0:00 Anthropic is now spending two-point-three times its payroll on compute.0:04 That means for every dollar they spend on an engineer's salary, they're spending two dollars and thirty cents on the machines for that engineer to use.0:14 Last week, in episode 102, we talked about that odd bug in GPT-5.5's Codex, and how it revealed the sheer complexity under the hood.0:22 Well, today OpenAI announced its successor, GPT-5.6 Sol Ultra, and it lands right in the middle of this new reality where the cost of AI is starting to eclipse the cost of the people building it.0:35 The conversation on Hacker News this week isn't just about what these new tools can do...0:41 it's about what they cost.0:42 And whether anyone can actually afford them.0:45 So, let's get into the headlines.0:47 First, the big one.0:49 OpenAI's Tibo Sottiaux tweeted this morning, simply, "Ultra will be in Codex." This refers to GPT-5.6 Sol Ultra, and it lit up the discussion boards immediately.0:59 Early access reports from corporate users are trickling in, suggesting this is a significant step up in capability for coding assistants.1:08 But, as we'll get into, there's a heated debate about what "Ultra" actually is—a new model, or just a new way of using the old one.1:16 Then there's this great, counterintuitive post from Dustin Bluck, the developer behind the Castro podcast app.1:23 He wrote a piece about trying to build customer loyalty through really detailed, thoughtful human support.1:30 And his conclusion?1:31 It mostly backfired.1:33 He found that his honest, in-depth answers to user emails were, in his words, "deeply unsatisfactory to users." It's a fascinating look at the gap between what we think users want from support and what actually works.1:47 It pours a little cold water on the idea that just being more human is the answer.1:52 And for a complete change of pace, the sleeper hit on Hacker News right now is something called OpenPrinter.1:59 It's exactly what it sounds like: a fully repairable, refillable, open-source ink printer.2:05 In a world of disposable tech and predatory ink cartridge models, this thing is a breath of fresh air.2:11 It's designed to last, to be fixed, and to use cheap, refillable ink.2:16 The community is absolutely loving it—it's a classic Hacker News story of engineering a better, more sustainable solution to a problem everyone hates.2:25 It’s a reminder that not all innovation has to come from a multi-billion-dollar foundation model.2:32 Sometimes it's just about fixing the damn printer.2:35 Finally, the slow-burn conversation around Anthropic's Claude Fable 5 continues.2:40 Users are still wrestling with regional lockouts, subscription changes, and the general friction of getting access to the latest models.2:49 It highlights this ongoing tension in AI.2:51 On one hand, you have these incredible tools being announced.2:55 On the other, you have the messy reality of regulation, geopolitics, and business models getting in the way of people actually using them.3:04 It’s a global rollout hitting a very non-global wall.3:08 So what does it all add up to?3:10 You have a new, more powerful—and presumably more expensive—GPT model.3:14 You have a stark warning that even dedicated human interaction has its limits.3:19 And you have this massive, underlying story about the astronomical cost of just keeping the lights on in an AI-native company.3:27 Okay, let's dive deep into that cost, because the numbers are staggering.3:32 And then we'll connect it to what's happening with GPT-5.6.3:36 The whole conversation was kicked off by a new analysis from Tom Tunguz.3:40 And his headline number is the one I opened with: Anthropic spends two-point-three times its payroll on compute.3:48 He breaks it down.3:49 Per engineer, that's an estimated five hundred and fifteen THOUSAND dollars a year.3:54 Just on compute.3:55 That's against a fully-loaded salary of about two hundred and twenty-four thousand dollars.4:01 Think about that.4:02 For decades, the most expensive part of a software company was the people.4:07 The brilliant, hard-to-find engineers.4:09 Now, for the first time, at least at the frontier of AI, the cost of the tool is more than double the cost of the artisan.4:17 Where have we seen this before?4:19 The pattern feels familiar, but the details are new.4:23 You could say it’s like the early days of the web, when companies had to build their own massive data centers.4:30 That was a huge capital expense.4:32 But this is different.4:33 This isn't a one-time buildout.4:35 This is an ongoing, operational cost that scales with the number of researchers you have.4:41 It's more like a pharmaceutical company where every single scientist needs their own, private, top-of-the-line MRI machine running twenty-four seven.4:51 The cost of doing research itself becomes the dominant factor.4:55 And Anthropic isn't entirely alone.4:57 Tunguz's data shows the median software company is already spending one hundred and thirty-seven thousand dollars per engineer on AI compute.5:06 The top one percent spend about eighty-nine thousand.5:09 So Anthropic is an outlier, for sure, but they're a leading indicator.5:14 They're showing us the future financial structure of a company that is ALL IN on AI.5:19 It's a world where your AWS or Google Cloud bill isn't just an expense line item; it IS the business.5:26 Tunguz projects three possible scenarios out to 2029.5:29 In the "Bear" case, token prices keep falling and efficiency improves, so costs stay manageable.5:35 In the "Base" case, costs for most companies level off.5:39 But in the "Bull" case—where the rest of the market starts to look more like Anthropic—the average compute spend per engineer could hit almost six hundred thousand dollars a year.5:51 That would be two hundred and thirty percent of their salary.5:54 This bull case is backed by projections from places like Goldman Sachs, who see a 24-fold rise in token consumption by 2030, driven by new "agentic" workflows where AIs are doing tasks in the background, constantly thinking, constantly burning tokens.6:11 Now, there are powerful counter-arguments.6:13 Token prices have been falling by about ten times per year for three years straight.6:19 And open-weight models are getting better and better, offering a much cheaper alternative.6:25 So which is it?6:26 Are costs going to the moon, or will efficiency and open source save us?6:30 The answer from the market this week seems to be...6:33 yes.6:34 Both are happening at once.6:35 And that brings us to GPT-5.6 Sol Ultra.6:38 This announcement couldn't be more perfectly timed to illustrate the tension.6:43 On one hand, here is a new, more powerful, more capable tool.6:47 It promises to make developers more productive, to write better code, to solve harder problems.6:53 And the immediate reaction from every company is: we need that.6:57 As Tunguz writes, "If a rival ships features faster, the AI bill stops being optional." You have to keep up.7:04 But what is Ultra?7:05 The initial hype suggested a whole new model.7:08 But the more skeptical and informed comments on Hacker News are painting a different picture.7:14 One user, Szpadel, who seems to have some inside knowledge, put it this way: "There is no ultra effort level implemented on the backend.7:23 It’s just an alias in the codex to a max effort setting and a single line addition to the prompt to use subagents proactively." Okay, let's translate that.7:33 It means "Ultra" might not be a new brain.7:36 It might be a command that tells the existing brain to think harder, for longer, and to spin up smaller, specialized sub-brains to help it work on a problem.7:46 It's like the difference between upgrading your car's engine versus just flooring the accelerator and turning on the nitrous.7:54 You're getting more performance from the same machine, but you are burning WAAAY more fuel to do it.8:00 And this is where the two stories—cost and capability—collide.8:04 This "Ultra" mode is exactly the kind of agentic workflow that Goldman Sachs predicts will drive that 24x increase in token usage.8:13 And we're seeing the consequences of this inside companies RIGHT NOW.8:17 Another user on Hacker News, writing from a corporate account, shared this gem: "Last year management was showing us scoreboards, praising leaders who used the most tokens...8:28 the message is that we should use cheaper models whenever we can." There it is.8:33 That's the entire story in two sentences.8:36 The whiplash from "Use this amazing new tool as much as you can!" to "Oh, wait, please stop using it so much, it's costing a fortune." It’s a pattern we’ve seen before, but never this fast.8:48 Think about cloud computing.8:50 It took years for companies to go from "move everything to the cloud!" to developing complex FinOps teams to control the spending.8:58 With AI, that cycle seems to be happening in months.9:02 One quarter, you're a hero for driving adoption.9:05 The next, you're on the hook for a seven-figure overage.9:09 So you have this incredible new capability being dangled in front of every developer.9:14 At the same time, you have the CFO walking down the hall asking why the cloud bill just tripled.9:20 This is the fundamental tension of applied AI in 2026.9:24 It's not a technical problem anymore, it's an economic one.9:27 It’s a battle between the desire for progress and the reality of the budget.9:32 This week sets up a conflict that's only going to get more intense.9:37 We now have models that are so powerful, the primary question is no longer "what can it do?" but "what can we afford to let it do?".9:45 The excitement around GPT-5.6 is real, but so is the quiet anxiety in the accounting department.9:51 The race for artificial intelligence was always about processing power.9:56 But we're now in an era where the budget for that power—not the brilliance of the engineer—is becoming the final gate on innovation.
