跳到主内容
@wquguru
精选90SaaStr 博客(RSS)产品与增长

Agent API 用量激增致成本飙升,自建 Postgres 镜像替代方案

We Got a $240,000 Estimate for Agent API Access. Our Agent Suggested a $5 Postgres Instance.

原文
发到 X
推荐理由

直接给出了应对 Agent API 计费的具体技术架构调整方案(自建 Postgres 镜像)及成本权衡逻辑,创业者可立即用于优化自身产品的单位经济与集成策略。

We run SaaStr with 3 humans and 21+ agents. 10K, our AI VP of Revenue, started as a simple dashboard and now runs most of our outbound, inbound and revenue ops. Across all the apps it touches, it makes 35,000 to 40,000 API calls a day.

我们运营 SaaStr 团队由 3 名人类员工和 21 个以上的 AI 智能体组成。我们的 AI 营收副总裁 10K,最初只是一个简单的仪表盘,如今负责处理我们大部分的外展、 inbound(入站)和营收运营工作。在它涉及的所有应用中,每天产生 35,000 到 40,000 次 API 调用。

None of our vendors ever charged us for that. Now Salesforce, Atlassian and HubSpot are all adding charges for agent access, and one estimate we got puts our cost at as much as $240,000 a year.

我们的供应商从未为此向我们收费。现在 Salesforce、Atlassian 和 HubSpot 都在为智能体访问添加费用,我们获得的一个估算显示,我们的成本可能高达每年 240,000 美元。

Amelia and I spent most of the latest episode of The Agents on what this looks like as a buyer. We also covered an analysis 10K ran on 22,000 paid SaaStr AI Annual attendees, which showed 49.7% of them changed jobs in the last 16 months.

Amelia 和我在《The Agents》最新一期中花了大量时间讨论作为买家这看起来是什么样子。我们还介绍了 10K 对 22,000 名付费参加 SaaStr AI 年度会议的参会者进行的分析,结果显示其中 49.7% 的人在过去的 16 个月内换了工作。

Salesforce, HubSpot and Atlassian Are All Charging for Agent Access

Salesforce、HubSpot 和 Atlassian 都在为智能体访问收费

Salesforce is rolling it out now, with a banner at login telling us we’ll pay more for agent access. Atlassian already did it, and it’s one of the reasons they’ve seen growth. HubSpot is doing it as well, but so far it looks aimed at their own first-party agents. They don’t appear to be charging yet for third-party agents like 10K, which I’d prefer.

Salesforce 正在推出这一举措,登录时会有横幅提示我们将为智能体访问支付更多费用。Atlassian 已经实施了这一政策,这也是他们实现增长的原因之一。HubSpot 也在这样做,但目前看来主要针对的是他们自己的第一方智能体。对于像 10K 这样的第三方智能体,他们似乎尚未开始收费,我对此更倾向于前者。

I get why. Every seat-based vendor is seeing contraction because of agents. We have one API seat at Salesforce and we don’t need 20 or 80. If I were running a B2B company that had been around for years and saw seat expansion at risk, we’d be having the same conversation and we’d probably land on the same answer.

我理解他们的做法。每个基于席位(seat-based)的供应商都因为智能体的出现而看到用户收缩。我们在 Salesforce 只有一个 API 席位,并不需要 20 个或 80 个。如果我是一家经营多年的 B2B 公司的负责人,看到席位扩张面临风险,我也会进行同样的对话,并且很可能得出相同的结论。

And in B2B, price increases have mostly worked. Customers complain about the 11th increase in a row, and more of them stay than leave. Bending Spoons built a business on this. They buy Evernote, double or triple the price, and make much more from the customers who stay than they lose from the ones who churn.

在 B2B 领域,涨价大多奏效。客户会抱怨连续第 11 次涨价,但留下来的人比离开的人多。Bending Spoons 就是基于此建立起来的商业模式。他们收购 Evernote,将价格翻倍或翻三倍,从留下的客户那里获得的收益远大于因流失客户而损失的收益。

One Estimate: $240,000 a Year to Keep Running 10K the Way We Run It

一项估算:按我们目前的方式运行 10K,每年需花费 240,000 美元

We don’t know yet exactly what each vendor will charge. The one estimate we have is as much as $240,000 a year to keep using 10K the way we do now. We’re not a Fortune 10 company, and we’re not going to pay that, or quadruple what we pay a vendor, for API access.

我们还不知道每家供应商具体会收取多少费用。我们手头唯一的估算是,保持像现在这样使用 10K,每年最多可能需要 240,000 美元。我们不是财富 10 强公司,我们不会支付这笔钱,也不会让供应商的费用翻四倍来获取 API 访问权限。

Some of this is on us. 10K was never constrained because nobody charged for API calls, so we used them like gas in a country where it costs 5 cents a gallon. And 40,000 calls a day costs almost nothing at what a modern database like Supabase charges.

这部分责任在我们自己。由于过去没有人对 API 调用收费,10K 从未受到限制,所以我们像在一个每加仑汽油只要 5 美分的国家里使用汽油一样随意使用它们。而在 Supabase 等现代数据库的收费标准下,每天 40,000 次调用的成本几乎可以忽略不计。

10K’s First Suggestion: Mirror the Data to a $5 Postgres Instance

10K 的第一个建议:将数据镜像到一个 5 美元的 Postgres 实例上

Once we knew metering was coming, Amelia had 10K install a tracker on its own API usage for a week. 10K came back and said a lot of those calls could be cut, down to a bare minimum, as a short-term fix.

一旦我们得知计量收费即将实施,Amelia 就让 10K 在其自身的 API 使用量上安装了一个追踪器,为期一周。10K 反馈称,许多调用可以削减,作为短期解决方案,可降至最低限度。

Its longer-term suggestion was to mirror the system of record into our own Postgres database and stop calling the vendor as much. It offered to set that up the same day.

其长期建议是将记录系统(system of record)镜像到我们要自己的 Postgres 数据库中,并减少对供应商的调用。它表示可以在当天就完成设置。

That has costs. We’d have a system of record plus a mirror, they’d need to sync, and they’d get out of sync sometimes. But a $5 Postgres instance with no API limits against $240,000 a year is an easy call for an agent. Other people have written about this as the big risk for systems of record. 10K proposed it as soon as it heard about the pricing.

这需要成本。我们将拥有一个记录系统加上一个镜像,它们需要同步,并且有时会不同步。但对于一个代理来说,在零 API 限制的 5 美元 Postgres 实例与每年 24 万美元之间做出选择是显而易见的。其他人曾将此描述为记录系统的重大风险。10K 在听到定价方案后便立即提出了这一建议。

Marketo Gave Our Agent 10 to 20 Minutes of API a Day

Marketo 每天仅给我们的代理提供 10 到 20 分钟的 API 使用时间

We’ve already lived the extreme case. Marketo would go down on us after 10 or 20 minutes each day because of API limits. It was a mission-critical app that our agent could only reach for 20 minutes a day, so we migrated off it.

我们已经经历了极端情况。由于 API 限制,Marketo 每天会在我们使用 10 或 20 分钟后停止响应。这是一个任务关键型应用,我们的代理每天只能访问 20 分钟,因此我们迁移离开了它。

I think metered pricing will get the same reaction from agents, just more slowly.

我认为代理会对计量定价产生同样的反应,只是速度会更慢一些。

We Might Not Pick Any of Our Current Stack Starting From Scratch

如果我们从零开始构建,可能不会选择当前技术栈中的任何组件

Staying with vendors we already have is one decision, and the Bending Spoons math may hold there for a while. Buying new is a different decision. I doubt an agent would recommend any system of record that charges materially for API access. If we started from zero today knowing what agent access might cost, we might use none of our current stack. I’m not sure the incumbents see the opening that creates.

继续使用现有的供应商是一个决定,而 Bending Spoons 的数学模型在那里可能暂时仍然成立。购买新服务则是另一个决定。我怀疑代理会推荐任何对 API 访问收取实质性费用的记录系统。如果今天我们从零开始,考虑到代理访问可能产生的成本,我们可能根本不会使用当前的任何技术栈。我不确定 incumbents(现有巨头)是否看到了由此产生的机会。

Some vendors are going the other way. Dharmesh Shah talks all the time about how important it is for HubSpot not to charge for general agent access. I’d bet a lot of people in revenue roles there disagree with him. At a sponsor and VIP event we held around Dreamforce, the CEO and CTO of Aurasell told us the last thing they’d do is charge extra for agent access. They started as a next-generation CRM competitor and are now more of a layer on top of Salesforce and HubSpot, and they want people using it as much as possible.

一些供应商正走向相反的方向。Dharmesh Shah 经常谈论 HubSpot 不对通用代理访问收费的重要性。我敢打赌,那里从事收入相关工作的许多人并不赞同他的观点。在我们围绕 Dreamforce 举办的一场赞助商和 VIP 活动中,Aurasell 的 CEO 和 CTO 告诉我们,他们绝不会对代理访问收取额外费用。他们最初是作为下一代 CRM 竞争对手起步的,现在更像是 Salesforce 和 HubSpot 之上的一个层,他们希望人们尽可能多地使用它。

Muse Is Free and Does Work We Used to Buy $50K to $100K Apps For

Muse 免费且能完成我们过去花费 5 万至 10 万美元购买的应用程序所做的工作

Muse is subsidized by Meta. It probably costs them at least a couple of dollars a month per user to run the VMs, and it comes with LLMs, a database and memory built in.

Muse 由 Meta 补贴。运行虚拟机每月每个用户可能至少花费他们几美元,并且它内置了 LLM、数据库和内存。

We use it for customer tracking and competitor tracking, and we’re starting to run ads in it. Those are things $50,000 and $100,000 B2B apps do for us today.

我们用它进行客户跟踪和竞争对手跟踪,并且我们开始在其中投放广告。这些正是今天价值 5 万和 10 万美元的 B2B 应用程序为我们做的事情。

David on our team doesn’t use 10K and doesn’t log into our AI SDRs. Amelia programs those for him. But he started building his own sales workflows in Muse without anyone asking, because he already had a Meta account and it was easy.

我们团队的David不使用10K,也不登录我们的AI SDR。Amelia为他配置这些工具。但他开始在没有任何人要求的情况下在Muse中构建自己的销售工作流,因为他已经有Meta账户,而且操作很简单。

It took Amelia weeks to find a use case because Muse couldn’t run ads. That shipped this week, for Facebook platform ads only. She’s running it head to head against Claude and Replit, which is how she runs our ads today, and we’ll share results next episode.

Amelia花了数周时间才找到一个用例,因为Muse无法投放广告。该功能本周上线,仅支持Facebook平台广告。她正将其与Claude和Replit进行A/B测试,这也是我们今天投放广告的方式,我们将在下一集分享结果。

The Cloud Index Is Up 18% and Seat-Only Vendors Are Still Getting Crushed

云指数上涨18%,而仅按席位销售的厂商仍遭重创

A third of the way through this year, every B2B software stock was down no matter how the company was doing. The cloud index is now up 18% for the year.

今年已过三分之一,无论公司业绩如何,所有B2B软件股票均处于下跌状态。目前云指数今年已上涨18%。

It’s very bifurcated. Vendors that only sell per seat and have nothing agentic are still struggling, and Monday and HubSpot are in that group. Okta, Atlassian and Cloudflare are on fire because they’re AI and agent friendly.

市场分化严重。仅按席位销售且没有任何代理(agentic)能力的厂商仍在挣扎,Monday和HubSpot就属于这一类。Okta、Atlassian和Cloudflare则表现强劲,因为它们对AI和代理友好。

22,000 Paid Attendees, and Half Changed Jobs in 16 Months

付费参会者达22,000人,其中一半人在16个月内换了工作

Amelia was doing outbound to our paid ticket holders and ran the list through ZoomInfo and Clay. Then she had 10K analyze it.

Amelia向我们付费票持有者进行外联,并将名单通过ZoomInfo和Clay进行处理。然后她让10K对其进行分析。

The data set is 22,000+ paid attendees from SaaStr AI Annual 2025, 2026 and early 2027 buyers, with free passes, sponsors and speakers removed. The caveat is that these are people who paid and traveled, many from across the world, to an event that has changed a lot. People who are less AI focused may be staying home.

数据集包含来自SaaStr AI年度大会2025、2026及2027年初的22,000多名付费参会者(买家),已剔除免费通行证持有者、赞助商和演讲嘉宾。需要注意的是,这些人是付费并前往现场参会的人,许多人来自世界各地,而该活动本身已发生巨大变化。那些对AI关注度较低的人可能选择留在家中。

1. AI-native titles and companies are up 4x in one year. Between May 2025 and May 2026, attendees with AI in their title or at a company with AI in its name became about four times more common.

1. AI原生职位和公司数量一年内增长4倍。2025年5月至2026年5月期间,头衔中包含AI或在公司名称中包含AI的参会者变得更为常见,比例约为原来的四倍。

2. 25% of our sponsors are hiring a go-to-market engineer. That was true the morning we recorded.

2. 25%的赞助商正在招聘上市工程师(go-to-market engineer)。我们在录制当天早上确认了这一情况。

3. CEOs own AI, then GTM leaders, then COOs. That’s 10K’s ranking of who is most likely to own AI at these companies.

3. CEO掌握AI主导权,其次是GTM负责人,然后是COO。这是10K对这些公司中最有可能负责AI的人员排名的结果。

4. 49.7% of executives changed jobs in the last 16 months, and 63% of CMOs did. We’ve always known our event buyers, who are usually CMOs, tend to be gone by the next Annual. We’d never measured it. I’d plan on your CMO not being there in two years and on half your executives turning over.

4. 在过去16个月中,49.7%的高管更换了工作,CMO中这一比例高达63%。我们一直知道我们的活动买家(通常是CMO)往往到下一届年度大会时已经离职。但我们从未对此进行过量化测量。我预计两年后你的CMO将不在场,且半数高管会发生更替。

5. More than 75% of AI-native attendees had never been to SaaStr before. I’d have guessed 50/50, with half being pre-AI B2B execs who moved over.

5. 超过75%的AI原生参会者此前从未参加过SaaStr。我原本猜测会是五五开,即一半是之前从事B2B业务后转向AI领域的高管。

The news the week we recorded was similar. The CEO of MongoDB left a $53 million pay package after 11 months for an enterprise leadership role at Meta. Snowflake’s first sales leader left for Cognition, which just crossed $1 billion in ARR.

我们录制新闻的那一周,情况类似。MongoDB 的首席执行官在任职11个月后离职,放弃了5300万美元的薪酬包,转而加入 Meta 担任企业领导职务。Snowflake 的首位销售负责人也离开了公司,加入了 Cognition,后者的年度经常性收入(ARR)刚刚突破10亿美元。

Jack Altman from Benchmark argued the upside when he joined us on 20VC. If executives are this mobile, you can recruit people you couldn’t have reached two years ago. Amelia hears it from sponsors too. The number one thing they ask her for help with, outside of SaaStr, is hiring.

来自 Benchmark 的 Jack Altman 在参加 20VC 节目时曾指出其中的机遇:如果高管流动性如此之高,你就能招募到两年前根本无法接触到的人才。Amelia 也从赞助商那里听到了同样的声音。除了 SaaStr 之外,他们请求她帮忙最多的事情就是招聘。

What We Changed This Week

本周我们的变更

  • 10K now tracks its own API calls. We had never looked.
  • We’re cutting the calls 10K flagged as unnecessary.
  • The Postgres mirror is on the table for any vendor whose agent pricing lands near the $240,000 estimate.
  • 10K 现在会自行追踪其 API 调用次数。此前我们从未进行过此类监控。
  • 我们将削减 10K 标记为不必要的调用。
  • 对于代理定价接近24万美元估算值的任何供应商,Postgres 镜像方案都在考虑范围内。

更进一步:量化金融体系

看懂新闻只是起点——沿量化金融路径,把它变成能交付的工程能力

进入量化体系 →

相似阅读

关联信息,但可能不是同一事件