$1.20 per Million Tokens: When AI Inference Costs Break Through the Floor
Bilingual Article · English First / 中文版见下文
On July 30, 2026, OpenAI did something that forced the entire AI industry to recalculate its numbers.
Just three weeks before the GPT-5.6 series officially launched, OpenAI announced an 80% price cut for its entry-level Luna API. Input pricing fell from $1 to $0.20 per million tokens, while output pricing dropped from $6 to just $1.20 per million tokens.
For developers, that means the output cost fell from roughly RMB 40.6 to RMB 8.1 per million tokens.
Terra, positioned as the "daily workhorse," also received a 20% price reduction, bringing input and output prices down to $2 and $12 per million tokens respectively. The flagship Sol retained its existing pricing of $5 input and $30 output per million tokens, but gained a new Fast mode capable of delivering up to 2.5× higher speed.
This is more than an ordinary price adjustment.
It represents a fundamental restructuring of AI inference economics.
The Model Is Helping Cut Its Own Costs
Why can OpenAI afford an 80% price reduction?
OpenAI's explanation points to an increasingly important trend: AI models are beginning to participate in optimizing the infrastructure on which they run.
According to the reported details, GPT-5.6 Sol helped optimize parts of OpenAI's production systems. It rewrote portions of GPU kernels, designed and executed hundreds of token-generation experiments, and participated in training-monitoring work.
The resulting improvements reportedly reduced end-to-end serving costs by around 20%, while inference-decoding optimization increased token-generation efficiency by more than 15%.
In other words, the model helped optimize the software running on the hardware that runs the model.
It sounds almost like AI is "lifting itself by its own bootstraps."
But the underlying trend is much more important.
AI cost reduction is gradually moving beyond "engineers optimizing AI systems" toward "AI helping engineers optimize the systems themselves."
The engineer's role is beginning to shift from simply writing low-level kernels to designing environments in which AI can help write, test, benchmark, and optimize those kernels.
Once that efficiency flywheel starts turning, lower prices no longer need to be a one-time promotional tactic.
They can become a structural trend.
The Token Price Index: From $2.20 to $1.16
OpenAI's price reduction is not happening in isolation.
The broader token market is experiencing a significant downward repricing.
According to the Token Spending Index published by Silicon Data, after rising sharply during the first half of 2026, token prices began declining in June. From a peak of approximately $2.20 per million tokens, the market had fallen by roughly 40%.
Jefferies data reportedly showed inference prices falling to around $1.16–$1.18 per million tokens between August 6 and 8, marking a new low for the year.
The Token Spending Price Index itself reportedly fell to 1.1635, down 43.7% from its May 28 peak of 2.0651.
The signal is clear:
AI inference is becoming cheaper across the industry—not just at a single vendor.
Several forces are driving this decline.
DeepSeek V4 Flash has pushed certain task-level costs toward extremely low levels. Xiaomi's MiMo-V2.5 has reportedly cut prices dramatically in some scenarios, while Tencent Cloud has also introduced substantial reductions for its DeepSeek V4 series.
The growing cost-performance advantage of lower-priced models is forcing even major international AI providers to reconsider how their own pricing structures are built.
The Developer Dividend of the "Cost-Sensitive" AI Era
The most important consequence of falling token prices may not be cheaper APIs.
It may be what developers can now afford to do with them.
After Luna's price reduction, its input price became comparable to the previous-generation GPT-5.4 nano, while its output price became even cheaper by the figures cited above.
Yet Luna is positioned for substantially broader workloads, including tool use, longer-context processing, and multi-step workflows.
That changes the economics of AI development.
A model that once looked too expensive for continuous background work can suddenly become affordable for:
Code review
Backend monitoring
Data processing
Automated testing
Tool calling
Workflow orchestration
Agent-based applications
The same subscription budget can potentially support substantially more model usage.
OpenAI has also indicated that some internal workloads could become dramatically cheaper after moving certain automated review tasks to GPT-5.6 Luna.
The larger trend is unmistakable:
Lower token prices are turning intelligence from a scarce resource into deployable infrastructure.
But lower prices do not automatically guarantee adoption.
Reported OpenRouter data showed that, during the first week after the price changes, combined usage of the new and old DeepSeek V4 Flash versions reached around 11.31 trillion tokens, while GPT-5.6 Luna recorded approximately 3.54 trillion tokens.
That gap illustrates an important point.
In an AI market where models are increasingly similar in capability, every fraction of a cent per million tokens can influence developer behavior.
Price still matters.
And that is precisely why the next stage of AI competition may no longer be fought purely at the model level.
It may be fought at the token distribution layer.
Unexhub's Bet: Making Price Signals Flow
This is where the strategy behind Unexhub becomes particularly interesting.
Rather than treating tokens merely as an internal billing unit, Unexhub is positioning them as a tradable computing resource within a token aggregation and distribution marketplace.
The basic idea is straightforward:
aggregate supply, connect demand, dynamically route requests, and let market competition influence price.
For smaller AI suppliers, real-time settlement can reduce the pressure created by long capital-recovery cycles. Suppliers can receive settlement based on token consumption rather than waiting through lengthy billing periods.
For early suppliers, zero commission can lower the barrier to entering the marketplace.
And the API bidding mechanism introduces another important element:
Dynamic pricing based on supply and demand.
Under a traditional fixed-price API model, peak congestion and off-peak idle capacity may effectively carry the same price.
A market-based model can work differently.
When demand rises, prices can increase, giving suppliers an incentive to release more capacity.
When demand falls, prices can decline, giving developers and smaller businesses an opportunity to use additional capacity at a lower cost.
That creates something more than a discount.
It creates a price signal.
From "Cheap APIs" to a Real Token Market
OpenAI cutting Luna's output price to $1.20 is a vendor-level pricing decision.
Unexhub is attempting something different:
to make pricing increasingly responsive to the broader market.
That distinction matters.
A vendor determines the price of its own model.
A marketplace can potentially aggregate multiple suppliers and allow users to compare and route demand across them.
Unexhub's model is built around aggregating AI suppliers across different scales, including major model providers such as Gemini, ChatGPT, Claude, DeepSeek and GLM, while also bringing together AI services supplied by smaller teams and independent providers.
For developers and enterprise users, the practical value is not simply having more models to choose from.
It is having more supply competing for the same demand.
According to Unexhub's current positioning, developers can pay based on actual token usage, with the platform offering access to competitive market pricing that can reach as low as 30% of prevailing market prices in eligible scenarios.
For suppliers, the proposition is equally straightforward: one integration, access to additional distribution channels, and real-time settlement of usage-based revenue.
The combination of millisecond-level routing, aggregated model supply, dynamic pricing, and real-time settlement is designed to make token consumption behave more like a market than a traditional software subscription.
$1.20 Is Not the Finish Line
The significance of Luna's $1.20 output price is not simply that developers are paying less.
It is that the economic foundation of AI inference is changing.
Models are becoming more capable.
Infrastructure is becoming more efficient.
Token prices are falling.
Competition between providers is intensifying.
And developers are gaining more leverage over where and how they purchase inference capacity.
The next question is therefore no longer simply:
"Which model is the smartest?"
It is increasingly:
"Which model delivers the best intelligence per dollar—and where can that capacity be purchased most efficiently?"
The AI industry is moving toward a world where intelligence is increasingly abundant, while efficient access to intelligence becomes the real competitive advantage.
At that point, token aggregation and distribution platforms may become an important piece of the infrastructure.
Unexhub is betting on exactly that direction: reshaping the global token economy by building a computing-power trading center for token aggregation and distribution.
For developers, the potential benefit is straightforward: more model choices, competitive routing, usage-based payment, and lower inference costs.
For AI suppliers, the value is equally practical: broader distribution, simpler integration, and real-time settlement.
The cost of intelligence is falling. The next battle is about how efficiently that intelligence can move.
And at $1.20 per million output tokens, Luna may not be the bottom. It may simply be another marker on the way down.
中文版|Chinese Version
每百万 Token 仅 1.2 美元:当 AI 推理成本跌穿地板
双语文章|English Version Above · 中文版从这里开始
2026 年 7 月 30 日,OpenAI 做了一件让整个 AI 行业重新算账的事。
就在 GPT-5.6 系列正式发布前三周,OpenAI 宣布将入门级模型 Luna 的 API 价格大幅下调 80%。
输入价格从每百万 Token 1 美元降至 0.2 美元,输出价格则从 6 美元直接降至 1.2 美元/百万 Token。
换算成人民币,输出价格从约 40.6 元降至约 8.1 元。
定位为"日常工作主力"的 Terra 也同步降价 20%,输入和输出价格分别降至每百万 Token 2 美元和 12 美元。
旗舰级 Sol 则维持原有价格:
输入:5 美元/百万 Token
输出:30 美元/百万 Token
新增 Fast 模式,速度最高提升约 2.5 倍
这已经不是一次普通的价格调整。
它更像是一次 AI 推理经济学的底层重构。
模型开始帮助自己降本
为什么 OpenAI 能够把 Luna 的价格一次性降低 80%?
一个值得关注的答案是:AI 模型本身正在参与优化运行它的基础设施。
按照相关披露,GPT-5.6 Sol 曾参与 OpenAI 生产系统的优化,包括重写部分 GPU 内核、设计并运行数百项 Token 生成实验,以及参与训练监控等工作。
相关优化带来的结果是,GPU 内核优化使端到端服务成本下降约 20%,推理解码优化则让 Token 生成效率提升超过 15%。
换句话说:
模型开始帮助优化运行模型的底层程序。
听起来有点像"左脚踩右脚,原地起飞"。
但真正值得关注的,是背后的产业变化。
过去是:
工程师优化 AI。
现在正在逐渐变成:
AI 协助工程师优化 AI 所运行的系统。
工程师的角色,也正在从单纯编写底层代码,转向设计让 AI 能够参与编写、测试、Benchmark 和优化底层代码的工程环境。
一旦这个效率飞轮真正转起来,降价就不再只是一次营销活动。
它可能会变成一种结构性的长期趋势。
Token 价格指数:从 2.2 美元跌到 1.16 美元
OpenAI 的降价并不是孤立事件。
整个 Token 市场正在经历一次明显的价格重估。
根据 Silicon Data 发布的 Token 支出指数,今年上半年 Token 价格经历快速上涨后,从 6 月开始掉头向下。
此前每百万 Token 约 2.2 美元的价格峰值,随后下降约 40%。
杰富瑞的数据则显示,8 月 6 日至 8 日期间,AI 推理价格已经下降至约 1.16~1.18 美元/百万 Token,创下年内新低。
Token 支出价格指数也下降至 1.1635,相比 5 月 28 日 2.0651 的高点,累计下降约 43.7%。
这意味着:
AI 推理正在变得越来越便宜,而且不是某一家厂商单独降价。
背后是多重力量共同作用。
DeepSeek V4 Flash 将部分任务成本压到了极低水平;小米 MiMo-V2.5 在部分场景大幅降价;腾讯云 DeepSeek V4 系列同样进行了大力度价格调整。
国产模型不断提高的性价比,正在迫使海外 AI 巨头重新计算自己的成本和价格体系。
"成本敏感"时代,开发者开始真正受益
Luna 降价后,输入价格已经接近上一代 GPT-5.4 nano 的水平,而输出价格甚至更低。
但两者承担的工作范围并不完全相同。
Luna 可以进一步承担工具调用、长上下文处理以及多步骤工作流等任务。
这意味着:
过去看起来成本过高、不适合持续运行的 AI 能力,现在开始可以进入更多实际生产场景。
例如:
代码审查
后端监控
数据处理
自动化测试
工具调用
工作流编排
AI Agent 应用
当 Token 单价不断下降,同样的预算可以支持更多模型调用。
OpenAI 也已经将部分自动化审查任务迁移到 GPT-5.6 Luna,并预计能够大幅降低相关成本。
因此,真正发生变化的可能并不只是"API 更便宜了"。
而是:
智能正在从一种稀缺资源,逐渐变成可以大规模部署的基础设施。
当然,价格下降也并不意味着所有模型都会获得同样的市场份额。
据相关 OpenRouter 数据,在价格调整后的首周,DeepSeek V4 Flash 新旧版本合计调用量达到约 11.31 万亿 Token,而 GPT-5.6 Luna 约为 3.54 万亿 Token。
差距接近三倍。
这说明一个非常现实的问题:
当模型能力逐渐接近时,每百万 Token 的成本差异,会直接影响开发者的选择。
价格仍然重要。
而下一阶段的竞争,也可能因此从单纯的"模型竞争",逐渐进入Token 分发与交易层面的竞争。
Unexhub 的算盘:让价格信号真正流动起来
在 Token 价格不断下跌的时候,有人看到的是利润空间被压缩,有人看到的却是一个新市场正在形成。
Unexhub 正在搭建一个以 Token 聚合与分发为核心的算力交易中心。
它的核心思路并不是简单地"卖 API",而是尝试把 Token 从传统的内部计量单位,进一步变成一种可以流通、比较和交易的算力资源。
简单来说,就是:
聚合供给 → 连接需求 → 动态路由 → 市场竞争 → 价格发现。
对于中小型 AI 供应商而言,实时结算能够缓解传统 API 分销模式中资金回笼周期较长的问题。
供应商提供多少 Token,就可以按照实际调用情况同步进行收益结算。
而早期入驻零佣金的策略,本质上是在降低供应商进入市场的门槛,同时帮助平台建立初始流动性。
真正值得关注的,则是 API 竞价模式。
从固定价格,走向动态价格
传统 API 模式往往是固定价格:
无论高峰期拥堵,还是凌晨大量算力闲置,价格可能都不会发生变化。
但市场化机制可以让价格跟随供需变化。
当高峰期需求上涨:
价格上涨 → 供应商获得更高收益 → 更多算力可能进入市场。
当低峰期需求下降:
价格下降 → 开发者获得更低成本 → 闲置算力获得新的需求。
这就不只是"优惠"。
它实际上是在建立一种价格信号机制。
DeepSeek 的峰谷定价,是模型厂商主动调整自己的价格。
而 Unexhub 想进一步做的,是让多个供应商进入同一个市场,让不同供应商之间的供需关系共同影响价格。
前者是:
厂商定价。
后者则更接近:
市场定价。
从"便宜 API"走向真正的 Token 市场
OpenAI 将 Luna 的输出价格降至 1.2 美元,是一次厂商层面的价格决策。
而 Unexhub 所尝试的方向不同。
它希望让价格越来越多地由更广泛的市场供需关系决定。
这一区别非常关键。
一家模型厂商只能决定自己模型的价格。
而一个聚合市场理论上可以同时连接多个供应商,让用户根据价格、性能、可用性等因素进行选择和路由。
目前,Unexhub 已经围绕多个层级的 AI 供应商进行聚合,包括 Gemini、ChatGPT、Claude、DeepSeek、GLM 等大型模型,同时也连接个人开发者、中小企业等提供的 AI 模型与服务。
对于开发者和企业用户而言,这种模式的价值不只是"多了几个模型可以选择"。
真正重要的是:
让更多供应商围绕同一批需求展开竞争。
按照 Unexhub 当前的市场定位,开发者可以按照实际 Token 使用量进行支付,并在符合条件的场景下获得低至市场价格约 3 折的使用成本。
对于供应商来说,价值也更加直接:
一次接入更多分发渠道
获得更多潜在调用需求
实时结算调用收益
降低传统渠道合作的资金周转压力
再结合毫秒级路由、模型聚合、动态价格以及实时结算,Token 消费就不再只是传统意义上的软件 API 购买,而开始具备一定的市场交易属性。
1.2 美元不是终点,而可能只是一个新的起点
Luna 每百万 Token 1.2 美元真正值得关注的,并不只是"开发者又省钱了"。
更深层的变化是:
AI 推理的经济基础正在改变。
模型越来越强。
基础设施越来越高效。
Token 价格越来越低。
供应商之间的竞争越来越激烈。
开发者也开始拥有越来越大的选择权。
因此,未来真正的问题可能不再只是:
"哪个模型最聪明?"
而会变成:
"哪个模型能够提供最高的 Intelligence per Dollar?在哪里可以最低成本地获得这些智能?"
AI 正在进入一个新的阶段:
智能本身越来越丰富,而高效获取智能的能力,正在成为新的竞争优势。
这也正是 Token 聚合与分发平台可能发挥价值的地方。
Unexhub 正在尝试重塑全球 Token 逻辑,打造以 Token 聚合、分发和交易为核心的算力交易中心。
对于开发者而言,其核心价值在于:
更多模型选择 + 动态路由 + 按 Token 实付 + 更具竞争力的价格。
对于 AI 供应商而言,则是:
更广泛的分发渠道 + 更低的接入门槛 + 实时结算调用收益。
当 AI 推理成本持续下降,真正值得竞争的就不再只是"谁能把价格降得更低"。
而是:
谁能让这些越来越便宜的 Token,更高效地流向真正需要它们的人。
每百万 Token 1.2 美元,或许不是 AI 推理成本的终点。
它可能只是价格持续下探过程中,又一个新的坐标点。