OpenRouter：从种子到 Stripe——多模型路由与十万亿代币经济

Anjney

我觉得有一种新的价值单位正在互联网上流动，叫作 token。未来十年，整条互联网价值链都得面对：token 越值钱，坏人就越想把手伸进去。

There’s a new type of unit of value that’s being streamed across the internet called a token. Over the next ten years, the entire internet value chain is going to have to deal with the fact that the more valuable tokens got, the more bad actors are going to try to get their hands on those tokens.

Anjney

在线支付从八九十年代走到超过万亿规模，需要整套新的反欺诈基础设施。代币经济这边，我们今天大致还在那个起点；未来五年大概到约 5 万亿美元量级，十年内到 10 万亿我不会意外。

Online payments started roughly in the eighties and nineties and grew to over a trillion dollars—we needed entirely new payment solutions for online fraud. Where we are today on tokens is roughly there; over the next five years we’re expecting the token economy to get to roughly 5 trillion dollars, and over the next ten years I’d be shocked if we weren’t at 10 trillion dollars of token flow.

Anjney

而且 Alex 描述的那些坏事，未来十年会越来越多由 AI 智能体来干——不只是坏人，还有攻击代币流的坏智能体。没有护盾，10 万亿 GMV 里欺诈占比可能高得吓人，信任也可能撑不住。

All the bad things Alex described as being perpetuated by humans right now are going to be perpetuated by AI agents over the next ten years. It’s not just bad human beings—it’s all the bad agents that are going to be attacking the token flow. Without that shield, a huge percentage of this 10 trillion dollars in GMV could be fraud and abuse—and we might never get there if people just don’t trust tokens.

Swyx

我们在 Anjney 家里——旧金山很多大创业故事都从这儿开始。恭喜你在 Cursor、Mistral 等等上面的布局；手上事真多。

Okay, we are here in Anjney’s house, which is where all big startups in San Francisco start. Congrats on Cursor, Mistral—God knows what else. You got so much stuff going on.

Anjney

确实很多。最近我最兴奋的之一，大概就是 OpenRouter。

There’s a lot going on. Well, OpenRouter has been one I’m most excited about recently.

Swyx

还有 Alex，第一次上这个播客。恭喜这一路。我翻你早年产品笔记，有一条「sub 作为产品原则」——你怎么想「世界上该存在什么」？

And we have Alex, first time on the pod. Congrats—what a journey. Looking back at your past posts, one of the earliest product principles I saw was “sub” as a product principle. Explain how you think about what should exist in the world.

Alex

「Sub」那篇大概是 2023 年初。把产品想成「订阅数据」和「发布数据」的交汇；市场是现成例子——供给方把商品发布到某个 SKU，SKU 像订阅主题，消费方按需取用。人类消费是离散、临时的；智能体和推理消费者却是持续消费、还不断换 SKU。OpenRouter 介于普通 API 体验和市场之间：我们造 model slug、有 auto router，各种可订阅的产品 SKU，让消费方持续拿价值、做决策。

The “sub” piece, early 2023, is about products as an intersection between subscribing to data and publishing data. Marketplaces are an easy example—suppliers publish a product to a SKU; the SKU is like a sub topic a consumer subscribes to. Humans consume in a discrete, ad hoc way. Agents and consumers of inference consume continuously and change the SKUs they consume from all the time. OpenRouter blends a normal API experience and a marketplace—we create model slugs, have the auto router, all kinds of product SKUs you can subscribe to, and continuously derive value and make decisions based on those consumers.

Swyx

现在大家默认要换模型、不会死磕原生 SDK；你们起步时远非共识。各自的「顿悟时刻」是什么？Anjney 你在 AI Engineer 讲过 Alpaca。

This is more consensus now, but not when you started—demand for swapping models, not only native SDKs. For each of you, what was the realization moment? You’ve given a talk at AIE about Alpaca as one of your inspiring moments.

Alex

2022 年底城里几乎只有 OpenAI，再加上 Cohere，以及零星开源权重尝试。2023 年 1 月 Llama 出来时很兴奋——某些基准能压过 GPT-3——但你没法真正「聊」；缺的是一点 RLHF。Alpaca 是我看到的第一个补上这一步的：斯坦福团队用合成数据微调 Llama，成本大约 600 美元，大概 7B 或 13B。我在飞机上用，很多时候和 ChatGPT 几乎分不清。若做模型这么便宜，就有全新数据变现方式——把高价值数据压进模型当服务卖，成本还会再降；也给了小团队追随前沿实验室的路径。一旦有爆款应用和可模仿的框架，生态该立刻长出来——于是需要一个市场，当 LLM 的「大本营」：谁在用、用多少、为什么。

End of 2022, OpenAI was the only game in town—OpenAI, Cohere, and a smattering of early open-weight attempts. When Llama came out in January 2023 it was really big; it outperformed GPT-3 on one or two benchmarks, but you couldn’t chat with it. Alpaca was the first that fixed that with RLHF—Stanford generated synthetic data, tuned Llama for about $600; maybe 7B or 13B. On an airplane I often couldn’t discern ChatGPT versus Alpaca. If it was that easy, we had a new way of monetizing data—turn valuable data into a service for $600, and that cost would go down. It also let a single developer follow frontier labs. Whenever you have a breakout app and a framework to imitate it, an ecosystem should arise—so you need a marketplace as a home base for LLMs: usage, who, and why.

Swyx

最接近的是 Hugging Face？

The closest would be Hugging Face?

Alex

当时是。但 HF 没有闭源模型，也谈不上「直接用」和「谁在用」的数据。OpenRouter 和 HF 的差异对我很关键——尤其当我在学人们为什么选不断冒出的小模型。

Hugging Face was the closest, yeah. But Hugging Face didn’t have the closed-source models, you couldn’t really use the models the same way, and there wasn’t data about who was using them. Those differences felt critical while I was learning why people chose the different little models emerging over time.

Swyx

Anjney，你并不陌生「要更多模型多样性」——当时你在 Anthropic 相关旅程里也有几年。你和 Alex 怎么认识的？

Anjney, no stranger to wanting more model diversity—you were a couple of years into your Anthropic journey. What was your introduction to Alex?

Anjney

认识大概要再往前推十三年；OpenRouter 的「握手」就在这儿发生。我们大二左右在 Stanford Review——Peter Thiel 当年创办的、偏自由意志主义校园报——第一次碰面。当时没有专门科技版，正值网络中立法案争论；我想开科技版，Alex 是少数说「酷啊」的人。大概 2011 或 2012，在 Old Union。后来职业交集最大的是我在 Discord 管平台、疫情期间加密与 NFT 爆发，Alex 在 OpenSea。钓鱼、社工、Katana 式 DDoS 都砸在平台上；我开始在斯坦福教 Security at Scale（CS 153）。高峰期 Discord 上跑着数十亿美元量级的 NFT GMV，几乎都来自 OpenSea——「DAO 的 D 就是 Discord」。

The introduction was maybe thirteen years before that; the OpenRouter handshake happened right over there. Alex and I met as sophomores at the Stanford Review—the libertarian newspaper Peter Thiel started. There was no dedicated technology section; net neutrality was being debated. I proposed a tech section; Alex was one of the only people who said yes. 2011 or 2012, Old Union. The most professional overlap was when I ran the platform at Discord during the crypto/NFT boom and Alex was at OpenSea—phishing, social engineering, Katana-style DDoS. I started teaching Security at Scale at Stanford, CS 153. At peak, several billion dollars in NFT GMV ran through Discord, all from OpenSea. The D in DAO is Discord.

Anjney

大约一年后 OpenAI 给 Discord 早期接入 GPT——是带 RL 的那版（GPT-3.5 系）。我们做了内部 Discord bot。两个用例：一是 Clyde，帮你搭服、拉朋友；二是内容审核。审核会直接拒答——护栏太粗。我们说需要权重、要在约 2.5 亿月活上可靠；对方说我们是闭源公司。那是我第一次意识到企业需要开源模型、能力可控，以及最终需要控制面来编排开源模型。直到大约半年后 Llama，再半年我领投 Mistral A 轮——Guillaume 和 Llama 团队那拨。差不多那时听说 Alex 在做 OpenRouter：这两个世界要撞上。

A year later OpenAI gave Discord early access to GPT—the RL version of three. We made an internal Discord bot. Two use cases: Clyde, a party friend for server setup, and content moderation. Moderation would refuse our prompts—guardrails. We said we needed the weights for reliability at 250 million MAU; they said they’re closed-source. That was my first realization enterprises would need open models, more control, and ultimately a control plane. No good open alternatives until Llama about six months later; six months after that I led the Series A into Mistral. Around then I heard Alex launching OpenRouter—these worlds were going to collide.

Anjney

Discord 每个服像微型部署，规则由社区自订；全球外包审核团队五千人以上。理想是把服规喂给 LLM 做定制审核——但很多服规直接撞 OpenAI 政策。哈利·波特同人服一提商标内容就拒；侦探小说暴力章节也会拒。精度根本不够。

Every Discord server is like a mini deployment with user-created rules. We had a 5,000+ person global content-moderation contractor team. The idea was to give server norms to the LLM for custom moderation—but many norms violated OpenAI’s rules. A Harry Potter fan community would trigger trademark refusals; detective stories with violence would refuse. Not precise enough.

Alex

还有写侦探故事、某一章很暴力时模型直接拒帮。那不是 LLM 结构宿命——一定有别处可选。拒答或烂结果时能换模型，这种张力也推我去做市场。

Another one: someone writing a detective story with a violent chapter—the LLM would refuse to help. That’s not structurally inherent to LLMs; there must be a choice so I can switch when I get a refusal or a bad result. That tension also drove me toward a marketplace.

Swyx

早期融资时别人吃不吃这套？最大阻力是什么？

When you were raising or starting, did people get it? What were some of the struggles?

Alex

最大反对意见是「大模型必胜」——缩放律加网络效应会归到一家，像 Google 搜素垄断，你们只能捡渣。有趣的是 Google 搜素赢面极大；但 LLM 不只是 UI，还是建新业务的方式。若真出现 Google 级垄断，量级会像东印度公司再乘上离谱倍数，因为整个经济都挂在一家上。而且做出好竞品的经济学更容易去中心化，所以那种结局既可怕也不太像。

The biggest objection was “big model wins”—scaling laws and network effects accrue to one company, a Google-style monopoly, and you’d fight for scraps. Google won search by a huge margin. But LLMs aren’t merely a UI; they’re ways of building entirely new businesses. A Google-level monopoly would be like the Dutch East India Company times a quadrillion because the whole economy depends on one monopoly. It didn’t seem like a sane outcome—and it’s less likely because the economics of creating good competitors are much more decentralizable.

Anjney

Alex 说的都对；我从另一面看。缩放律在我这儿一直是 OpenRouter 值钱的「特性」不是「缺陷」。我是 Anthropic 最早一批投资人之一，研究生读机器学习，圈子里都信 Bitter Lesson——OpenAI 和 Anthropic 已经是算力缩放的两个样板。决定一起做 OpenRouter 时，我已投了 Mistral、Black Forest Labs、Luma，跨模态都有。叙事说「只会剩一家像 Google」——也许，但我不信；而且多个研究团队在狂创新。

Everything Alex said is true, and I came at it from a completely different perspective. Scaling laws were always a feature, not a bug, for why OpenRouter would be valuable. I was one of the first investors in Anthropic; I went to grad school for ML; the bitter lesson holds—OpenAI and Anthropic were two proof points. By the time we teamed up on OpenRouter I’d already invested in Mistral, Black Forest Labs, and Luma across modalities. The “only one company dominates like Google” narrative—maybe, but I don’t believe that, and there was extraordinary innovation across several research teams.

Anjney

共享痛点是：研究团队极擅长想新能力，却不是开发者心态。checkpoint 出来后怎么办？早期 Anthropic、BFL、Mistral 的默认路径惊人地像——训完挂个 API，然后……鸦雀无声。Claude 第一个 checkpoint 内部做好约一年才对外；ChatGPT 出来后才觉得该放。Claude 1 博客大概三个开发者示例：一个 Discord bot、一个是我太太 Vivian 的 Juny Learning，还有 Notion——全是朋友圈，说明分发几乎是临时起意。密钥、开通、端点、版本——科学家觉得那是水管细节。

The shared problem: research teams are fantastic at new capabilities but not developer-mindset-oriented. What happens after the checkpoint? Early Anthropic, BFL, Mistral were similar—checkpoint done, put it out as an API, then crickets. The first Claude checkpoint was done a year before external release; after ChatGPT they decided to ship. The Claude 1 blog had like three developer examples—a Discord bot, Vivian’s startup Juny Learning, and Notion—friends of Anthropic; last-minute planning. Key management, provisioning, endpoints, versioning—scientists go “that’s plumbing.”

Anjney

我投的实验室有时砸进以十亿计的训练，early access 还是安静——因为光有权重做不出东西，需要一堆 plumbing。若要有人能跟 Google 竞争，就需要 OpenRouter 这种分发平台：DeepMind 训完一键推到 Docs、Android 上亿设备；在此之前，模型实验室得自己扛，非常吓人。Anthropic 到第一个 1000 万美元收入好像花了不止十二个月；对比 BFL，OpenRouter 可以说「上线当天我们能给你送去一百万开发者」——一小时量级的台阶。今天 OpenRouter 上开发者超过一千万（去重很难）。哪怕一天一千人试用反馈，对实验室也往往已经超过他们自己摸得到的量。Mistral 第一版权重甚至是磁力链接种子，没有 API——他们不是基建人。

Labs I funded would sometimes spend billions training, then crickets in early access—it’s hard to use a checkpoint; you need plumbing. A distribution platform like OpenRouter was critical if we wanted competition to Google. DeepMind finishes a checkpoint and blasts it across Docs to a billion Android devices overnight. Until OpenRouter, labs had to think about all that themselves. At Anthropic it took more than twelve months to first $10M revenue. With BFL, OpenRouter could say “day you launch, we can send 1 million developers”—a step function in an hour. Today it’s over 10 million developers on OpenRouter, hard to dedupe. Even 1,000 developers on day one is more than many labs knew how to reach. Mistral’s first checkpoint was torrent weights—a magnet link, no API. They weren’t infra people.

Alex

除了开发者体验，我们给模型做的「营销」和实验室自卖完全不同、观感也不同。市场像一间大黑屋，角落全黑，用户摸索着抓桌上的东西拼公司。模型不是能把特性列在网页上的产品——包括开源权重也是黑盒。需要有人打灯照亮角落，而且最好是中立第三方；路由与发现对供给方、实验室、托管方都越来越关键。

In addition to developer experience, the marketing we do for different models is totally different from what a model lab does for itself. We’re a neutral layer looking at the market like a big dark room with obscure corners; users walk in feeling around for objects to build companies. Models aren’t products where you enumerate features on a webpage—they’re black boxes, including open weights. You need to shine lights so people see what makes a model good, and the company shining that light should be a neutral third party. Routing and discovering models becomes critical for providers, labs, and servers.

Anjney

最大挫败感之一：很多 VC 没有一线操作经验。我到 a16z 大约一年前刚跑完 Discord 平台，知道把模型做成能用的软件有多难。有人不点名地说 OpenRouter「只是个市场」「别人 API 上的薄包装」。你根本不懂：哪怕只把三个 API 在生产里编排好，工程与社区设计的工作量都不会「默认发生」。Alex 从 OpenSea 时代就懂飞轮——系统思维，而不只是训练流水线。那时机器学习还多是笔记本下载数据、画 loss；部署、收集轨迹、闭环是后来的事。投资 OpenRouter 时我没费劲教育别人，直接投。大约一个月后 Matt Murphy（Menlo）好像给标了约 10×——具体数字我记不清——他们意识到战略价值。把一切贬成 wrapper 是最蠢、最还原论的框架，说明没在规模上交过产品。

One of my biggest frustrations is VCs without operating experience. I’d arrived at a16z about a year after running Discord’s platform. Investors looking at OpenRouter called it “just a marketplace,” “just a wrapper on other people’s APIs.” You have no idea how strategic orchestrating even three APIs in production is—the engineering and community design don’t happen by default. Alex understood flywheels from OpenSea—systems thinking most ML people don’t. Grad-school ML was download a dataset, look at loss curves—“great, I made AI.” Deploy, collect feedback, continuous loops came later. In the OpenRouter investment I didn’t educate other VCs—I just invested. About a month later Matt Murphy marked it up by ~10×; Menlo realized the strategic value. “Everything is a wrapper” is a stupid reductive framework from someone who’s never deployed product at scale.

Swyx

人人都是别人的 wrapper；有些 wrapper 很值钱。2023 年我提 AI Engineer 时最大反驳就是「没价值，你该去训模型」。

Everyone’s a wrapper on everything—and some wrappers have value. When I coined AI engineer in 2023, the number-one pushback was there’s no value; you should just train models.

Anjney

模型上线当天 OpenRouter 端点经常冲上 Hacker News 首页——外人不知道有多难，而这事一次又一次发生。

The day a model launches, an OpenRouter endpoint for that model frequently tops Hacker News on day one. People don’t realize how hard that is—and it happened over and over.

Swyx

今天你们有规模和信誉；早期飞轮怎么转起来的？

Today you have scale and reputation. When you were early, what was the bootstrap like?

Alex

拉回 Discord 早期——其实是 OpenSea 故事。我们因为 Axie Infinity 服连上：当时 Discord 最大服之一，你们不断抬人数上限。要给项目做市场，得先确保社区真想要。一对一推销有效，但在社区里大家同时跟你说话杠杆高得多。用户不断分享 OpenSea 链接，是「有事发生」的强信号。早期 LLM 的痛是 OpenAI 拒答或不完成任务，以及没法定制——被卡死的社区最值得扎进去学。

Early Discord days—technically an OpenSea story. We connected over the Axie Infinity server, one of Discord’s biggest; you were constantly bumping limits. To create a marketplace we needed the community to want it. One-to-one works, but a community where everyone talks to you at once is higher leverage. Users sharing OpenSea links constantly was a clear indicator. In early LLM days the gap was OpenAI refusing to finish prompts or complete tasks, and inability to customize—those blocked communities were the most useful to dive into.

Anjney

OpenSea × Discord 集成冲刺时，七个人都觉得「外链跳到 OpenSea」没问题；Alex 一个人说 UX 不行——应该嵌进 Discord iframe 里直接结账。技术上跳转两边 PM 都打勾，但对 OpenSea「留人在 Discord」几乎是对抗性的；他把用户体验放前面。那种既懂开发者流又死磕 UX 的组合很少见——飞轮两面转起来很难停。这也解释了为什么 Stripe 要买 OpenRouter：懂 ML 社区、开发者体验和用户体验的人太稀缺。

On OpenSea–Discord integration Zoom sprints, everyone aligned on bouncing out to OpenSea. Alex said no—bad UX; embed an iframe and check out inside Discord. Technically it checked PM boxes on both sides, but keeping the user in Discord was adversarial to OpenSea—and he put UX first. Someone technical who also prioritizes best UX is rare; those two sides of the flywheel are hard to stop. Nobody should be surprised Stripe decided they had to buy OpenRouter—that combination of ML community, developer experience, and user experience is really rare.

Alex

OpenRouter 之前我想做「自带模型」实验——AI 版 MetaMask。当时正经走 API 的 AI 应用，大概和用 JavaScript 硬调模型的小游戏一样少；浏览器里由用户控的桌面托管应用调 LLM，一度像可能路径。我用 Plasmo 做了 Chrome 扩展 Window AI；Plasmo 作者开始给仓库提代码，那人是 Louis Vichy——后来的 OpenRouter 联合创始人。扩展让网页按用户配置去调模型，形态不对，但学到很多：必须是 API，要有发现与开发者体验——需要图表、例子、空间，让人和智能体都能探索。OpenRouter 就这么来了。

Before OpenRouter I wanted a bring-your-own-model experiment—a MetaMask analogy for AI. There were probably as many AI apps hitting an LLM via API as games doing it in JavaScript; web apps calling LLMs through a user-controlled desktop-managed browser layer seemed possible. I built Window AI on Plasmo. The Plasmo creator started contributing on GitHub—Louis Vichy, co-founder of OpenRouter. Users configured which model a page used; wrong form factor for LLMs, but the learning was: this has to be an API with real developer and discovery experience—visuals, graphs, examples; explore as a human and as an agent. That’s how OpenRouter came to be.

Anjney

被低估的一点：我们当时紧挨加密社区——事后看，加密像生成模型的彩排。我在 Discord 当平台负责人，大约 80% 注意力在加密 NFT，另有约 20% 和朋友 David Holz 周末打万智牌——他在做文本生图的 Discord bot，叫 Midjourney。Axie 见顶后 Midjourney 是增速最快的社区之一；我们故意把 Midjourney 服做成与模型交互的主场——单机 midjourney.com 留存很差：空白输入框像 DALL·E 2，打 cat/dog 就僵住；在服里能看别人提示词再变奏，参与爆炸。从 0 到约 1000 万月活比 Axie 顺滑；四选一那套也是 RLHF 反馈环。Tom Brown、David 和我也一起打牌——概念一直在这圈朋友里转。Discord 成了早期应用培养皿：Midjourney、内部 Claude bot、ElevenLabs TTS……在 OpenRouter 给世界一个公共店面之前，Discord 几乎是搭在加密基建上的培养皿店面。

Underappreciated: we were adjacent to crypto—crypto was a dress rehearsal for generative models. As Discord head of platform, ~80% attention was crypto NFT volume; ~20% with David Holz—Magic on weekends—he was building a text-to-image Discord bot called Midjourney. After Axie petered off, Midjourney was one of the fastest-growing communities. We made the Midjourney server the primary place to interact—midjourney.com alone had terrible retention: empty field like DALL·E 2, type cat or dog, paralysis. In Discord you saw others and riffed; engagement was off the charts. Zero to ~10M MAU was smoother than Axie; best-of-four is the RLHF loop. Tom Brown, David, and I played Magic—these ideas were in one friend group. Discord became a petri dish—Midjourney, Claude as a Discord bot, ElevenLabs TTS—before OpenRouter gave a public storefront.

Alex

OpenRouter 并不以 Discord 为主体验。图像要靠「一眼看见别人怎么用」；LLM 需要更多 UX、计费点、底座集成与治理——数据策略、团队权限——Discord 服撑不住。

OpenRouter isn’t Discord-centric the way Midjourney is. Visually seeing prompts helps images; for LLMs you need a lot more UX, charge points, base integration, governance for data policies and teams—more than a Discord server can provide.

Anjney

Midjourney 是终端应用，Discord 有约 2.5 亿月活消费者，当宿主说得通。它从上线到约 1 亿美元收入年化不到八个月。随后 Stable Diffusion——我们泡在 LAION Discord——忽然开发者能自建 Midjourney 了；此前 Midjourney 没同级 API。有开源替代闭源应用时，OpenRouter 这种「形状」的价值就爆炸：你不必自己搞定托管权重。

Midjourney was an end-user application; Discord’s ~250M monthly consumers made it a sensible host. Launch to ~$100M revenue run rate was less than eight months. Then Stable Diffusion—in the LAION Discord—suddenly a developer could create their own Midjourney; Midjourney didn’t have an API of that quality. When open model alternatives to closed apps appear, the shape of OpenRouter becomes extraordinary—any developer can show up without figuring out how to host the weights.

Swyx

收个尾：大约 NeurIPS 前后有「Mistral 价格战」——Mixtral 8×7B 出来，价格大概掉 80%。对我是托管竞争第一次认真出现。

Cap off the Mistral side—there was a “Mistral price war” around NeurIPS. They launched Mixtral 8×7B and price went down like 80%. First real competition to host Mistral.

Alex

十二月那阵。模型一出，马上有人认真说这是「世上最好的模型」——就我记忆，开源权重第一次被这么严肃地叫。当然有 hype，也有些场景像在压 GPT-4。推理供给一片乱；我们把它理顺，让托管商拼价格，开发者在一个地方拿到最好价。这是供给方市场第一次清楚证明能给终端开发者加价值。

December. We saw that model and immediately people said it was the best model in the world—first time an open-weights model was called that in real seriousness. It was hype, and there were examples outperforming GPT-4. The inference landscape was messy; we cleaned it up so providers competed on price and you got the best price in one spot. First clear example of a provider marketplace adding value to end developers.

Anjney

NeurIPS 午餐会我们见过 Guillaume；刚宣布投资。他用法式淡定说「模型还行，没那么好」——反差巨大。他也提到很多人觉得它强过 GPT-4，部分因为速度：MoE 做得极高效，在帕累托前沿上。LLM 有时更快会让你以为更聪明；严格 N=7 评估未必，但人类偏好会把「快」读成「聪明」。人类其实一直在路由：先问快模型，不够再手动升级——后来变成 OpenRouter 自动路由。

We met a few days after Mistral at a NeurIPS luncheon with Guillaume. He’d just launched; in typical French fashion—“it’s an okay model, not that good.” He also said part of why people thought it beat GPT-4 was speed—an MoE they’d made super efficient, on the Pareto frontier. Sometimes when LLMs are faster you think they’re smarter; on N-of-seven evals maybe not, but human preference felt smarter because it was so fast. Humans were the routing mechanism—ask the fast model first, upgrade manually—which eventually becomes OpenRouter as router.

Swyx

你也投了 Arena。Midjourney 那种 ABCD 反馈环你懂，为什么没做成 Arena？Arena 为何没做成 OpenRouter？他们还有基于 ELO 的路由项目没商业化。

You’re also an investor in Arena. Midjourney’s A/B/C/D feedback loop mattered—how come you didn’t build Arena, and Arena didn’t build OpenRouter? LM Arena had a router on ELOs they never commercialized.

Alex

Arena 比 OpenRouter 早（校园项目再公司化）。一家公司很难两头做：一边吃数据卖数据，另一边默认不能。OpenRouter 默认不训练、不看你的 prompt/completion（除非组织选择开启）；我们在数据政策与隐私上很保守。LM Arena 的商业模式更朝向实验室。

Arena started before OpenRouter—school project then company. It’s hard for one company to do both: one takes data and sells it; the other really can’t by default. When you set up OpenRouter there’s no training, no prompts aside from provider policy—OpenRouter can’t see your prompts or completions unless you opt in. We’re conservative on data policy, security, and privacy. LM Arena’s business model is oriented around the labs.

Anjney

我当过 Arena 前五个月左右的过渡/创始期 CEO，帮 Anastasios 和 Wei-Lin 从伯克利 spin out；也在 OpenRouter 之前投了它。外人老拿两家比，我觉得奇怪——使命完全不同。Arena 创办实体叫 AI Reliability Institute，是评估服务：让模型评估比「手指蘸口水」可靠；那是他们博士工作——纠正收集偏差的统计方法、style control 等。Arena 的最高期望客户是实验室研究员；OpenRouter 的最高期望客户是把研究结果做成上线应用的开发者。签 OpenRouter 条款前几周我还打电话想和 Arena 拼开放 prompt 数据集——因为两边数据根本不一样。三万英尺看相邻，路线图当时指向很不同的方向。

I was interim / first CEO of Arena for the first five months helping Anastasios and Wei-Lin spin out of Berkeley, and I invested before OpenRouter. Outside comparisons felt strange—missions completely different. The founding entity was the AI Reliability Institute—an eval service to make evaluation more reliable than finger-in-the-wind. That was their Berkeley PhD work—statistical methods for correcting eval estimates, style control. Arena’s highest-expectation customer was a researcher at a lab; OpenRouter’s was a developer who ships an application. Weeks before the OpenRouter term sheet I called about pooling an open prompt repo—so different we didn’t have API prompts of what developers want. At 30,000 feet adjacent; roadmaps pointed different directions.

Alex

即使在 AI 时代，聚焦仍被低估：不只产品更好，世界也知道你的牌子到底解决哪类问题——真痛的人会找「最在乎这件事」的品牌。

Even in the age of AI, focus is underrated and critical—not just because you get a better product, but because the world knows what your focus is. If this really matters, go with the brand that cares most about it.

Anjney

强调一下：Anthropic 早期并不轻松，起步大概落后 OpenAI 一百亿美元量级。种子备忘录从第一天就是「负责任地商业化 AI 结对编程」——排除当时很炫的图像/视频。今天你看到结果：大约五年内走到万亿美元量级公司叙事。最高期望客户是谁、如何持续超出预期，是 OpenRouter 和 Anthropic 能成的原因之一。主评估一直是代码、长程智能体编程。Claude Instant / Claude 2 对外营销也曾偏散文与长上下文，但内部主线仍是编程。

To underscore focus: early Anthropic wasn’t easy—starting ~$10B behind OpenAI. From day one the seed memo was responsibly commercialize an AI pair programmer—excluding shiny image/video. Today you see the results—a trillion-dollar company within five years. Highest-expectation customer focus is part of why OpenRouter and Anthropic succeeded. Main evals were always coding, long-horizon agentic programming. Claude Instant and Claude 2 marketing leaned prose and long context, but that was always the plan inside.

Alex

做过没上线的原型：一个很消费级的「调模型即服务」——丢两三个 YouTube 视频，抽字幕，微调成「像视频里的人说话」；最初用例甚至是 AI Steve Jobs（尽管不允许）。市场主线在涨，我们没大力外测。也有人劝我们做：推理都从你过，就地日志、蒸馏到更小更快模型。我们选择更深地聚焦中立市场，和愿意帮企业做定制的推理商合作，而不是自己再开一条生意。

We made prototypes we didn’t launch. One was tuning-as-a-service in a consumer form factor—give us a few YouTube videos, extract transcripts, tune a model to talk like the people; AI Steve Jobs was the initial use case even though it’s not allowed. The model marketplace was growing; we didn’t test widely. Others would pitch: store everything, tune a smaller cheaper model. We leaned into focus—partner with inference providers who want to help companies do that; there’s value in being a neutral marketplace.

Alex

2024 年初我们试过把多模型结果融合，原型叫 MOM——Mixture of Models：你选或我们选几个模型，最后再用「集合里最强」的模型融合，中间结果铺在看板式界面。很早期的 LLM council / agent swarm 味道。细节决定成败，会吸走焦点，技术也可能太早——当时融合结果有时并不比最好的那个单模强，因为第一名遥遥领先。后来前三、前四更接近，各有「神经多样性」；RL 扩大了各实验室的创意表面积。我们删光代码；2026 年初研究更有戏，我自己做架构方案时多模咨询再融合，再问各模「融合是否更好」——它们都说是。抽检不错，就开始正式做 fusion。

In early 2024 we thought it’d be interesting to fuse multiple models—prototype MOM, Mixture of Models: pick a couple or we’d pick, fuse at the end with the smartest of the set, intermediate results in a Kanban-looking product. Very early LLM council / agent-swarm idea. Devil’s in the details; it takes focus; tech can be early—the fused result was sometimes the same as the best model because #1 was so far ahead. Over time top three or four got closer, still neurodivergent; RL expanded creativity surface area. We deleted all the code. Early 2026 research looked promising; I ran a personal experiment—architecture plan to all models, fuse, ask if fused beat each individual—they all said yes. Spot check good; that’s how we built fusion.

Swyx

商业模式很早就定了：抽成。增长拐点？有人说大约每周 9% token 量？以前「一百兆 token」还觉得夸张，现在你们……

You decided the business model very early—you take a cut. Major milestones that inflect growth? Growing like 9% week on week in token volume? You had a State of AI with a hundred trillion tokens—cute, because now you’re doing…

Alex

token 量大约每周 9% 听着差不多。我们现在大约一天十兆——还更多。很多拐点围着模型发布。到 2024 年 5 月主线还偏散文，代码不够、应用建不起；Dream Tavern 曾是头部应用之一（创始人后来在 Cognition / Devin 做产品）。2024 年中 Claude 3.5 Sonnet 出来，代码能力跃迁，用量与用户猛增，账单开始吓人，人们找「更便宜但等价」的选项；随后开源权重侧又有强模型接棒——像实验室冲锋、账单休克、开源两三个月后补位的拉锯，发生过好几轮。我们也拆出编码智能体排行：Cline、Roo Code 等很吃这一套。2025 年底排行榜上多是 IDE / 终端智能体；然后 OpenClaw 出现——新形态带来非开发者（生产力、互联网创作者），架构上还有周期性 heartbeat 探活，心跳不该付贵模型的钱，于是 auto router 突然对很广用户有用，指数起飞。Hermes 等也吃透 auto router、技能与记忆。记忆 / 技能 / 沙箱我们本可做，但开发者当时更想自己架构这些层；抽象很难对所有人成立。排行榜随时间像一部 AI 产业电影——Wayback 看排名和应用榜就够。Karpathy 有句像成年礼：不怎么读 Local LLaMA 了，直接看 OpenRouter 排行榜。

~9% week-on-week token volume sounds about right. We’re doing ten a day—more. A lot is oriented around model launches. Through May 2024 focus was prose—coding wasn’t there; Dream Tavern was a top app—the creator now runs product at Cognition/Devin. Mid-2024 Claude 3.5 Sonnet leaped coding; volume and users surged; people looked at invoices and wanted cheaper equivalents; open weights delivered effective options a couple months later—several swing cycles. We broke out coding agents—Cline vs Roo Code loved that leaderboard. End of 2025, coding apps were IDEs or terminal agents; then OpenClaw—a new form factor for productivity/internet creators, plus heartbeat calls you don’t want to pay a lot for—so auto router became really useful and it rocketed. Hermes leaned into auto router, skills, memory. Memory/skills/sandboxes we could have done, but developers wanted to architect those; hard abstractions. The leaderboard over time is a movie of AI—Wayback the rankings. Coming-of-age: Andrej Karpathy said he no longer reads Local LLaMA; he goes to OpenRouter’s leaderboard.

Swyx

Stripe 这话题怎么起的？

How does the Stripe conversation start?

Alex

和 Stripe 长期合作很多项目；我们在反滥用、token 欺诈上砸很多力气。我发过：上个月拦下的美元金额大约是前一个月的 10 倍，手法也在分化——盗卡、违反 ToS 转售流量、被黑账号、公司被入侵自己还不知道、私下转卖推理、以及失控智能体把账单打爆（不是黑客）。Trust & Safety 用模型和流程去拦。做网关或「广义卖推理」就会成欺诈靶子；若卖很具体的智能任务、让用户自带推理，中招概率低得多。所以行业会从「推理加价转售」走向按任务/增强收费、BYOK 式自带推理。定价会像 Datadog 事件页那样连续、多样；消费级则更订阅、更简单。

We had a longstanding relationship with Stripe across projects. We invest a lot in countering abuse and token fraud. I posted: we blocked 10× as much dollar volume last month as the month before. Types diversify—stolen cards, reselling traffic against ToS, hacked accounts, fully compromised companies that don’t realize it, accounts reselling inference, accidental runaway agents. Trust and safety works those categories. If you’re a gateway or selling generalized inference, you’re a fraud target; discrete intelligence products with BYOK-style bring-your-own inference are less so. Companies will move from markup-on-inference toward tasks/enhancements. Datadog’s pricing page is a good look at the future—continuous event pricing; consumer apps simpler subscriptions.

Anjney

Alex 把我四年前因 Discord 就确信会在规模上爆炸的事讲清楚了。Midjourney 早期免费试用：大约十张图是「啊」的激活点。有天凌晨用户洪水，David 连打未接——他关了免费试用：有人按地理 IP 看，中国侧在把免费试用当订阅转售。连 Midjourney 这种专用应用都会中招。从那时起我清楚：token 是新的价值单位；越值钱攻击越多。Midjourney 好像再也没开回免费试用。这也是我开 Security at Scale 的原因之一。支付走过万亿；代币侧五年约 5 万亿、十年 10 万亿的量级在我看来不意外——而当时 Midjourney 年化还不到约 3 亿美元就已经打得很凶，说明需要全新系统。

Alex described something I knew at scale four years ago from Discord. Scaling Midjourney, free trial to first ~10 generations was the magic activation. One morning a flood of users—three missed calls from David—he shut the free trial: geolocation showed someone in China reselling Midjourney free subscriptions. Even a specialized app. Big picture: a new unit of value—tokens—and the more valuable, the more bad actors. I don’t think Midjourney’s turned free trial back on. That’s why I started teaching Security at Scale. Online payments grew past a trillion; on tokens, ~$5T in five years and ~$10T in ten wouldn’t shock me. Midjourney was under ~$300M revenue run rate and abuse was already aggressive—we’d need entirely new systems.

Anjney

董事会提到 Stripe 想合作时立刻通了。十年前在 Kleiner 投 Stripe 时，Patrick 和 John 的叙事是：不像传统工具先做完 KYC/AML，他们先吞欺诈成本当获客，五分钟五五行代码开收；数据攒起来后做出 Stripe Radar。人们当支付公司——其实更像安全公司；更便宜的通道很多，Stripe（以及欧洲的 Adyen）能守住，靠的是多年欺诈检测。价值大流量跨世界，就需要新防护：这是 Stripe × OpenRouter 对前沿 AI 生态的安全故事。

When you brought up Stripe wanting to partner in a board meeting, it made so much sense. At Kleiner ten years ago we invested in Stripe—Patrick and John: unlike Braintree-style day-one KYC/AML, bite fraud cost as CAC, five lines of code, accept payments in five minutes; over time data → Stripe Radar. People think payments; Stripe is really a security company. Cheaper transmission exists; Stripe and Adyen win on fraud detection. Every time value streams at scale you need new protection—the Stripe–OpenRouter story is a security story for the frontier AI ecosystem.

Anjney

第二点常被低估：现在由人干的坏事，未来十年会由 AI 智能体递归放大。实验室研究员只看自己训练的智能体如何失控，那只是冰山一角。需要新的治安官——戴牛仔帽那种——看见跨实验室、跨部署、跨开发者的坏行为，给整个代币经济做盾。否则 10 万亿里欺诈占比可能极高，信任崩了就到不了那儿。智能体欺诈像海啸，基建几乎还不存在。

Second: everything bad humans perpetrate now will be perpetrated by AI agents—recursive scale of bad actors. Researchers only see how agents they’re training go rogue—a fraction of bad behavior. You need new sheriffs in town with cowboy hats who see bad agent behavior across labs, deployments, and developers, and build a shield for the token economy. Without that, a huge percentage of $10T GMV is fraud/abuse—and we might never get there if people don’t trust tokens. Agentic fraud is about to hit like a tsunami; this infrastructure barely exists.

Swyx

最后留给你们：大家能期待 OpenRouter 和 Stripe 带来什么？

Last word—what can people expect from OpenRouter and Stripe?

Alex

这是加速打市场、更快上探企业的好方式；也像 Anjney 说的，信任安全与「接受 token / 让人自带推理 / 让开发者继续建在推理之上」有清晰的 better-together。OpenRouter 品牌、产品、路线图和名字都会留——接下来大约六个月，多数事仍是独立时会做的那些，只是整体更快。更长线的事希望很快能说。

This is a really good way for us to accelerate go-to-market and go upmarket more quickly. As Anjney described, there’s a clear better-together on trust and safety—making it easy to accept tokens, bring-your-own inference, and help developers build on inference. We have a strong brand and we’re keeping it—product, roadmap, name stay the same. Next six months, most things are what we would have done independently, except everything moves faster. Longer term, hopefully I can comment soon.

Swyx

希望以后能做续集。谢谢你们的时间，也恭喜——这是我在 AI 里见过最美好的兄弟情谊之一。

Hopefully a follow-up. Thank you for being so generous with your time, and congrats on the partnership—one of the most beautiful bromances I’ve seen in AI.

Anjney

才刚开始。从 Stanford 到现在，代币经济还需要很多新治安官。

Just starting out. From Stanford to here—lots more to do. Lots of sheriff work for the token economy. We need new sheriffs for sure.

Alex

谢谢。

Thank you.

Swyx

谢谢。

Awesome. Thank you.