Goal

After OpenAI’s 80% price reduction:

GPT-5.6 Luna’s input pricing became close to DeepSeek V4 Pro levels.

However, Luna can still provide:

  • Tool usage;

  • Long-context processing;

  • Multi-step workflows.

Low-cost models are no longer limited to simple tasks.

They are entering:

  • Code review;

  • Monitoring;

  • Automation;

  • Agent workflows.


When DeepSeek increased pricing, some users criticized the move.

But another perspective is important:

A model processing more traffic than hundreds of models combined is no longer simply a “cheap model.”

It has become infrastructure.

Infrastructure cannot operate indefinitely at unsustainable prices.


Research estimates show:

DeepSeek V4 Flash remains significantly cheaper than GPT-5.6 Luna, with blended costs roughly 65% lower and certain task costs reduced by approximately 57%.

The gap still exists.

But it is narrowing.

OpenAI is moving downward.

DeepSeek is moving upward.

The two pricing curves are approaching each other.

And Token marketplaces like Unexhub provide the infrastructure needed during this transition.


The Future Is Not a Price War — It Is Price Discovery

The real future question is not:

“Which model is cheaper?”

It is:

“Which model provides the highest intelligence value per Token at this moment?”

When Tokens become dynamically priced like electricity:

AI selection becomes a market decision.

Not a marketing decision.


DeepSeek’s pricing strategy created a temporary benchmark for the industry.

But that benchmark will continue moving with:

  • Supply;

  • Demand;

  • Infrastructure costs;

  • Model efficiency.

The role of platforms like Unexhub is to make this process:

  • Transparent;

  • Real-time;

  • Competitive;

  • Tradable.


Token Is Becoming a Commodity — And Commodities Need Markets

The AI industry is entering a new phase.

From:

Selling models

to:

Trading intelligence resources.

From:

Fixed pricing

to:

Market-based pricing.

From:

Choosing a model

to:

Choosing the most efficient Token performance.

Unexhub is building a global Token aggregation and compute trading center:

  • Connecting AI suppliers worldwide;

  • Supporting major models including Gemini, ChatGPT, Claude, DeepSeek, GLM and other AI systems;

  • Enabling developers and enterprises to access optimized AI resources;

  • Providing intelligent routing and competitive pricing.

Developers can pay according to actual Token usage and access optimized market pricing through competitive supply.

AI suppliers can connect once and receive usage-based settlement through unified distribution channels.

With features such as:

  • Multi-model aggregation;

  • Intelligent routing;

  • Dynamic pricing;

  • Enterprise AI integration services;

Unexhub aims to create a more efficient AI resource marketplace.

The future AI economy will not only ask:

“Who built the smartest model?”

It will ask:

“Who can deliver the best intelligence at the most efficient Token cost?”

And that is where the Token economy begins.

一次降价,一次涨价,同桌而坐:Token 市场正在寻找新的均衡价格

2026 年 7 月 30 日,OpenAI 宣布 GPT-5.6 Luna API 大幅降价 80%。

其中,输入价格从每百万 Token 1 美元降至 0.2 美元,输出价格从每百万 Token 6 美元降至 1.2 美元;中端模型 Terra 同步降价约 20%,而旗舰模型 Sol 暂时维持原有价格。

仅仅一周后,2026 年 8 月 6 日,DeepSeek 在开发者后台发布公告,表示计划近期对 DeepSeek API 服务进行整体调价,预计涨幅较大。

一个主动降低价格,一个开始提高价格。

看似方向完全相反,但实际上,它们正在共同推动同一个结果:

AI Token 市场正在从厂商定价时代,进入市场均衡价格形成时代。


OpenAI 降价的背后:AI 开始帮助自己降低成本

OpenAI 对 GPT-5.6 Luna 降价给出的解释非常特别:

这不是简单的市场让利,而是 AI 系统本身帮助企业降低了运营成本。

根据 OpenAI 公开信息,GPT-5.6 Sol 已经参与自身生产系统优化工作,包括分析真实生产流量中的负载情况,利用 Triton 和 Gluon 重构部分计算内核,并针对推测解码草案模型设计大量架构实验。

这些优化带来的结果是:

  • 端到端服务成本下降约 20%

  • Token 生成效率提升超过 15%

换句话说:

AI 优化了 AI 的运行环境,而优化后的环境又反过来降低了 AI 服务价格。

这是人工智能产业正在出现的新循环:

模型能力提升 → 基础设施优化 → 单 Token 成本下降 → 服务价格下降 → 用户规模扩大。

不过,从市场角度来看,OpenAI 此次降价并不仅仅来自内部效率提升。

更大的压力来自竞争。

过去半年,以 DeepSeek V4、Kimi K3 等为代表的高性价比模型快速进入全球开发者市场,不断挑战传统闭源模型的价格体系。

数据显示,2026 年 7 月下旬,全球 AI 模型调用市场中,国产模型占比已经超过 60%。

在 OpenRouter 最新周度排名中,调用量排名前五的模型全部来自中国厂商:

  • DeepSeek V4 Flash

  • 小米 MiMo-V2.5

  • 腾讯 Hy3

  • DeepSeek V4 Pro

  • 智谱 GLM-5.2

硅谷模型第一次在全球调用榜前五中全部缺席。

OpenAI 的降价,本质上是技术进步与市场竞争共同推动的结果。


DeepSeek 涨价的背后:成功带来的基础设施压力

如果说 OpenAI 降价是为了竞争,那么 DeepSeek 调价则来自另一种压力:

需求增长太快。

2026 年 8 月 4 日,OpenRouter 周度数据显示:

DeepSeek V4 Flash 单周 Token 调用量达到 7.22 万亿,排名全球第一。

更加惊人的数据来自海外开发平台 OpenCode:

8 月 1 日当天,DeepSeek V4 Flash 单日处理 Token 数量达到约 8 万亿。

作为对比:

OpenRouter 聚合超过 400 个模型,全平台每天处理 Token 总量约为 6.6 万亿。

也就是说:

一个模型一天的调用量,超过了一个聚合平台所有模型的总流量。

支撑这一巨大需求的是极具竞争力的价格:

DeepSeek V4 Flash:

  • 输入价格约 1 元 / 百万 Token

  • 输出价格约 2 元 / 百万 Token

相比 Anthropic Claude Opus 4.8 等旗舰模型,价格差距达到几十倍,部分场景甚至超过 80 倍。

这种极低价格带来了巨大的用户增长,但也带来了新的问题:

基础设施压力。

8 月 4 日,DeepSeek V4 Flash API 曾因访问量暴增出现容量不足情况。

原因很简单:

模型可以通过架构优化降低单次计算成本,但无法突破物理限制。

更多请求意味着:

  • 更多 GPU

  • 更多服务器

  • 更多数据中心投入

DeepSeek 并不是想涨价。

而是市场需求推动它重新寻找可持续价格。


从模型价格,到 Token 商品价格

OpenAI 降价,DeepSeek 涨价。

这两件看似相反的事情,其实揭示了一个更深层趋势:

Token 正在从模型厂商的计费单位,变成 AI 时代的新型商品。

过去:

用户购买的是模型。

比如:

  • Claude

  • GPT

  • Gemini

  • DeepSeek

价格由模型厂商制定。

用户只能选择:

贵一点的旗舰模型,或者便宜一点的基础模型。

但随着 AI 应用规模扩大,新的问题出现:

真正决定价值的,不再只是模型名称,而是:

  • 每 Token 成本

  • 实际任务效果

  • 响应速度

  • 稳定性

  • 调用效率

未来竞争的核心,不是:

“哪个模型最强?”

而是:

“哪个 Token 在当前市场环境下性价比最高?”


Token 交易市场正在形成新的基础设施

这也是 Unexhub 试图解决的问题。

未来 AI 服务市场需要的不只是 API 聚合,而是一个真正围绕 Token 流通、交易、结算建立的新型算力市场。

Unexhub 的核心逻辑,是将 Token 从固定价格商品,转变为可交易资源。

主要包括三个方向:

第一,即时结算机制

传统模式下,中小 AI 模型供应商通常面临资金回收周期问题。

提供算力和 Token 后,需要等待较长周期才能获得收益。

Unexhub 通过 Token 调用结算体系,让供应商能够根据实际调用量快速获得收益,提高供给端流动性。


第二,供应侧激励机制

为了吸引更多模型供应商进入生态,Unexhub 为早期合作供应商提供更低的平台服务成本。

通过降低接入门槛,让更多:

  • 大型模型厂商

  • 独立开发者

  • 中小 AI 企业

能够参与 Token 供应市场。

谁能够聚集更多优质 Token 资源,谁就拥有更强的话语权。


第三,API 动态竞价机制

传统 API 定价是固定模式:

模型厂商设定价格,用户被动接受。

而未来 Token 市场可能采用动态价格发现机制:

  • 需求低峰期,用户获得更优惠价格

  • 高需求时期,供应商获得更合理收益

Token 不再只是账单上的数字,而会像能源、电力一样,根据市场供需变化实时调整。


工程师视角:价格战不是终点,价格发现才是核心

OpenAI 降价之后,GPT-5.6 Luna 的价格已经接近部分国产模型。

但 Luna 依然具备:

  • 工具调用能力

  • 长上下文处理能力

  • 多步骤任务执行能力

这意味着:

低价模型正在进入过去只有旗舰模型才能覆盖的应用领域。

与此同时,DeepSeek 涨价也说明:

一个模型如果承担了全球级别的调用量,就不再只是“低价模型”。

它正在成为基础设施。

基础设施必须满足:

  • 稳定运行

  • 持续扩容

  • 商业可持续

因此,未来 AI 市场不会简单陷入价格战。

真正重要的是:

价格发现。

当 Token 可以像商品一样自由流通,市场会自动寻找最合理价格。

用户关注的不再是:

“哪个模型最便宜?”

而是:

“当前任务下,哪个 Token 最值得购买?”


Token 正在成为 AI 时代的新商品

GPT-5.6 Luna 降价 80%。

DeepSeek V4 Flash 调整价格。

一个向下,一个向上。

两条价格曲线正在靠近同一个目标:

Token 的市场均衡价格。

未来 AI 竞争的核心,将从:

“谁拥有最大的模型”

转向:

“谁能够提供最高性价比的智能资源”。

Unexhub 正是在这一趋势下构建 Token 聚合与分发的算力交易中心。

平台已连接不同规模的 AI 服务供应商,包括 Gemini、ChatGPT、Claude、DeepSeek、GLM 等主流模型生态,同时也支持个人开发者和中小企业提供优质 AI 能力。

通过供应商竞价机制、智能路由和 Token 计费模式,企业用户和开发者可以根据实际需求选择最优模型组合。

在 Unexhub:

  • 开发者按照实际 Token 使用量付费

  • 部分场景最低可享市场价格约 3 折

  • 供应商可一键接入渠道

  • 根据调用量实时获得收益分成

未来,AI 不只是模型之间的竞争。

它将成为一个由 Token 驱动的全球智能资源市场。

模型决定能力,Token 决定流通,而交易市场决定价格。

这场 AI 基础设施革命,才刚刚开始。