Transcript 0:00 On Monday, I told you that Ox Alpha had turned all of us into an evaluation team for an anonymous AI lab. Well, the lab is anonymous no more. ZAI has revealed that Ox Alpha was a preview of GLM 5.3 Flash, as we said. 0:16 Okay, hike eaters, exhale. We know who built it, but that was never the most interesting question. During that launch, I became much more curious about the growing competition between OpenCode and OpenRouter. 0:29 One is an open source coding agent loved by many, while the other is a marketplace and browsing layer for hundreds of models. Why are they beginning to look like competitors? 0:39 And if they are competing, why would OpenCode have any chance? Because Stripe just agreed to acquire OpenRouter in a deal reported to be somewhere around 8 billion. Who could come for OpenCode next? 0:52 And could anyone buy without damaging the reason developers chose it? [upbeat music] Attention Span. Welcome to Attention Span by Turing Post. I'm Ksenia, and today we're following the Ox Alpha traffic one layer up. 1:08 The model mystery is over, but the fight over who gets to choose and serve the model is only beginning. To understand the rivalry, it helps to follow one collision request through the whole system. 1:21 First, there is a model lab, let's say ZAI, who builds GLM, or Anthropic, who builds Claude, or OpenAI, who builds ChatGPT. The lab develops the model, but it does not necessarily operate every server that runs it. 1:35 That creates a second business, the inference provider. These companies operate GPU capacity, load model weights, and sell access by the token. 1:45 The same open weight model may be available through several providers with different prices, speed, and reliability. Above them sits a gateway such as OpenRouter. 1:55 It gives developers one account and one API for many models and providers. 2:00 The developer usually chooses the model, while OpenRouter chooses which provider should serve it according to price, latency, availability, and data policy. 2:10 OpenRouter also offers an optional outer router that can choose the model, but only when the developer explicitly selects it. Finally, there is the agent harness, and OpenCode began here. 2:24 The harness gives the model access to repository, files, tools, a terminal, and tasks, then manages the loop while the model tries to complete the task. These layers often feel like one product. 2:37 A request can travel from OpenCode, through OpenRouter, to a provider running a ZAI model, so four companies may participate before even one line of code changes. 2:47 The competition begins because OpenCode is no longer stayed inside the harness layer. OpenCode still allows users to connect OpenRouter, call a model company directly, or run a local model. 3:00 But it now also sells model access through services of its own, then provides pay-as-you-go access to model and provider combinations that OpenCode says it has tested for coding. 3:12 OpenCode says it sells this inference of cost with a separate payment processing fee. Go is a $10 monthly subscription. For some models, the included limits are valued at up to $60 per month at public API prices. 3:26 OpenCode says it can offer this because it receives bulk discounts and reserves capacity in advance. Both products work through standard API endpoints, which means they can be used by other compatible agents. 3:40 So OpenCode is not only selling inference inside OpenCode, it is beginning to sell inference to the wider agent market. OpenRouter also says it passes through provider prices without adding an inference markup. 3:54 It charges 5.5% when customers purchase credits. In return, it offers a much larger catalog, provider focus, routing controls, and policies for teams. 4:05 This overlap is still small, but OpenCode has enough traffic to make it interesting. 4:12 Its site claims more than 16 million monthly developers, and CEO JV said in July that OpenCode was processing around 7 trillion tokens per day. 4:22 OpenRouter reported more than 10 trillion per day around the Stripe deal, so the difference is not huge. These figures were collected differently, so it's hard, and we don't know if we can treat them as a ranking. 4:34 But what they do tell us is that OpenCode has enough concentrated demand to negotiate directly with model labs and infrastructure suppliers. And this is the first important part of this story. 4:47 OpenCode began as the place where developers used models. It is now becoming one of the companies that buys, packages, and resells access to them. And it was interesting to read Dax Rodat's posts this week. 5:01 He's a co-founder of OpenCode and one of its main builders, so what he says, with plenty of teasing along the way, can reveal where the company is heading. 5:10 Dax argued that some inference companies rent GPUs, serve tokens, and add margins above 40%. He suggested that the companies supplying their GPU capacity may eventually compete with them. 5:23 He then claimed that OpenCode is already the number one buyer on the planet for some models, followed by, "Watch what we do next." 5:32 When someone asked why OpenCode could not rent GPUs and serve the models directly, Dax said the company was working toward exactly that. 5:40 He estimated that the first batch would require roughly 40 million upfront, so this is most likely their next step. 5:47 OpenCode could reserve a large amount of GPU capacity, select the small number of models that work well for coding, and serve those models itself or with an infrastructure partner, then would sell access by usage, while Go would package the same capacity into a subscription. 6:04 Raising GPUs is only the beginning. Someone still has to deploy the models, improve serving efficiency, respond to hardware failures, and pay for the capacity when demand falls below expectations. 6:16 And OpenCode would have to take responsibility for most of those costs. 6:21 Its bet would be that coding traffic is now large and predictable enough for OpenCode to do this more cheaply than buying individual tokens from an inference provider. 6:31 OpenRouter usually sits above the inference providers. It receives a request, selects an available provider, and forwards the request there. 6:41 If more companies begin serving models directly, OpenRouter may gain more providers to choose from. 6:46 The problem for OpenRouter begins when a customer as large as OpenCode signs a direct capacity agreement and the request no longer passes through OpenRouter. We can now describe the competition more accurately. 7:00 OpenRouter aggregates demand horizontally. It receives requests from many applications and many kinds of work, then routes them across a large market of models and providers. OpenCode aggregates demand vertically. 7:13 It controls one category, coding, but that category is large enough to support its own model selection, pricing, and infrastructure agreements. Open Router's advantage is breadth. 7:26 If a new model or provider appears, Open Router can add it without rebuilding the applications connected to its API. It can also move requests when a provider becomes slow, expensive, or unavailable. 7:38 OpenCode does not need that much breadth. It only needs a smaller group of models that cover the coding tasks its users perform most often. 7:48 And if coding traffic concentrates around five or 10 models, controlling the price and availability of those models may be enough. 7:56 OpenCode can negotiate directly, place them inside Zen and Go, and recommend them during setup. This is why the default is important. Most developers will not compare 20 providers every time they fix a bug. 8:10 They will choose a service once and continue using it until something gives them a reason to switch. And Open Router wants that service to be Open Router. OpenCode increasingly tries to move it to Zen and Go. 8:25 That brings us back to Ox Alpha, GLM-5 Country Flash was free during the Anaios preview. 8:32 OpenCode's public page eventually attributed around 45 trillion tokens, 500,000 users, and more than 13 million completed sessions to the model, although those totals continued growing after their year. 8:46 ZAI supplied the model and subsidized the inference. OpenCode and Open Router supplied distribution. Developers supplied millions of real coding sessions. 8:56 This interesting lineup suggests another role for companies such as OpenCode. They can become launch partners for model labs. A lab can offer discounted or free capacity. 9:06 OpenCode can place the model inside the coding workflow, and if developers respond well, the lab gains adoption and OpenCode gains better commercial terms. 9:16 The company controlling the workflow does not need to build the model to influence which model succeeds. This is much broader shift that I want to point out. 9:26 Until now, we have often treated AI applications as customers of model labs, inference providers, or gateways. They buy intelligence and build a product around it. 9:36 OpenCode suggests that a sufficiently large application can move in the opposite direction. 9:42 It can combine the demand of its users, negotiate directly with suppliers, and eventually operate part of the infrastructure itself. 9:50 Maybe even a legal agent could do this for legal models, or like a research platform could do it for scientific workloads. It all depends on the size of the audience. 9:59 And we may not end up with one universal marketplace where every application buys every model. Large applications may create smaller model markets inside their own workflows. 10:11 Now, to the question if anyone is going to buy OpenCode. It's very appealing. But buying OpenCode would not mean buying its code. 10:20 The appealing part is the reputation, but that also part of the problem, because it's open source and the code is MIT licensed and available to everyone. 10:30 A buyer would actually be purchasing Anomaly, the team behind OpenCode, along with the brand Zen, Go, and the official product used by millions of developers. 10:39 The company has already rejected several offers, according to Dax. But let's think who might be the most logical buyer. 10:48 The most logical buyer is actually two companies with the recent huge acquisitions, Stripe with Open Router and NVIDIA with Hugging Face. 10:56 Stripe and Open Router would gain the coding workplace they currently route traffic into but do not control. OpenCode would connect the agent model gateway metering and payment system. 11:07 NVIDIA and Hugging Face would gain something different, direct demand. Hugging Face provides models. NVIDIA provides infrastructure. And OpenCode would provide millions of coding sessions consuming both. 11:21 NVIDIA is so much about open source recently, and OpenCode feels just like a right piece for their puzzle. Of course, we can consider model labs such as OpenAI, xAI, DeepSeek, or whoever. 11:35 They could also potentially buy the distribution, but that would immediately weaken OpenCode's neutrality, and every recommendation would begin to look like a house recommendation. 11:45 Cloudflare is another plausible fit because it already has Workers AI and AI Gateway but lacks a major coding environment of its own. This is the problem with every potential buyer. 11:56 OpenCode is valuable because it can influence model choice without appearing captured by one supplier. 12:02 A company, another company, can buy Anomaly, but it cannot be certain that the community and its trust will come with it. So who gets to choose the model? The question that we asked in the beginning. 12:14 It might be the wrong question because at some point the choice may no longer feel like much of a choice. 12:20 Couple of years ago, I said that we were all becoming AI curators, choosing the models and tools we trusted, choosing the people and companies we trust. But OpenCode points to the next stage. 12:32 Curation becomes procurement. If OpenCode reserves capacity for a small group of models, makes them cheaper through Go, and places them prominently during setup, most developers will just go with it. 12:45 They will not run their own evaluation. They will choose from the options OpenCode has already tested, negotiated, and priced. 12:51 And the recommendation we see inside the agent may therefore be shaped months earlier during negotiations over GPU capacity. What looks like a model choice may begin as an infrastructure commitment. 13:03 Open Router offers a broad market of models and providers. 13:07 OpenCode may create a much narrower market around coding where the important competition happens earlier, when OpenCode decides which models receive testing, reserve capacity, favorable pricing, and prominent placement. 13:21 And by the time the developer opens the menu, much of the choosing may already have been done. And for some companies, it might be a slippery slope. Thank you for watching. I wonder what do you think about this shift. 13:33 Leave me your comments, and I'll see you next week. [outro music]