RIP Golden Age of Agent Experimentation 2026-2026

Today on the AI Daily Brief: The big implications of Anthropic's new pricing plan. And before that in the headlines: The US AI envoy lands in Beijing. The AI Daily Brief is a daily podcast and video about the most important news and discussions in AI. All right, friends, quick announcements before we dive in. First of all, thank you to today's sponsors KPMG, Granola, Robots and Pencils, and ZenCoder. To get an ad-free version of the show, go to patreon.com/aiDailyBrief, or you can subscribe on Apple Podcasts. Ad-free is just $3 a month. If you want to learn more about sponsoring the show, send us a note at sponsors@aidailybrief.ai. 3 more quick things on the learning program side of the house. Agent OS, which is a platform-neutral agent operating system building program that is free, self-paced, self-directed, is live now, and there are more than 5,000 of you doing it, which is just awesome to see. Enterprise Claw, which is the supported enterprise-grade version of that sort of agent OS type program, is signing up folks for Cohort 3. And if you are interested in building the next generation of this podcast and all of these cool things around it, I have a job live for what I am calling a growth engineer. To be clear on the terminology, this does not mean you have to be an engineer by training, but you do have to be a Codex Cloud Code style builder who can think in growth terms. That is a full-time role. It is live now. And you can find links to all of this at aidailybrief.ai. Well, the AI envoy has touched down in Beijing, and Jensen Huang ended up making the flight. Earlier this week, the White House announced that over a dozen executives would be joining the president on his trip to China, with those executives representing the industries that would form the core of trade negotiations. Despite NVIDIA being a major bargaining chip in US-China relations, Huang was reportedly excluded from the invite list. White House sources said that after Huang's exclusion made headlines, the president personally reached out to extend a last-minute invite. Now, the full details of the trade agenda are still a little uncertain. Trump likened the deal to Nixon's summit with Chairman Mao in 1972, commenting, "I will be asking President Xi, a leader of extraordinary distinction, to open up China so that these brilliant people can work their magic and help bring the People's Republic to an even higher level." Meanwhile, the commentary is all swirling around a renegotiation on the status of Taiwan. As with anything in current global politics, this week could be a wild nothingburger, or it could have incredible ramifications, particularly for tech and AI, which of course we'll cover in more detail if and as there's anything to discuss. Next up, we turn to markets, where Cerebras has raised over $5 billion in the largest IPO so far this year. Interestingly, it apparently almost didn't happen. The IPO was priced on Wednesday after the close, with the final stock price set at $185. This was significantly above the advertised range of $150 to $160, even after Cerebras supersized the offering on Monday. The final price implies a market value of $40 billion. Now, CoreWeave was the last AI-related IPO, but the hype around Cerebras blew them out of the water. Bloomberg reports that Cerebras saw orders for 20 times more than the available shares. And as an additional twist, a last-minute acquisition bid almost derailed plans to go public. Sources told Bloomberg that chipmaker Arm and their majority owner SoftBank approached Cerebras for a takeover late last month. Cerebras reportedly rejected the bid, preferring to go it alone in the public market. And it's kind of hard to second-guess that decision now that they've raised what would have been one of the largest IPOs in any of the past few years. The stock was set to begin trading on Thursday morning after I record, so we'll check in and see how the market receives Cerebras on the Friday show. Moving over to the political side of the house, a new Gallup poll has found that voter sentiment is breaking hard against data centers. 7 in 10 Americans now oppose data center construction in their local area, including 48% who said they're strongly opposed. Just 7% said they were strongly in favor. The survey was taken in mid-March and is the first time Gallup has polled specifically on data center sentiment. Gallup designed their polling questions to mirror polling on nuclear power plants, which has been conducted since 2001. In their March survey, they found that more people oppose data centers than nuclear plants. In fact, the opposition is so intense that it exceeds the all-time high of nuclear opposition at 63%. The polling also dug into the reasoning behind the sentiment, with 46% of people said that they worry a great deal about the environmental impact of data centers, with an additional 24% stating they worry a fair amount about that. Half of opponents cited excessive use of resources, with 18% mentioning the use of water and electricity. 16% cited pollution, including noise pollution. 22% of opponents said that they were worried about the impact on the local quality of life, including increased traffic and land use concerns, and around another 20% said that they had economic concerns, citing higher utility bills, cost of living increases, and use of taxpayer funds in subsidizing construction and operation. Interestingly, only 12% said that they were worried about jobs or AI replacing human workers, And similarly, only 14% of opponents cited a generic negative view of AI, and only 13% cited AI safety as a reason to block data center construction. Now, on the opposite side of the ledger, by far the most popular reason to promote data center construction was local economic benefit. 66% gave this explanation for their support, with 55% explicitly mentioning job opportunities. No other reason came even remotely close. Interestingly, despite the broader partisan divide in this country, Gallup didn't really find all that much distinction between Democrats and Republicans on this issue. There was also no meaningful difference based on age, race, education level, or urbanicity. Now, what's super interesting to me is the extent to which the stated opposition has to do with environmental concerns as opposed to economic concerns. Understandably, many advocates are frustrated with the misinformation. Matt DeLuca writes, so 70% of Americans don't understand that data centers will improve schools and local services while not adding any traffic or negligible noise. Electricity increases and water complaints are all myths. Timothy B. Lee writes, "It's wild that we're having an environmental panic about data centers, a carbon-free industrial facility that during ordinary operation consumes no natural resources besides electricity and water and emits no waste products besides heat." He continues, "Almost every industrial facility requires electricity and water as inputs and produces waste heat. If you think this is bad, you aren't just opposed to AI, you're opposed to the industrial revolution." And yet, for all of the AI people complaining about this, there's no amount of being annoyed at environmental misinformation that's gonna change people's perspectives. I think Mads Campbell has it right when she wrote, "Data centers are fundamentally a marketing problem. Railroads used to be one too. Towns fought them over noise, land use, and fear of outsiders. Eventually people realized railroads meant commerce, jobs, and growth. Stadiums get backlash every single time they're proposed, but teams flip the narrative by turning them into entertainment districts with restaurants and nightlife. Data centers need the same shift. Right now the public narrative is just doom and gloom." Power consumption, water usage, giant buildings. People do not feel personally connected to the upside. There needs to be a massive shift in rhetoric from infrastructure people tolerate to infrastructure that benefits your community. In another tweet, she said, no one would care about data centers in their neighborhood if it came with free electricity and Wi-Fi. The second Americans benefit from the infrastructure, everyone becomes pro-data center instead of against it. I don't know how this could be any more obvious to people, and I think that both the AI companies who are funding this and the data center construction companies themselves should be embarrassed at how badly they're losing hearts and minds. I genuinely believe that you could turn around people's perspectives on this with a tiny fraction of the resources spent to actually construct the damn things in the first place. And importantly, that doesn't mean just running ads trying to correct misperceptions about environmental impact. It means legitimately providing clear and tangible benefits. Anyways, that's a rant deserving of a full episode, but we got a couple more stories before we get out of here. OpenAI is now supporting multiple pieces of AI regulation as part of a broader shift in policy positioning. Earlier this year, OpenAI laid out their policy views in a white paper called Industrial Policy for the Intelligence Age. The document was poorly received— speaking of rants, you might remember my episode about that one— although the big complaint was actually different than mine, with many claiming that OpenAI's lobbying efforts to stop AI regulation didn't align with their proposals. In a new interview with Axios, OpenAI's Chief Global Affairs Officer Chris LeHane discussed a rethink on how the company interfaces with policy and the American voter. LeHane acknowledged that AI companies will be crushed by public sentiment if they don't find a way to redistribute AI wealth. Similar to how Alaska shares oil and gas revenue with citizens. He said, "People need to feel like they're gonna have a piece of this and participate in it. You can't talk beyond people or above people. You need to talk with people and involve them in the conversation." Now, alongside the interview, OpenAI has thrown their weight behind regulatory efforts in Illinois. They are supporting both the Kids Online Safety Act and SB 315, a bill that mirrors the regulations in New York and California. It seems that OpenAI is now on board with state-by-state regulation as long as it doesn't result in a patchwork of different standards. Axios reporter Ashley Gould commented, "In: state consistency. Out: praying hopelessly for a federal standard." Lastly today, just a fascinating one that I think is a good reminder of how entrenched anti-AI feeling is becoming. Shloms on X posted, "I just generated an image in the style of a Monet painting using AI. Please describe in as much detail as possible what makes this inferior to a real Monet painting." Hundreds and hundreds of people were willing to do exactly that. In a follow-up tweet they wrote, "Please focus on the specific visual elements that distinguish the AI image from a real Monet painting like the below." So in each of these tweets, there was a Monet-style image. In the second it was a real Monet, and in the first it was— you guessed it— also a real Monet. The social experiment, then, was that Shloms posted a real Monet, a specific close-up crop of one, but said it was AI, and asked people to say why it sucked. And boy did he get responses. Hundreds and hundreds of responses. "It is Monet-ish, yes, and certainly pretty, but at the risk of sounding pretentious, it, in my opinion, would not hold up next to the real thing seen face to face in a museum." Another account writes, "There's a certain harshness, no soft blending of colors, no depth, no symbiosis of the elements." Another account, "It's all borked nonsense with no sense of space." Another account, "Monet shows more distinction in how he uses colors and outlines of objects." Another, "It's not a physical painting created by a well-known artist. That's it. That's the difference." et cetera, et cetera, et cetera. Indeed, the depth that people go to point out the specific ways in which it sucks compared to Amané is really fascinating. Now, the point in this case is not to dunk on people's artistic knowledge, not at all. It's to remind us as we're talking about things like regulations and like Gallup polling on data centers, that those of us who actually care about AI and the potential for it to have a positive impact on the future, have a hell of a lot of work to do. That's gonna do it, however, for the headlines. Next up, the main episode. One of the most important AI questions right now isn't who's using AI, it's who's using it well. KPMG and the University of Texas at Austin just analyzed 1.4 million real workplace AI interactions and found something surprising. The highest impact users aren't better prompt engineers, They treat AI like a reasoning partner. They frame problems, guide thinking, iterate, and push for better answers. And the good news? These behaviors are teachable at scale. If you're trying to move from AI access to real capability, KPMG's research on sophisticated AI collaboration is worth your time. Learn more at kpmg.com/us/sophisticated. That's kpmg.com/us/sophisticated. Today's episode is brought to you by Granola. Granola is the AI notepad for people in back-to-back meetings. You've probably heard people raving about Granola. It's just one of those products that people love to talk about. I myself have been using Granola for well over a year now, and honestly, it's one of the tools that changed the way I work. Granola takes meeting notes for you without any intrusive bots joining your calls. During or after the call, you can chat with your notes, ask Granola to pull out action items, help you negotiate, write a follow-up email, or even coach you using recipes, which are pre-made prompts. Once you try it on a first meeting, it's hard to go without. Head to granola.ai/aidaily and use code AI Daily. New users get 100% off for the first 3 months. Again, that's granola.ai/aidaily. Today's episode is sponsored by Bolt.new. Bolt.new is agentic engineering on multiplayer mode. Designers, product managers, and engineers build in the same environment, and the design system agent keeps every screen on brand. No more Frankenstein UIs stitched from a dozen prompts. Whether you're shipping internal tools, moving from prototype to production, or replacing a legacy admin panel, Bolt.new takes your team from concept to deployed app. One personal recommendation: hit plan mode before you build. I had a project I'd half described in 3 different prompts and plan mode made me actually think through it with Bolt.new before a single line got written. It saved me from rebuilding the same screen probably about 4 times. Build better apps faster. Start with the link in the description. Here's a harsh truth. Your company is probably spending thousands or millions of dollars on AI tools that are being massively underutilized. Half of companies have AI tools, but only 12% use them for business value. Most employees are still using AI to summarize meeting notes. If you're the one responsible for AI adoption at your company, you need Section. Section is a platform that helps you manage AI transformation across your entire organization. It coaches employees on real use cases, tracks who's using AI for business impact, and shows you exactly where AI is and isn't creating value. The result? You go from rolling out tools to driving measurable AI value. Your employees move from meeting summaries to solving actual business problems. And you can prove the ROI. Stop guessing if your AI investment is working. Check out Section at sectionai.com. That's S-E-C-T-I-O-N-A-I.com. Welcome back to the AI Daily Brief. Today we are discussing the most recent changes to how paid Claude plans work. This has caused a ton of angry discussion among developers, but I think that what it reflects is actually a much bigger shift. A couple of weeks ago, we published a show called The AI Subsidy Era Ends, and this story is really that coming home to roost. So first of all, let's talk about what happened. On Wednesday, Anthropic announced a big change to their pricing model. As of June 15th, paid users will be able to claim a monthly credit for what Anthropic calls programmatic usage. Basically, there are kind of two different ways that people use Claude and Claude Code. One is what Matt Pocock calls human in the loop, where you are sitting interacting with Claude via Claude AI, Claude Code, or Claude Cowork, the apps. And then there's what he calls the away from keyboard uses, which include the Claude agent SDK, Claude -p, which calls up Claude code from the CLI, or Claude GitHub Actions. Previously, people had been able to use their Claude subscriptions for either of these categories of uses, and their subscription limits such as they were didn't matter whether they were being used up in one area or the other, at least sort of. Throughout the year, there have been back and forths and questions on exactly what type of behavior Anthropic did and didn't want to support, with Matt himself jokingly posting at the beginning of April, "I don't know what the fuss is about. Anthropic's rules on using subscriptions are very simple. Claude code, okay. Claude's online platform, okay. Agent SDK running in personal software, okay-ish. Agent SDK running in commercial software, not okay. Claude code running in CI, not sure. Oh, maybe it's not so simple. Agent SDK running in CI, not sure. Claude-P running in CI, not sure, etc., etc." The point is that even before this, people were confused. It certainly seemed, however, ever since changes that seemed to impact how people could use their Open Claw subscriptions, that the team at Anthropic was clearly nudging you away from what Matt called those away from keyboard use cases and towards the human-in-the-loop use cases. Basically, it has seemed to many like Anthropic wants you to interact with their models through their own harnesses. Anthropic sort of acknowledged this confusion, writing in the X thread, we've heard your questions about SDK and Claude-P usage, sharing your subscription rate limits with Claude Code chat. They say then that this is a way to address that. Now, for what it's worth, what Matt is calling human-in-the-loop usage, Anthropic is calling interactive use, and what he is calling away from keyboard usage, they're calling programmatic use. Now, the way that the Claude Devs account tried to present this change is that it was actually just a bonus, that effectively nothing is changing about your subscription limits for how you use the normal interactive Claude accounts, i.e., when you're sitting there in front of Claude AI, Claude Code, or Claude Cowork. But that on top of that, you are now getting this bonus credit where you didn't before, with the bonus credit being equal to the amount that you're paying for your plan. So if you're paying $20 for a Pro account, you get a $20 credit to be used for that programmatic API-based usage. So Anthropic is basically trying to say that this is good in 3 ways. First of all, it's good because it clarifies things. Second of all, it's good because it means with that clarity that you are in fact allowed to use your Claude subscription for third-party tools. No more confusion about that. They write, this means that third-party tools built on the agent SDK like Conductor and OpenClaw will work with your Claude plan. And they say it's a good thing because you're getting this free credit. Lydia Hallie from Anthropic tried to visualize it, showing that Anthropic has just simply now cut things into two different categories of use with a separate accounting for when you start to hit overages and have to start paying the going API rate. So this is good, right? You're just getting free credits on top of your existing plan. Not so for many developers, and the issue at core is token subsidies. The TL;DR is that for a very long time, power users have been able to use dramatically more tokens than the equivalent value of API credits in their plans. What I mean by that is that the most powerful users of Claude Max subscriptions, for example, are in many cases not using $200 worth of tokens like the $200 a month plan would suggest, but are in fact using thousands of dollars worth of tokens if those tokens were charged for in the same way that the API charges. No one knows the exact ratio, but the fact that there is a subsidy is very clear. And the problem is that as it turns out, lots of people have built their third-party projects around an assumption of that subsidy persisting. Now what Anthropic is saying with this announcement is that that subsidy is persisting when it comes to people who are using Claude Code through the Claude Code or Claude Cowork harness, but it is not persisting if you're trying to use your Claude subscription for third-party experiences. Those experiences are now going to be billed straight on API rates. A big chunk of the developer community absolutely crashed out at this. AI YouTuber and T3 Code builder Theo writes, I can't help but feel personally burned by the Claude code changes announced today. We put so much work into wrapping the atrocious Claude agent SDK in T3 Code. It was the only path they supported, so we made it work. It was hell. Now our users are getting their rate limits cut by 40x despite us doing everything right. I listened to the Claude Code team. I had my issues with their direction, but I trusted them and took them at their word. I will never make that mistake again. Until we see significant change, it is safe to assume any statement from an Anthropic employee is a lie on a timer. The rug will be pulled no matter how many promises are made beforehand. That was one of about 100 tweets from Theo about this. If you use any of the following with your Claude sub, he writes, your usage just got cut by 25x. T3 Code, Conductor, Zed, John, Claude P in your CI scripts to call Claude code from other tools. They're disguising this as free credits. Don't fall for it. And honestly, relative to some of the other tweets out there, the O's were restrained. Robin Evers, who calls himself an AI coach for founders, said, Anthropic literally effing disgusts me. I said this before and I'll say it again, they hate developers. Their marketing and legal teams write the most intentionally misleading copy and rules. They sell dumpster fires as wins to the clueless and they're getting away with it. Never been happier to have canceled my subscription. Luca Lotto writes, "Really a shame that Anthropic intentionally harms their own developer community." And indeed this sentiment, that Anthropic was in fact abandoning the people that got them to where they are, was pretty prominent. Mario Zetschner writes, "I don't think Anthropic is dumb. Pretty sure they have data telling them that dev and social sentiment doesn't matter anymore, as they've reached corporate decision makers' hearts and minds. Enjoy your obligatory enterprise cloud code." In other words, the argument here is that they simply don't care about this class of user because they have bigger fish to fry with their enterprise clients. Mario also said, I do think initial growth was largely based on a focus on absolute nerds. Getting the nerds to get the normies, then dropping the nerds is a strategy as old as computers. A lot of the businesses affected are ones like T3 that Theo was talking about, which are effectively alternative harnesses for using Claude code. That's what T3 Code is, and that's also what Conductor is, a third-party coding harness focused on multi-agent workflows. Now, the argument implicit in a lot of these critiques is that these types of developers, even when they're building interfaces that compete with core Anthropic interfaces, are so valuable to the overall ecosystem by giving people different doors into that ecosystem that they should be actively supported and encouraged. Theo again writes, kind of crazy that Anthropic spends more time trying to lock out better apps and harnesses instead of just fixing flawed code. Now, on top of the specific critique in terms of how Anthropic runs its business and what they think about developers, The other big issue that people had with this was feeling gaslit and talked down to. Gauntlet founder Austin Allred retweeted Lydia Hallie's explanation of the change and said, "It's dramatically more expensive, but you get a few free credits towards the new expense to soften the blow. Your users aren't dumb. Stop trying to dress everything up in 10 layers of PR speak and just say what it is." Robin Evers again writes, "It's not the SDK change, but their pathetic attempt at gaslighting everyone into thinking this is a huge win for users." "Free API credits, yippee! Plus 50 weekly limits for 2 weeks. But if you don't use Cloud Code directly, you get banned or pay API pricing. This is not a win, it's a rug pull." Peter Yang took the volume of the discussion down a notch and kicked the thoughtfulness quotient up when he wrote, "I think one thing I learned about engaging your customers is you just have to be honest and upfront about your constraints, especially with developers. They really just want communication that tells it like it is. It's very hard to earn their trust, and once lost, it's very hard to earn it back." So that is all the critique and controversy. However, I will say that when I saw this announcement, I felt pretty similarly to developer Nick Dobos when he wrote, "Everyone with an ounce of sense knew this was coming. Did y'all seriously expect to run your insane custom scripts on top of Claude at the massively discounted subscription rate forever?" So let's talk about what the possible explanations for this are and which combination of them have it right. Let's talk first about the idea that Anthropic doesn't want third-party developers around. It is certainly the case that they have been more assertive about this than, for example, has OpenAI. Back in January, VentureBeat wrote an article called Anthropic cracks down on unauthorized Claude usage by third-party harnesses and rivals. Even 5 months ago then, identifying that Anthropic clearly had a bias towards getting people to use their harnesses, not their models inside other people's harnesses. Given how much we've talked about this year, The idea that harnesses, i.e., the products and experiences through which you use a model, are every bit as significant as a model when it comes to a company's success, I don't necessarily find this all that surprising. Also, at the beginning of the year, we saw Anthropic cut off xAI's access to its models for coding. Dario has made it clear that it seems crazy to him that they would just allow their competitors to use their models. And then of course there was everything around OpenClaw, where at no point has Anthropic suggested they were even a little bit interested. In making it easy to use third-party harnesses that compete with their tools. Indeed, whether you think that they're incorrect or not, I don't think it's an unreasonable take that Anthropic actually believes that they are making a concession here to officially allowing these third-party harnesses with the trade-off that they have to pay the API rate. Now, zooming out and not just talking about the competition between Claude Code and Cowork as a harness and all the other harnesses, many of Anthropic's moves also suggest that they have a pretty voracious appetite when it comes to owning the application layer. Just this week we've gotten updates for Claude for Legal, Claude for Finance, and the announcement yesterday of the new Claude for Small Business. Compare that to OpenAI, who doesn't have those sort of dedicated focuses and prepackaged plugins around these different sectors and professional functions. Whether it's the right approach or not, Anthropic clearly has a bias to be a little bit more Apple-ish in controlling the end-to-end experience. And to use the Apple-Google analogy, there are obviously people who feel very strongly on both sides of that argument. Developers and builders tend to be more aligned with the Android type of approach that's more open and allows people to do more, while end consumers who just like great experiences often find themselves in the Apple ecosystem because they like the consistency that, that end-to-end control comes with and don't really bother themselves all that much with all the taxes that creates for the developer ecosystem around that closed end-to-end experience. There's also at least a little bit of an argument that this is Anthropic flexing their pricing power. Anthropic is just increasingly dominant when it comes to enterprise usage. The Information spoke to a few enterprise customers and found that even mid-sized firms are readily signing up for 6- and 7-figure Anthropic bills. New data from RAMP backs this up, where the latest edition of the RAMP AI Index found that Anthropic had overtaken OpenAI in terms of business usage for the first time. 34.4% of RAMP's customers are now paying for Anthropic subscriptions compared to 32.2% for OpenAI. Anthropic's adoption rate has quadrupled over the past year while OpenAI's is basically flat. But honestly, as much as I do believe that Anthropic has a particular perspective that wants people in its own controlled ecosystem, I don't think that the core issue here is really about some acrimony towards third-party developers or even competing harnesses. I think the extremely obvious thing going on here is what I called a couple of weeks ago the end of the AI subsidy era. We have moved from a paradigm of caring about seats to caring about tokens. When you shift from assisted AI to agentic AI, the business model cannot be based on seats anymore, which try to normalize consumption across a big set of people. Agents that can do work in the background can consume the equivalent amount of AI usage in the form of tokens that dozens or even hundreds of those seats would account for. If we had unlimited compute infrastructure, it would be one thing, but we don't. We are crashing right now headlong into the inevitable market consequence of there being massively more demand for tokens than there is supply of tokens. And Anthropic has been telegraphing these changes for months. Everyone knows they've struggled with compute. Everyone knows they've struggled with latency issues. Everyone who's tried to use Claude or Claude code knows how frequently they've been a victim of their own success with the system going down. And for months, all signs have been pointing in the direction. That at some point soon, whatever subsidies there were were going to end. Claude Code's Tariq, for example, back in February wrote, "We want to encourage local development and experimentation with the agent SDK and Claude-P, but if you're building a business on top of the agent SDK, you should use an API key instead." Let's try to put some real numbers around this. As I said, it's not exactly clear what the full extent of the subsidy was, but to just take a couple of examples, Umang Jaipura wrote, in a recent session, Claude Code spent about $31 worth of API tokens, counting it as 7% usage over a 5-hour window. That backs out to about $100 per hour of usage for a $100 a month subscription, approximately the base comp of a mid-level software engineer. That is a huge subsidy for subscription usage. No wonder they don't want to subsidize programmatic usage. It will massively cannibalize API revenue. Last month, meanwhile, Anthropic estimated that the average cost per developer for enterprises is $13 per day or between $150 and $250 a month. And that's just for developers using Claude Code in the Claude Code Harness. And that's just the average. That's not trying to capture people running third-party services on top of the agent's SDK. Back in January, it was reported that Cursor estimated that a $200 subscription would allow $2,000 in compute, but some sources suggest that it could be as high as $5,000. In other words, the general consensus has been that if you maxed out your Anthropic plan, they were effectively offering a 10x to 20x type of subsidy. They are not yet taking that subsidy away from their main accounts. What they are doing is taking that subsidy away for usage that happens outside of their core ecosystem. And it just seems very clear to me that this is more about capacity constraints than it is about any sort of relationship with developers. Now, a lot of commentary was some version of this tweet from Wes Winder who wrote, they don't have a huge compute constraint anymore after the SpaceX deal though, which, sorry to pick on Wes, is so insane as a belief set that it kind of boggles the mind. Anthropic has unlocked a couple hundred thousand H100s and H200s, but when you are dealing with stratospherically growing demand increasing inference costs while also still having to figure out how to train the next generation of competitive models, With all of the required pre-training compute, even Colossus-1 is a drop in the hat. Nityesh from Evry got it closer to right when he wrote, been thinking about this announcement. I think it connects directly to something I've been thinking about since listening to the Dylan Patel episode on the Dwarkesh podcast back in March. Their core thesis, we are going into a semiconductor shortage till around 2030. It's not just AI chips, it's everything underneath memory, logic, ASML machines, TSMC fab capacity, power infrastructure. Every layer of the stack is constrained, and these are years-long bottlenecks. What happens when supply is fixed but enterprise demand is exploding? Consumers get rationed. Every big company is racing to deploy AI agents and copilots, happy to pay for Claude code on a usage-based model. This means Anthropic can sell the same Opus compute to an enterprise for way more than the $200 a month Mac subscriber. Now, one important point that he makes that I agree wholeheartedly with, and which is why I think OpenAI has done a good job so far to not try to revel in this moment, comes when he writes, "OpenAI is not going to do that for a few more months because they're not seeing the crazy ADX growth like Anthropic, but they will also follow a similar model in the less than 1-year range. So we might need to prepare for a world where consumers rely more on open-source models running on their own hardware until that gets supply constrained as well." Simply put, this situation was a complete inevitability. I simply don't believe that you can persist in a scenario of extremely limited compute and effectively unlimited and growing demand and expect tokens to stay subsidized. No matter how valuable it is to have the affinity of those who are once subsidized. And I agree, I think that OpenAI will be forced to make a very similar decision before very long at all. Do I think Anthropic handled the comms poorly? Big time. They are doing a spectacular job right now of alienating pretty much everyone with their communications in one way or another. But the real bummer here is not just that certain services that lots of people like using are going to cost more. It's that in general, As George from Product Management World puts it, the token maxing era is ending. We have had this very brief 6-month golden window where the coding models were good enough that we could create anything our little brains thought of without really having to worry about the costs. And now, unfortunately, because of the laws of physics and supply and demand, we're gonna have to start to consider the cost of what we do. That's a bummer, but it is the reality. I would say for now, if you're doing something that Anthropic doesn't particularly like or competes with their core services, run on over to Codex. And soak up as much subsidy as you can before they have to change. And honestly, and this is the dreariest part in some ways, for those of you who are using Claude, even directly in Claude code in the way that they want you to, I would not assume that you're going to get a 10x subsidy forever either. It seems more likely than not that market forces will over the next, I don't know, year, force even those subsidies to go away as well. That is the story for now. The times they are a-changing. And that's going to do it for today's AI Daily Brief. Appreciate you listening or watching as always, and until next time, peace.