
Bilingual Article | English First, Chinese Version Below
From Per-Minute Billing to Token-Based Pricing: Speech Transcription Is Entering the AI Compute Marketplace
For the past two years, developers building speech-transcription products have been dealing with a persistent mismatch in how AI usage is measured.
Chat models are billed by Token. Speech transcription, however, has traditionally been billed by audio duration.
Transcribe a one-hour meeting with Whisper, and the invoice may show a charge of $0.36. The number is easy to understand, but it reveals very little about the actual workload.
You do not know how much compute was consumed, how complex the audio was, or how the transcription cost compares with the downstream summarization and reasoning tasks.
Unlike a text-model invoice that may show “2,000 input Tokens, 500 output Tokens, and a cost of $0.02,” duration-based transcription gives developers only a fixed relationship between time and price.
One minute of silence and one minute of rapid, overlapping conversation may receive the same listed rate.
That simplicity is useful, but it also hides the differences in actual processing complexity.
OpenRouter’s latest audio integration introduces another option: speech transcription and text generation can now be accessed through the same gateway, while developers can choose between duration-based and Token-based pricing models.
Behind that technical change is a broader shift.
Speech, text, images, and other AI workloads are beginning to move toward a more unified system of metering, routing, and cost accounting.
OpenRouter: Two Billing Models Behind One Unified API
On July 22, OpenRouter introduced a dedicated audio transcription endpoint.
Developers can use the same Bearer Key as Chat Completions, submit base64-encoded audio, and receive a structured JSON response containing the transcription result and usage information.
Chat and transcription can therefore share:
One API key;
One Base URL;
One OpenAI-compatible request format;
One centralized usage and billing interface.
The core Base URL remains:
https://openrouter.ai/api/v1
For teams that already use an OpenAI-compatible transcription client, migration may require only a change to the Base URL, API key, and selected model rather than a complete rewrite of the application.
That matters because the real cost of integrating AI is not limited to inference.
Maintaining separate SDKs, authentication systems, retry mechanisms, usage dashboards, and provider invoices also consumes engineering and financial resources.
A unified gateway reduces those hidden costs.
More importantly, OpenRouter’s transcription catalog presents developers with two distinct pricing paths.
Route One: Whisper Models Billed by Audio Duration
The first route is the familiar Whisper-style billing model.
Models such as openai/whisper-1 are priced according to audio duration. The longer the recording, the higher the cost.
This approach is simple and predictable.
It is particularly suitable for workloads such as:
Meeting recordings;
Customer-service calls;
Podcast transcription;
Interview processing;
Subtitle generation;
Voice archives.
When a company knows it must process 10,000 minutes of audio, duration-based pricing makes budget planning straightforward.
The finance team can multiply the total number of minutes by the listed rate and obtain a reasonable estimate before the workload begins.
For fixed-volume transcription, this simplicity is valuable.
However, duration-based billing treats time as the main measure of work.
A quiet recording with long pauses and a high-speed multilingual discussion may be billed at the same rate if their duration is identical.
The price is connected to the clock, not necessarily to the amount of useful content generated or the complexity of the processing task.
Route Two: Newer STT Models Billed by Token
The second route is Token-based pricing for newer speech-to-text models.
Instead of charging only according to audio duration, the model reports input and output Token usage. The response can also include the actual request cost through usage-related fields.
This approach creates a more granular cost record.
Speech that generates a shorter transcription may consume fewer Tokens. Dense conversations, long outputs, or more complex processing may consume more.
The relationship is no longer simply:
More seconds = higher cost.
It becomes closer to:
More processing and more generated content = higher cost.
This distinction is especially useful when transcription is only the first stage of a larger workflow.
A typical enterprise pipeline may include:
Audio transcription → Speaker and topic extraction → RAG retrieval → Summarization → Reasoning → Report generation
If each stage is measured in Tokens, the business can calculate the end-to-end cost using a more consistent unit.
The company no longer needs to compare a transcription invoice measured in minutes with a summarization invoice measured in Tokens.
Both can be entered into the same usage ledger.
For developers and finance teams, that consistency may be more important than saving a fraction of a cent on a single request.
How to Find Transcription-Capable Models
Speech-to-text models may not always appear in the default model directory because transcription is a specialized output capability rather than a standard chat response.
Developers may need to apply a transcription-related output filter when querying the model catalog.
Once the appropriate model is selected, the API returns structured data containing the recognized text and request-level usage information.
That request-level visibility allows costs to be assigned more precisely to:
Individual customers;
Product features;
Departments;
Projects;
Meetings;
Call-center teams;
Internal cost centers.
Instead of receiving one large monthly audio invoice, an enterprise can see how much each workflow actually consumed.
This makes cost optimization far more actionable.
Token Pricing Is About Measurement Alignment, Not Just Changing Units
The difference between per-minute and per-Token billing is not simply a conversion from one unit to another.
It represents a change in measurement granularity.
Duration-based pricing assumes that every minute of audio carries roughly equal processing value.
Token-based pricing attempts to connect cost more closely with the amount of information processed and generated.
Consider two one-minute recordings.
The first contains a few short voice commands in a quiet environment.
The second contains several speakers talking rapidly, interrupting one another, and switching between technical terms.
Under a flat duration-based model, the listed cost may be the same.
Under a Token-based model, the second recording may generate more input or output Tokens and therefore incur a higher cost.
This creates a “pay according to workload” logic that text-model users already understand.
For short and simple audio—such as a three-second command or a clean monologue—Token-based pricing may be more economical.
For long, dense, noisy, or multilingual recordings, Token usage may rise and narrow the cost advantage.
The important change is not that Token billing will always be cheaper.
It is that developers can finally choose the pricing structure that fits the workload.
One Token Ledger for the Entire AI Workflow
Another important benefit is accounting consistency.
When chat, transcription, retrieval, and reasoning can all be measured through Token-related usage records, teams gain a more complete view of application economics.
They no longer need to ask:
How many Tokens is ten minutes of audio roughly equivalent to?
Instead, the workflow can be tracked from beginning to end:
How much did speech recognition consume?
How much did data extraction consume?
How much did RAG retrieval and reasoning consume?
How much did the final report cost?
Which model or provider contributed most to the bill?
This creates a common financial language across different AI functions.
For a small prototype, that may seem unnecessary.
For an enterprise processing millions of audio files, customer conversations, or meeting records, it becomes essential.
A unified ledger allows the company to optimize not only individual model prices, but also the total cost of a completed task.
OpenRouter’s Pricing Structure Remains Relatively Transparent
OpenRouter separates the underlying model price from its platform service fee.
The first layer is the price of the model or infrastructure provider. OpenRouter generally presents that price without adding a separate markup directly to each Token.
The second layer is the platform fee collected when users purchase credits.
The listed rate is approximately 5.5%, or around 5% for cryptocurrency payments, with a minimum fee for smaller transactions.
For BYOK users who bring their own provider keys, the first one million monthly requests may be processed without an additional platform fee. Usage beyond that range is charged according to the applicable percentage-based structure.
Failed requests are generally not billed as successful inference.
This model makes the division of costs easier to understand:
The supplier receives payment for model inference;
OpenRouter earns revenue from aggregation, routing, billing, and platform access.
By mid-2026, OpenRouter reportedly served more than eight million developers and enterprise users and processed up to 25 trillion Tokens per week.
Its growth demonstrates the demand for a unified access layer.
However, unified access does not automatically create a true market price for AI capacity.
OpenRouter Solves “How to Call,” but Not Fully “How to Price”
OpenRouter answers an important technical question:
How can developers call chat, transcription, and other models without maintaining several different systems?
Its answer is one key, one Base URL, and one compatible interface.
But the underlying price is still primarily determined by suppliers.
A provider publishes a price, and the developer decides whether to use it. A routing platform may compare several offers and choose a cheaper provider, but the prices are still generally fixed quotations.
They do not necessarily move in real time according to market supply and demand.
That is the layer Unexhub is attempting to develop.
Unexhub: Turning Token Capacity into a Tradable Resource
Unexhub is building a Token aggregation and distribution marketplace that connects model suppliers, compute providers, enterprise customers, and developers.
Its objective goes beyond offering one API for several models.
It aims to make Token capacity a resource that can be:
Aggregated from multiple suppliers;
Competitively priced;
Dynamically routed;
Accurately metered;
Rapidly settled.
The platform includes major domestic and international model ecosystems such as Gemini, ChatGPT, Claude, DeepSeek, and GLM.
It also creates distribution opportunities for independent developers, small and medium-sized AI companies, specialized model providers, and teams with underused inference capacity.
When several suppliers can serve a similar request, the platform can compare them according to price, latency, throughput, availability, and service quality.
The request can then be sent to an appropriate supplier through millisecond-level routing.
For developers, the experience remains simple:
One request, pay for the actual Tokens consumed, and let the platform handle supplier matching in the background.
Faster Settlement Can Improve Supplier Cash Flow
For smaller compute and model suppliers, one of the biggest barriers is the settlement cycle.
A supplier may complete large volumes of inference today but wait weeks before receiving payment.
During that period, it must continue paying for GPU resources, cloud infrastructure, bandwidth, maintenance, and technical operations.
Unexhub aims to shorten the cycle through usage-based metering and faster settlement.
Once the platform verifies the Token consumption and successful completion of a request, the corresponding supplier revenue can be calculated and distributed more quickly.
The practical benefits include:
Faster recovery of infrastructure costs;
Lower working-capital pressure;
Better cash-flow visibility;
More funds available for capacity expansion;
A stronger incentive to provide idle compute to the market.
Early suppliers may also qualify for promotional commission rates as low as 0%.
This reduces the initial cost of joining the distribution network and gives smaller providers a chance to reach global developer and enterprise demand without building their own sales, billing, and clearing infrastructure.
The more accurate claim is not that all supplier financial pressure immediately disappears.
Rather, faster settlement and lower early-stage channel costs can materially improve capital turnover.
API Bidding: Allow Prices to Respond to Supply and Demand
One of Unexhub’s more ambitious ideas is an API bidding mechanism.
Instead of keeping every model at one fixed price throughout the day, available capacity can be priced according to current supply and demand.
The logic resembles an electricity spot market.
During peak demand, electricity becomes more expensive because available supply is under pressure.
During low-demand periods, excess capacity can be sold at a lower price.
AI inference has similar patterns.
Weekday business hours often generate heavier enterprise traffic. Late nights and weekends may leave large amounts of compute underused.
With dynamic pricing:
Suppliers can earn more when demand is high;
Idle capacity can attract workloads through lower prices;
Developers can schedule non-urgent tasks during cheaper periods;
The market can discover a more realistic price for available capacity.
This can be especially useful for workloads that do not need to run immediately:
Batch transcription;
Subtitle generation;
Archive processing;
Dataset labeling;
Document extraction;
Offline evaluation;
Large-scale summarization.
Instead of manually comparing supplier prices, developers can allow the platform to select a lower-cost route automatically.
Can Token-Based Transcription Really Reduce Costs by 90%?
A claim such as “reduce costs by 90%” needs to be examined carefully.
It is possible in certain workloads, but it should not be treated as a universal result.
Consider a simple example using gpt-4o-mini-transcribe.
Its listed pricing is approximately:
$1.25 per million input Tokens;
$5 per million output Tokens.
Suppose a 9.2-second audio sample produces 113 total Tokens: 83 input and 30 output.
If that ratio is projected to one minute, the audio would generate approximately 737 Tokens, including about 542 input Tokens and 195 output Tokens.
The estimated cost would be:
Input: 542 ÷ 1,000,000 × $1.25;
Output: 195 ÷ 1,000,000 × $5;
Total: approximately $0.00165 per minute.
By comparison, whisper-1 at $0.006 per minute would cost more than three times as much in this specific example.
The potential saving is substantial, but it is not yet 90%.
Reaching a reduction closer to 90% would require additional advantages, such as:
Short and simple audio;
Lower-priced model routes;
High cache reuse in downstream stages;
Off-peak market discounts;
Batch processing;
Lower-cost suppliers available through dynamic routing.
Under the right combination of conditions, reducing the total workflow cost to roughly one-tenth of a traditional setup may be achievable.
However, long, noisy, multilingual, or highly complex audio may generate more Tokens and reduce—or eliminate—the cost advantage.
The honest conclusion is:
Token billing creates an optimization opportunity, not an automatic discount.
The developer must compare actual audio characteristics, model quality, latency, and total workflow cost.
The Most Valuable Benefit Is Choice
The significance of having both duration-based and Token-based models is not that one permanently replaces the other.
It is that developers gain the ability to choose.
For predictable, fixed-length audio, duration billing may remain the simplest option.
For short commands, clean monologues, or transcription embedded inside a larger Token-based workflow, Token pricing may offer better economics and accounting consistency.
A mature routing system could eventually make that choice automatically.
Before each request, it could evaluate:
Audio duration;
Estimated complexity;
Required accuracy;
Available providers;
Current market price;
Latency requirements;
Downstream workflow needs.
It could then route the request to either a duration-priced or Token-priced model according to the expected cost per completed task.
At that point, developers would no longer need to choose a billing system manually.
The infrastructure would optimize the decision in real time.
OpenRouter and Unexhub: Two Layers of the Same Transformation
OpenRouter and Unexhub address different parts of the same market evolution.
OpenRouter solves the access layer.
It brings chat, speech transcription, and other modalities behind one key, one Base URL, and one compatible API.
Its value is reducing technical fragmentation and making usage easier to track.
Unexhub focuses on the marketplace and settlement layer.
It aims to aggregate suppliers, enable competitive pricing, route requests in milliseconds, meter actual Token usage, and distribute revenue more quickly.
Its value is improving market efficiency.
For developers and enterprise users, Unexhub’s proposed benefits include:
Unified access to major domestic and international models;
Pay-as-you-go billing based on actual Token consumption;
Millisecond-level routing across suppliers;
More flexible selection between cost, speed, and quality;
Selected services available at prices as low as approximately 30% of typical market rates;
Lower integration, account-management, and reconciliation costs.
For suppliers, the platform can provide:
One-step access to a broader distribution channel;
Automated traffic allocation;
Usage-based revenue sharing;
Faster settlement;
Early-stage commission incentives as low as 0%;
A way to monetize idle or underused capacity.
One platform lowers the technical barrier.
The other attempts to improve price discovery and capital flow.
Together, they point toward the same future:
AI capacity is becoming standardized, measurable, routable, and tradable infrastructure.
Final Takeaway
The shift from per-minute billing to Token-based transcription is not merely a change of units.
It is a change in cost logic.
Under duration-based billing, the clock determines the charge.
Under Token-based billing, the amount of processing and generated content plays a larger role.
When speech transcription and text generation can be measured through a more consistent ledger, the cost structure of an AI application becomes easier to understand and optimize.
OpenRouter gives developers a choice between different transcription and billing models.
Unexhub points toward a broader marketplace in which Token capacity can be competitively priced, routed in real time, and settled according to actual usage.
In that environment, a significant cost reduction would not depend on one provider permanently cutting its list price.
It could emerge from a combination of:
Better model selection;
More granular measurement;
Competitive supplier routing;
Off-peak pricing;
Faster settlement;
Higher utilization of idle compute.
Token is moving beyond text generation and becoming a common unit for measuring AI services across speech, images, documents, and multimodal workflows.
Speech transcription billed by Token is only one visible part of that transition.
The deeper change is that AI compute is beginning to behave less like a collection of fixed-price APIs and more like a continuously routed, priced, and settled market.
中文版|English Version Above
本文为英中双语文章。以上为英文全文,以下为完整中文版本。
从按分钟计费到按 Token 定价:语音转录正在进入 AI 算力交易时代
过去两年,专做语音转录的开发者一直被一种“计量错位”所困扰。
聊天模型按 Token 计费,语音转录却长期按照音频时长收费。
用 Whisper 转写一场一小时的会议,账单上可能写着“0.36 美元”。这个数字简单易懂,却很难说明背后的实际工作量。
你不知道模型处理了多少有效信息,也无法看清这段音频究竟消耗了多少算力,更难将转录成本与后续的摘要、检索和推理费用放在一起比较。
文本模型的账单可能清楚地显示:
输入 2,000 Token,输出 500 Token,本次调用花费 0.02 美元。
而传统语音转录通常只告诉你处理了多少分钟。
一分钟静默和一分钟多人快速对话,可能对应相同的目录价格。
这种计费方式简单直接,却也掩盖了音频复杂度与实际处理成本之间的差异。
OpenRouter 新增的音频转录能力,为开发者提供了另一种选择:语音转录和文本生成可以通过同一个入口调用,同时支持按时长和按 Token 两种计费路线。
在这次接口变化背后,语音、文本、图像等 AI 工作负载正在走向更统一的计量、路由和成本核算体系。
OpenRouter:一个统一 API,两种计费模式
7 月 22 日,OpenRouter 正式推出音频转录端点。
开发者可以继续使用 Chat Completions 的 Bearer Key,提交经过 base64 编码的音频,并获得包含转录文本和实际用量信息的结构化 JSON。
聊天与转录由此可以共享:
一个 API Key;
一个 Base URL;
一套 OpenAI 兼容请求格式;
一个统一的用量与账单入口。
其核心 Base URL 仍然是:
https://openrouter.ai/api/v1
如果团队此前已经接入 OpenAI 兼容的转录客户端,迁移通常不需要推翻原有代码。
多数情况下,只需更换 Base URL、API Key 和模型名称,即可完成切换。
这项设计的价值并不只是少写几行代码。
AI 接入的真实成本,不只有模型调用费用,还包括多套 SDK、密钥管理、重试机制、监控系统和账单核对所占用的工程资源。
统一网关可以显著减少这些隐性成本。
更值得关注的是,OpenRouter 在语音模型层面同时呈现了两种计费路线。
第一条路线:Whisper 按音频时长收费
第一条路线是开发者熟悉的 Whisper 计费模式。
openai/whisper-1 等模型按照音频长度计算费用。音频越长,价格越高。
这种模式最大的优势,是简单、直观、容易预估。
它尤其适合以下场景:
会议录音;
客服通话;
播客转录;
访谈整理;
字幕生成;
历史语音归档。
当一家企业明确知道每月需要处理 1 万分钟录音时,财务人员可以直接用总时长乘以单价,提前估算转录预算。
对于音频数量与时长较为固定的业务,这种确定性非常有价值。
但按时长计费的核心,是把“时间”当成主要工作量单位。
一段包含大量停顿的安静录音,和一段高语速、多人交叉发言的讨论,只要时长相同,目录价格就可能完全一样。
成本由时钟决定,而不一定与实际信息密度、输出长度和处理复杂度直接对应。
第二条路线:新一代 STT 模型按 Token 计费
第二条路线,是新一代语音转文字模型采用的 Token 计费方式。
模型完成音频处理后,会返回输入与输出 Token 用量,响应中的 usage 相关字段还可以显示本次请求的实际费用。
与按分钟计费相比,这种方式提供了更细的成本记录。
输出内容越长、处理任务越复杂,Token 用量通常越高;音频内容越简单、生成文字越短,成本则可能更低。
计费逻辑不再只是:
时间越长,成本越高。
而更接近:
处理的信息越多、生成的内容越多,成本越高。
当语音转录只是完整 AI 工作流中的第一步时,这种方式尤其有价值。
一条典型的企业工作流可能包括:
音频转录 → 说话人和主题提取 → RAG 检索 → 摘要生成 → 深度推理 → 报告输出
如果每个环节都可以用 Token 记录,企业就能够通过相对统一的单位,计算整条任务链的综合成本。
财务团队不再需要把按分钟计算的转录费用,与按 Token 计算的摘要费用拼在两套表格中。
所有消耗都能进入同一本 Token 台账。
对于开发者和成本管理团队来说,这种统一性往往比一次请求便宜几厘钱更重要。
如何筛选支持转录的模型
语音转文字模型不一定会直接出现在默认模型目录里,因为 transcription 属于专门的输出能力,而不是标准聊天响应。
开发者在查询模型目录时,可能需要通过与 transcription 相关的输出条件进行主动筛选。
选择模型后,接口会返回结构化数据,包括识别文本和请求级别的实际用量。
这种细粒度的记录,可以帮助企业把费用精确分配到:
不同客户;
不同产品功能;
不同部门;
不同项目;
不同会议;
不同客服团队;
不同成本中心。
企业不再只能在月底看到一笔笼统的语音账单,而是可以追踪每一项工作流消耗了多少费用。
成本优化也因此从“猜测”变成了可执行的决策。
按 Token 计价的本质,是计量对齐
按 Token 和按分钟之间的差别,并不只是换了一个计量单位。
真正的变化,是计量粒度发生了改变。
按分钟计费默认每一分钟音频具有相近的处理价值。
按 Token 计费则尝试让成本与信息处理量和生成内容量建立更直接的关系。
假设有两段同样长达一分钟的录音。
第一段只有几条简短的语音指令,环境安静,停顿较多。
第二段包含多名参与者的快速讨论,彼此打断,并且涉及大量专业术语。
在固定时长计费模式下,两段录音的价格可能相同。
在 Token 计费模式下,第二段可能产生更多输入或输出 Token,因此费用也会更高。
文本模型中已经常见的“按工作量付费”逻辑,开始进入语音转录领域。
对于三秒语音指令、安静环境中的简短独白等场景,Token 计费可能更具成本优势。
对于时间较长、信息密度高、噪声较多或包含多种语言的录音,Token 用量则可能上升,使成本优势缩小。
真正重要的变化,并不是按 Token 一定更便宜。
而是开发者终于可以根据业务特点,选择更适合自己的计费方式。
一套 Token 台账,核算完整 AI 工作流
另一个容易被忽略的价值,是计量口径的统一。
当聊天、转录、检索和推理都能通过 Token 相关记录进行核算时,团队就可以获得更完整的应用成本结构。
过去经常需要估算:
10 分钟音频大概相当于多少个 Token?
现在则可以沿着整条任务链直接记录:
语音识别消耗了多少?
信息提取消耗了多少?
RAG 检索和推理消耗了多少?
最终报告总成本是多少?
哪个模型或供应商占据了最多预算?
这样,文本、音频和其他 AI 功能便拥有了一套相对统一的财务语言。
对于小型原型项目来说,这种能力可能并不显眼。
但对于每天处理海量录音、客户对话和会议材料的企业而言,它会直接影响预算控制、产品定价和毛利率计算。
企业优化的不再只是单个模型的价格,而是完成一项业务任务的总成本。
OpenRouter:模型价格与平台费用分开
OpenRouter 将模型费用和平台服务费分别呈现。
第一层是模型或基础设施供应商的价格。平台通常按照供应商目录价格展示,不在每个 Token 上额外叠加一层明显加价。
第二层是平台服务费用。
用户购买 credits 时,需要支付约 5.5% 的服务费;使用加密货币支付时约为 5%,小额交易还可能存在最低收费。
对于 BYOK,也就是自带供应商 API Key 的用户,每月前 100 万次请求可以获得免平台费额度,超过后再按照相应规则收取费用。
失败请求通常不会被当作成功推理计费。
这套结构让成本来源更加容易理解:
模型供应商获得推理服务收入;
OpenRouter 通过聚合、路由、账单与平台接入获得收益。
截至 2026 年中,OpenRouter 据称已经服务超过 800 万开发者和企业用户,每周处理最高约 25 万亿 Token。
它的增长证明,开发者确实需要一套统一的模型入口。
但统一调用,并不等于已经形成真正的 AI 算力市场价格。
OpenRouter 解决“怎么调”,却没有完全解决“怎么定价”
OpenRouter 回答了一个重要的技术问题:
如何在不维护多套系统的情况下,同时调用聊天、转录和其他模型?
它给出的答案,是一个 Key、一个 Base URL 和一套兼容接口。
不过,底层模型价格仍然主要由供应商制定。
供应商公布价目表,开发者决定是否购买。路由平台可以比较多家报价并选择更低成本的节点,但价格通常仍然是相对固定的供应商报价。
它不会持续根据市场供需自动形成实时成交价格。
这正是 Unexhub 希望进一步推动的方向。
Unexhub:让 Token 成为可交易的算力资源
Unexhub 正在建设 Token 聚合与分发的算力交易中心,用于连接模型供应商、算力提供方、企业用户和开发者。
它的目标不只是通过一个接口提供多个模型。
平台希望让 Token 算力成为一种可以被:
多方聚合;
公开竞价;
动态路由;
精确计量;
快速结算的市场资源。
Unexhub 聚合 Gemini、ChatGPT、Claude、DeepSeek、GLM 等海内外主流模型生态,也为个人开发者、中小型 AI 企业、垂直模型服务商和拥有闲置推理资源的团队提供分发渠道。
当多家供应商能够处理相近请求时,平台可以结合价格、延迟、吞吐、稳定性和服务质量进行比较。
随后,通过毫秒级路由,将请求发送到更合适的供应商。
开发者看到的操作依旧简单:
发送一次请求,按照实际 Token 消耗付费,供应商匹配与路由由平台自动完成。
更快结算,改善供应商资金周转
对于中小型模型和算力供应商而言,较大的障碍之一,是结算周期。
供应商今天已经提供了大量推理服务,却可能要等待数周才能收到款项。
在等待期间,GPU、云服务、带宽、模型维护和技术运维费用仍然需要持续支付。
Unexhub 希望通过按调用量计量和更快结算,缩短资金回收周期。
当平台确认请求成功完成并核对 Token 用量后,可以更快计算供应商应得收入,并推进收益结算。
具体价值包括:
更快回收基础设施费用;
降低营运资金压力;
提高现金流可见性;
为算力扩容提供更多周转资金;
增强供应商释放闲置资源的意愿。
早期入驻供应商还有机会享受平台最低 0 佣金的阶段性权益。
这能够降低中小型供应商进入分发网络的初始成本,使其无需自行搭建完整的获客、计费、路由和清算系统,也能触达更多企业与开发者。
更准确地说,快速结算无法让所有财务压力瞬间归零,却可以明显改善供应商的资金周转效率。
API 竞价:让价格跟随供需变化
Unexhub 正在推进的另一项机制,是 API 竞价。
平台不要求一种模型在所有时间维持完全固定的价格,而是允许算力报价根据实时供给和需求进行变化。
这套逻辑类似电力现货市场。
需求高峰时,可用电力减少,价格上升;需求低谷时,过剩资源可以通过降价获得更多订单。
AI 推理同样具有周期性。
工作日白天往往是企业 API 调用高峰,深夜和周末则可能存在大量闲置算力。
在动态定价机制下:
供应商可以在需求旺盛时获得更高收益;
闲置算力可以通过更低报价吸引任务;
开发者能够把非紧急任务安排在低价时段;
市场可以逐渐形成更真实的价格发现机制。
这类模式尤其适合无需立即完成的工作:
批量语音转录;
字幕生成;
历史录音处理;
数据集标注;
文档信息提取;
离线模型评测;
大规模摘要生成。
开发者无需手动比较数十家供应商,系统可以在后台自动匹配成本更低的可用路径。
“降本 90%”,这笔账应该怎么算?
“降低 90% 成本”听起来非常有吸引力,但它只可能出现在特定场景中,不应被理解为所有任务的固定结果。
以 gpt-4o-mini-transcribe 为例,其参考价格约为:
输入 1.25 美元/百万 Token;
输出 5 美元/百万 Token。
假设一个 9.2 秒的音频示例产生 113 个 Token,其中输入 83 个、输出 30 个。
按照相同比例推算,一分钟音频大约产生 737 个 Token,其中输入约 542 个、输出约 195 个。
估算费用为:
输入:542 ÷ 1,000,000 × 1.25 美元;
输出:195 ÷ 1,000,000 × 5 美元;
合计约为 0.00165 美元/分钟。
相比之下,whisper-1 按 0.006 美元/分钟计费,在这个特定样本中,Token 计费成本不到其三分之一。
节省幅度已经相当明显,但还没有达到 90%。
要进一步将综合成本压缩到原来的十分之一,通常还需要同时满足多项条件:
音频较短且内容简单;
使用价格更低的模型或供应节点;
后续工作流拥有较高缓存命中率;
在闲时获得 API 竞价折扣;
使用批量处理;
通过动态路由匹配到更低成本的供应商。
在这些条件叠加时,特定业务的端到端成本确实可能下降至传统方案的约十分之一。
但对于长时间、高噪声、多语言或信息密度很高的录音,Token 消耗会明显增加,成本优势可能缩小,甚至不再存在。
更准确的结论是:
按 Token 计费提供了降本空间,而不是自动提供固定折扣。
开发者仍然需要结合音频特征、识别质量、响应延迟和完整任务成本进行比较。
最大的权益,不是便宜,而是拥有选择权
两种计费模式并行存在的意义,不是其中一种必须彻底取代另一种。
而是开发者拥有了选择权。
对于时长稳定、预算需要提前锁定的音频,按分钟计费仍然简单可靠。
对于短语音、清晰独白,或者需要进入后续 RAG、摘要和推理流程的任务,按 Token 计费可能更具经济性,也更方便统一核算。
未来,成熟的路由系统甚至可以自动做出选择。
在每次请求之前,平台可以评估:
音频长度;
预估复杂度;
识别精度要求;
当前可用供应商;
实时市场价格;
延迟要求;
后续工作流类型。
随后,系统可以根据预计的单位任务成本,自动选择按分钟或按 Token 计费的模型。
届时,开发者甚至不需要手动决定采用哪种定价方式。
基础设施会实时完成优化。
OpenRouter 与 Unexhub:同一场变革的两个层面
OpenRouter 与 Unexhub,分别解决了这场行业变化中的不同问题。
OpenRouter 负责接入层。
它将聊天、语音转录和其他模态放进同一个 Key、Base URL 与兼容接口中,降低技术碎片化,也让用量更容易被统一统计。
Unexhub 关注交易与结算层。
它希望聚合不同供应商,支持竞价报价,通过毫秒级路由分配请求,按照实际 Token 用量计费,并更快完成收益结算。
对开发者和企业用户而言,Unexhub 提供的价值包括:
统一接入海内外主流模型;
按实际 Token 消耗付费;
在多家供应商之间实现毫秒级路由;
根据成本、速度与质量灵活选择;
部分模型服务最低可至市场常规价格的约 3 折;
减少接口维护、多账号充值和月底对账成本。
对供应商而言,平台则可以提供:
一键接入更广泛的分发渠道;
自动获得适合自身能力的调用流量;
按实际用量获得收益分成;
更快完成收入结算;
早期最低 0 佣金的阶段性激励;
将闲置或未充分利用的算力转化为收入。
一个降低技术门槛,一个改善价格发现和资金流转。
两者最终指向同一个方向:
AI 算力正在成为标准化、可计量、可路由、可交易的基础设施。
说到底
从“按分钟收费”转向“按 Token 计价”,变化的并不只是单位。
而是成本逻辑。
按分钟收费,价格主要由时钟决定。
按 Token 计价,信息处理量和生成内容量会更加直接地参与成本计算。
当语音转录与文本生成能够进入一套更统一的计量体系后,AI 应用的完整成本结构才更容易被看清和优化。
OpenRouter 将两种模型和计费路线的选择权交给开发者。
Unexhub 则指向更进一步的市场形态:Token 算力可以被多方聚合、实时竞价、自动路由,并按照真实用量完成结算。
在这种环境中,大幅降本不必完全依赖某一家供应商永久下调目录价格。
它可以来自多项机制的共同作用:
更合适的模型选择;
更精细的用量计量;
多家供应商竞价;
闲时价格优惠;
毫秒级动态路由;
更高的闲置算力利用率。
Token 正在跨越文本、语音、图像和文档等不同模态,逐渐成为衡量 AI 服务消耗的重要基础单位。
语音转录按 Token 计价,只是这场变化中一个清晰可见的切面。
更深层的改变是:AI 算力正在从一组彼此割裂、价格固定的 API,走向一个可以持续路由、动态定价并高效结算的市场。