token 交易市场从"卖水"转向"卖任务":OpenAI 企业收入跑至 $40B,算力成本衡量标尺生变

从“每百万 Token 多少钱”,转向“每成功完成一个任务多少钱”。


From Selling Tokens to Selling Tasks: How OpenAI’s $40 Billion Run Rate Is Rewriting the Economics of Compute

In August 2026, OpenAI’s annualized revenue run rate officially surpassed $40 billion, doubling in just eight months. More revealing than the number itself is where that growth is coming from: enterprise business continues to expand rapidly, with monthly revenue growth reaching 32%, outpacing the consumer segment.

OpenAI CFO Sarah Friar also confirmed in mid-August that enterprise revenue driven by ChatGPT had surpassed consumer revenue.

The $40 billion figure is impressive.

But the more important question is: how is that money actually being generated?


Who Is Paying for the $40 Billion?

The answer can be found across OpenAI’s product ecosystem.

Three forces are driving much of the growth:

  • The rapid expansion of Codex, OpenAI’s AI coding product;

  • The surge in enterprise API usage;

  • Large-scale deployment of AI across enterprise workflows.

Codex has reportedly reached 3 million weekly active users, roughly five times its level in January. At the same time, API throughput has reached approximately 15 billion tokens per minute.

These numbers point to a fundamental change in the AI business model:

Enterprise customers are increasingly paying not for the model itself, but for the work the model can complete.

That distinction matters.

A company does not ultimately care whether an AI system consumed 100,000 or 1 million tokens. What matters is whether it successfully completed the software migration, analyzed the financial documents, resolved the support tickets, generated the report, or deployed the feature.

In other words, the unit of economic value is gradually moving from tokens to tasks.


Token Pricing Is Losing Its Ability to Tell the Whole Story

Token remains the basic billing unit for most AI APIs. But it is becoming a less useful measure of actual economic value.

Artificial Analysis provided a revealing comparison in August: the estimated cost of completing one Intelligence Index task was approximately:

  • GPT-5.6 Sol: $1.04

  • Claude Opus 5: $2.03

  • Claude Fable 5: $2.75

The interesting part is that Sol does not necessarily have the lowest headline price for every billing category. For example, its output price of $30 per million tokens is actually higher than Opus 5’s $25.

Yet Sol can still achieve a lower cost per completed task.

Why?

Because the real bill depends on much more than the published token price.

Three hidden variables are reshaping the bill

1. Tokenizer efficiency

Different tokenizers can represent the same piece of text using different numbers of tokens. Fewer tokens for the same task can translate directly into lower consumption.

2. Reasoning-token economics

Different providers handle reasoning tokens differently. Some reasoning processes may be hidden from users, while others are reflected more explicitly in usage and billing.

3. Retry and failure rates

A task that succeeds on the first attempt is economically very different from one that requires multiple retries, tool calls, corrections, and human review.

That is why two models with seemingly similar API prices can produce dramatically different costs at the task level.

The more meaningful question is no longer:

How much does one million tokens cost?

It is becoming:

How much does it cost to successfully complete one task?

This is the logic behind the emerging Cost per Completed Task (CPCT) framework, which attempts to combine token costs, caching, tool usage, sandbox resources, failed attempts, and human review into a broader measure of the cost of accepted work.

The measurement standard for AI compute is therefore shifting—from cost per million tokens toward cost per successfully completed task.


Where Does Unexhub Fit In?

As OpenAI pushes enterprise AI revenue toward the $40 billion level, another structural change is taking place underneath the model market.

That change is the transition from vendor-controlled token pricing to market-driven pricing.

Under the traditional model, the AI provider determines the price, and customers simply pay the listed rate.

But the market is already becoming more fragmented.

The same model can be offered by different providers at significantly different prices. This has created opportunities for API resellers and intermediary services, but the rapidly growing gray market has also exposed serious problems, including dependence on promotional quotas, refund arbitrage, inaccurate token reporting, model routing failures, and questionable data practices.

The result is an ecosystem where pricing can become opaque and the interests of model suppliers, intermediaries, and developers are not always aligned.

Unexhub is designed around a different idea: building a more transparent token aggregation and distribution marketplace.

Instead of relying solely on fixed price lists, Unexhub aims to connect AI suppliers and demand-side users through a market-oriented infrastructure.

Three mechanisms are particularly important

Instant settlement

For smaller and independent AI suppliers, selling compute is not only about finding customers. Cash flow can be equally important.

With real-time settlement, suppliers can receive their share as usage occurs instead of waiting through lengthy settlement cycles. That can reduce working-capital pressure and make it easier for smaller suppliers to participate.

Early-stage zero commission

For suppliers entering the platform during the early stage, zero-commission incentives can lower the cost of acquiring demand and encourage more compute providers to join the marketplace.

This is more than a promotional discount.

The broader objective is to attract enough supply to make price discovery meaningful.

API bidding

API bidding introduces a more dynamic pricing mechanism.

Instead of treating the price of an AI API as permanently fixed, supply and demand can influence the effective price. During periods of excess capacity, developers may have access to more competitive pricing. When demand rises, suppliers can capture additional value from scarce compute capacity.

The goal is straightforward:

Turn token pricing from a static price list into a market-discovered price.


From Tokens to Electricity

There is another way to understand this transformation.

Token consumption is increasingly beginning to resemble electricity consumption.

Different applications need different amounts of computing power. A simple classification task does not require the same model or inference budget as a complex coding agent. A high-volume summarization workload should not necessarily consume the same resources as a mission-critical autonomous workflow.

In this environment, the market naturally becomes segmented:

  • High-volume, low-complexity workloads → lower-cost models

  • Coding agents → stronger reasoning models

  • Mission-critical workflows → premium models

  • Latency-sensitive applications → faster inference options

The point is not to make every model equally cheap.

The point is to make the right compute available at the right price for the right task.

This is where token aggregation becomes increasingly important.

If developers have to negotiate separately with dozens of model providers, compare pricing structures, manage different APIs, and monitor usage independently, the complexity itself becomes a cost.

An aggregation marketplace can potentially turn that complexity into a unified access layer.


The $40 Billion Number Is Only the Beginning

OpenAI’s $40 billion revenue run rate is not simply a milestone for one AI company.

It is a signal that the AI economy is moving into a different phase.

The first phase was about models:

Who has the smartest model?

The second phase became about tokens:

Who can provide inference at the lowest cost?

The next phase is increasingly about tasks and outcomes:

Who can complete the required work at the lowest reliable cost?

That shift changes how companies should evaluate AI infrastructure.

A model with a lower token price is not necessarily cheaper.

A model with a higher benchmark score is not necessarily more economical.

And a model with the lowest API price may still become expensive if it requires more retries, more tokens, more tool calls, or more human intervention.

The real metric is increasingly:

Cost × Reliability × Latency ÷ Completed Work

That is the economic equation AI infrastructure providers will eventually have to optimize.


Unexhub’s Role in the New Compute Market

Unexhub is positioning itself around this transition by building a global token aggregation, distribution, and compute trading center.

The platform is designed to bring together AI suppliers across different scales, including major model ecosystems such as Gemini, ChatGPT, Claude, DeepSeek, and GLM, as well as high-quality models supplied by independent developers and smaller organizations.

Instead of requiring every developer or enterprise to negotiate separately with multiple providers, Unexhub aims to aggregate supply and use routing and market mechanisms to connect available compute with actual demand.

The value proposition is particularly relevant for high-frequency AI workloads.

For developers, the platform emphasizes:

  • Token-based pay-as-you-use billing

  • Competitive market pricing

  • Millisecond-level routing

  • Access to multiple AI model providers through one platform

  • Potential pricing as low as 30% of prevailing market prices, depending on the applicable supplier, model, traffic conditions, and promotional terms

For suppliers, the platform focuses on:

  • One-step channel integration

  • Access to aggregated enterprise and developer demand

  • Real-time settlement of usage-based revenue

  • Greater flexibility to participate in market-based API pricing

The early-stage incentives are therefore better understood as part of the marketplace-building strategy rather than simply a temporary discount.

For developers, the potential benefit is straightforward: lower effective inference costs and easier access to multiple models.

For suppliers, the value lies in additional demand channels, faster settlement, and more flexible pricing opportunities.


From Selling Tokens to Selling Results

The AI industry is gradually moving beyond the question of:

“How many tokens did the model consume?”

The more important question is becoming:

“What did those tokens accomplish?”

OpenAI’s $40 billion enterprise-driven growth is one expression of that shift.

Task-level cost measurement is another.

Dynamic token pricing is another.

And platforms such as Unexhub are attempting to build the market infrastructure required to connect all three.

The future AI compute market may therefore look less like a traditional software pricing table and more like an energy market:

Supply → Routing → Price Discovery → Compute → Completed Task → Settlement

The ultimate commodity is not the token itself.

It is usable intelligence at a predictable cost.

And when the industry moves from selling tokens to selling completed work, the winners will not necessarily be the companies with the cheapest tokens.

They will be the ones capable of delivering the best combination of intelligence, reliability, speed, and cost per completed task.

The $40 billion OpenAI milestone may mark the beginning of that transition—not the end of it.


中文版|双语文章中文部分

从卖 Token 到卖任务:OpenAI 400 亿美元账单与算力规模的重写

2026 年 8 月,OpenAI 的年化营收运转率正式突破 400 亿美元,短短 8 个月实现翻倍。

更值得关注的其实不是 400 亿这个数字,而是这笔收入正在从哪里产生。

企业业务持续高速增长,月度环比增速达到 32%,明显跑赢消费者业务。OpenAI CFO Sarah Friar 也在 8 月中旬确认,由 ChatGPT 驱动的企业营收已经超过消费者业务。

400 亿美元当然是一个惊人的数字。

但真正值得研究的问题是:

这 400 亿美元究竟是怎么赚出来的?


400 亿美元是谁在付?

答案藏在 OpenAI 不断扩张的产品体系里。

目前推动增长的主要力量,大致来自三个方向:

  • Codex 等 AI 编程产品快速扩张;

  • 企业级 API 调用量持续增长;

  • 企业客户开始大规模部署 AI 工作流。

Codex 周活跃用户已经达到 300 万,相比 1 月增长约 5 倍;与此同时,API 吞吐量达到约 每分钟 150 亿 Token。

这些数字实际上揭示了一个更加重要的变化:

企业正在逐渐从“购买模型”,转向“购买模型完成的工作”。

这一区别非常关键。

一家企业最终并不关心一个任务究竟消耗了 10 万 Token 还是 100 万 Token。

它真正关心的是:

代码迁移完成了吗?

财务数据分析完成了吗?

客服工单处理了吗?

报告生成了吗?

软件功能上线了吗?

换句话说,AI 的经济价值正在从 Token 向 任务 转移。


Token 计价正在失去完整的价值解释能力

目前,Token 仍然是绝大多数 AI API 的基础计费单位。

但随着模型能力、推理链、工具调用和 Agent 工作流越来越复杂,仅仅观察“每百万 Token 多少钱”,已经很难准确判断一次 AI 任务究竟花了多少钱。

Artificial Analysis 在 8 月给出的一组测算非常有代表性:

模型完成一次 Intelligence Index 任务的估算成本
GPT-5.6 Sol约 $1.04
Claude Opus 5约 $2.03
Claude Fable 5约 $2.75

有意思的是,Sol 并不是所有计费项目中的标价都更低。

例如,Sol 的输出价格为 30 美元/百万 Token,实际上高于 Opus 5 的 25 美元/百万 Token。

但最终完成同一个任务,Sol 的实际成本却可能更低。

为什么?

因为真正决定账单的,从来不只是价目表。


三个隐藏变量正在重新定义 AI 成本

01|Tokenizer 效率

不同模型使用不同的分词器。

同样一段文本,不同 Tokenizer 可能产生完全不同数量的 Token。

如果模型能够用更少 Token 表达相同信息,那么即使单 Token 价格不占优势,最终任务成本仍然可能更低。


02|推理 Token 的成本

不同模型对于 reasoning token 的处理方式不同。

有些模型隐藏推理过程,有些模型则会以更加明确的方式体现推理消耗。

因此,表面上看起来相近的 API 单价,在实际任务中可能产生完全不同的账单。


03|失败与重试成本

一个一次成功的 Agent 任务,与一个需要不断重试、调用工具、修正代码、重新执行的任务,经济成本完全不同。

尤其对于企业级 Coding Agent 来说:

一次失败,可能意味着更多 Token、更多 GPU 时间、更多工具调用,以及更多人工审核成本。

所以真正应该问的问题已经不再只是:

100 万 Token 多少钱?

而应该逐渐变成:

成功完成一个任务,需要多少钱?


CPCT:AI 算力正在从“Token 成本”转向“任务成本”

这也是 Cost per Completed Task(CPCT,每完成任务成本) 思路越来越受到关注的原因。

它试图把影响一次 AI 任务真实成本的因素放到同一个框架里:

  • 输入 Token;

  • 输出 Token;

  • 缓存成本;

  • 工具调用;

  • Sandbox / 执行环境;

  • 失败尝试;

  • 重试;

  • 人工审核。

最后真正计算的,不再是“消耗了多少 Token”,而是:

为了获得一份最终被接受的有效工作成果,企业究竟花了多少钱?

这意味着 AI 算力成本的衡量标准正在发生变化:

从“每百万 Token 多少钱”,转向“每成功完成一个任务多少钱”。


Unexhub 正在解决什么问题?

当 OpenAI 的企业业务不断扩大,AI 市场的另一个结构性问题也越来越明显:

Token 的价格究竟应该由谁决定?

传统模式非常简单:

模型厂商定价。

开发者购买。

企业付费。

价格表决定一切。

但随着模型供应商、API 服务商和中间平台不断增加,同一个模型在不同渠道的价格已经可能出现明显差异。

例如,同一个 DeepSeek 模型,不同服务商之间可能出现数倍价差。

这也催生了大量 Token 代理、中转和 API 聚合业务。

但传统中转模式同时存在不少问题,包括:

  • 价格透明度不足;

  • 供应来源不稳定;

  • Token 统计口径不一致;

  • 模型路由质量差异;

  • 退款套利;

  • 数据安全风险;

  • 中间环节抽佣过高。

当中间层越来越复杂,真正承担成本的开发者和企业,反而很难知道:

自己支付的钱究竟有多少真正用于算力。


从固定价目表走向 Token 市场

Unexhub 的思路,是尝试建立一个更加透明的 Token 聚合、分发与算力交易基础设施。

核心并不是简单地“再做一个 API 中转站”。

而是希望把供给侧和需求侧放到同一个市场体系中,通过聚合、路由和动态竞价,让 Token 从固定价格逐渐走向市场化价格发现。

其中三个机制尤其值得关注。


01|即时结算:让供应商更快回笼资金

对于中小型 AI 模型供应商而言,真正的问题并不只是“有没有客户”。

现金流同样重要。

传统合作模式下,供应商提供算力之后,往往需要等待较长结算周期。

而即时结算的核心价值,就是让供应商能够根据实际调用情况更快获得对应收益。

对于独立开发者、中小模型团队和算力供应商来说,这意味着:

卖出多少算力,就能够更快回收多少资金。

这能够降低资金占用,也能进一步降低小型供应商进入市场的门槛。


02|早鸟零佣金:降低供给侧进入成本

对于平台早期阶段,零佣金或阶段性佣金优惠的意义,不只是“少收一点手续费”。

更重要的是:

降低 AI 供应商进入新市场的试错成本。

供应商可以在更低渠道成本的情况下测试市场需求、获取开发者客户,并验证自身模型或算力资源的商业价值。

而平台则能够通过早期激励吸引更多供应商加入。

供应越丰富,价格发现才越有可能真正成立。


03|API 竞价:让价格由市场决定

传统 API 最大的问题之一,是价格通常是固定的。

但算力本身并不是固定资源。

不同时间段、不同模型、不同地区、不同负载条件下,实际算力供需都可能发生变化。

因此,Unexhub 希望通过 API 竞价模式,让 Token 价格能够更加灵活地反映供需关系。

当供应充足时,开发者有机会获得更具竞争力的价格。

当需求快速上涨时,供应商也能够从稀缺算力中获得更高收益。

这意味着:

Token 不再只是一个写死在网页上的价格,而可能成为一个动态交易价格。


从 Token 到电力:AI 算力正在进入分层市场

Token 正越来越像电力。

不同设备需要不同电压。

不同任务,也需要不同等级的模型与算力。

一个简单的分类任务,没有必要调用最高端模型。

一个高频摘要任务,也没有必要长期使用旗舰级 Agent。

但对于复杂代码重构、金融分析、自动化决策等任务,更高质量的模型又可能带来明显的效率提升。

于是,一个更加合理的 AI 算力市场应该逐渐形成:

任务类型更适合的算力选择
高频分类、摘要低成本模型
批量数据处理高性价比模型
Coding Agent强推理模型
复杂企业工作流高可靠模型
高实时性业务低延迟模型

真正重要的不是:

“哪个模型最便宜?”

而是:

“哪个模型最适合当前任务?”

这也是 Token 聚合平台的价值所在。

如果开发者必须分别接入几十家模型供应商、维护几十套 API、比较几十套价格体系,那么管理成本本身就会变成新的算力成本。

而聚合平台的意义,就是试图把这些复杂性集中到一个统一入口。


从“卖 Token”到“卖任务”

AI 行业正在发生一个非常重要的变化。

过去的问题是:

这个模型每百万 Token 多少钱?

现在的问题正在变成:

这个模型完成我的任务,需要多少钱?

再往前一步:

这个模型能不能以最低的综合成本,把我的任务稳定完成?

这就是从 Token Economics 走向 Task Economics。

OpenAI 企业营收突破 400 亿美元,是这种变化的一个重要表现。

CPCT 等任务级成本指标,是第二个表现。

Token 动态定价,则是第三个表现。

而 Unexhub 所探索的,是把这些变化进一步连接起来:

模型供应 → 算力聚合 → 智能路由 → 价格发现 → Token 消耗 → 任务完成 → 收益结算

这条链路的最终目标,并不是单纯把 Token 卖得更便宜。

而是让企业能够:

用更合理的价格获得可用的 AI 智能。


Unexhub:把 Token 交易变成算力市场

Unexhub 正在构建面向全球市场的 Token 聚合、分发与算力交易中心。

平台聚合不同规模的 AI 供应商,覆盖包括 Gemini、ChatGPT、Claude、DeepSeek、GLM 等主流 AI 模型生态,同时也吸纳个人开发者和中小企业提供的优质模型与算力资源。

通过统一接入和智能路由,开发者与企业不必为每一个模型单独建立完整的接入体系。

在需求侧,平台强调:

  • 按 Token 实际使用量付费

  • 多模型统一接入

  • 毫秒级智能路由

  • 市场化价格选择

  • 面向高频 AI 工作负载的成本优化

  • 在符合具体供应商、模型、流量及活动条件的情况下,开发者价格最低可达到市场价格的约 3 折

对于开发者而言,这意味着可以在多个模型之间灵活选择,而不必被单一厂商的固定价格体系锁定。

对于企业而言,则有机会进一步降低多模型采购、API 管理和算力调度的复杂度。

而对于供应商,Unexhub 提供:

  • 一键接入渠道

  • 聚合企业与开发者需求

  • 实时结算调用分成

  • 更灵活的 API 定价空间

  • 通过竞价机制寻找更合适的市场价格

因此,早期优惠的价值并不只是“便宜”。

它更像是一种市场启动机制:

开发者获得更低的使用门槛,供应商获得更低的获客成本,平台则获得更丰富的算力供给。

当供需两侧真正形成规模,Token 的价格发现机制才有机会发挥作用。


400 亿美元不是终点

OpenAI 的 400 亿美元,不只是一个公司的营收里程碑。

它更像是 AI 商业模式进入下一阶段的一个信号。

第一阶段,行业竞争的是:

谁拥有最强模型?

第二阶段,竞争的是:

谁能够以最低 Token 成本提供推理?

而下一阶段,真正决定商业价值的可能变成:

谁能够以更低、更稳定、更可预测的成本完成任务?

这意味着:

Token 价格下降,并不代表 AI 算力市场的价值下降。

恰恰相反。

价格越低,使用量可能越高;使用量越高,企业越愿意把更多工作交给 AI;AI 完成的任务越多,企业愿意支付的总算力费用反而可能继续增长。

这就是典型的杰文斯悖论:

效率提升,并不一定减少资源消耗,反而可能刺激更多需求。

所以,未来 AI 算力真正的竞争,不一定只是“谁的 Token 最便宜”。

而可能是:

谁能以更低的综合成本,提供更高质量、更稳定、更快速的任务完成能力。


最终的账本正在被重写

从“模型价格”到“Token 价格”,再到“任务成本”,AI 的经济计量方式正在不断升级。

Token 是计量单位。

算力是生产资源。

任务是价值载体。

市场则决定最终价格。

OpenAI 正在用企业级 AI 工作流证明“卖任务”的商业价值。

模型厂商正在通过降价、缓存和动态定价争夺 Token 需求。

而 Unexhub 所探索的,则是建立连接这些参与者的市场基础设施:

让供应商提供算力,让市场发现价格,让路由匹配需求,让开发者按实际使用付费,让供应商根据真实调用获得收益。

这意味着 AI 算力的商业模式可能正在从:

“卖模型 → 卖 Token → 卖 API”

进一步走向:

“卖算力 → 卖任务 → 卖生产力”。

400 亿美元只是账本上的一个数字。

真正正在发生变化的,是这本账到底应该怎么算。