sudoferraz

been gone from youtube for 6 months in that time creem went from washing coffee cups in the bathroom sink to having our name on a sign in the lobby new office tour, a halloween card game that sent me home early, and a dog that's basically a tenant now youtu.be/9q0MBGsOiRg

anthropic filed its S-1, and now everyone is arguing about AI pricing again. I read the filing, listened to what dario has been saying, and went through the main posts. here's the actual data how anthropic charges today they call it "pay-for-intelligence". in practice, customers pay for the tokens they use, with different prices for different models for the first 9 months of 2026, anthropic made $33.2B: - $26.3B from consumption, 79% - $6.9B from subscriptions, 21% consumption was 75% of revenue in 2024 and 83% in 2025. the amended filing says 79% for the first 9 months of 2026, and anthropic expects it to stay "the substantial majority of total revenue for the foreseeable future" where outcome pricing comes up only in the risk factors. on p.62, anthropic says it might have to offer "alternative pricing models, such as outcome- or task-based pricing" the same section says agentic products are "structurally more compute-intensive", and if prices don't account for that, "our compute efficiency could deteriorate as agentic usage grows" a few other numbers worth knowing - cost of revenue was $4.8B for the 9 months, mostly compute used to serve customers - total compute and infrastructure spend was $18.8B - operating loss was $1.3B - Opus 5.5 reduced customer cost per task by about 40% compared with Opus 5 - net revenue retention was about 900% what dario said on dwarkesh patel's podcast in february, dario said API pricing "is more durable than many people think", but also: "Not every token that's output by the model is worth the same amount" his example: tokens telling you to restart your mac might be worth a few cents. tokens helping a pharma company improve a molecule "could be worth tens of millions of dollars" then he said: "At some point we're going to see 'pay for results' in some form, or we may see forms of compensation that are like labor, that kind of work by the hour. I don't know." how people are reading it - @guilleflorvs put dario's clip next to the revenue numbers, and it got more than 1.3M views - @martinvars compared it to telecom in the 90s, when every minute cost the same regardless of what the call was worth - @VictorLShammas compared it to a power company charging more when your electricity produces something valuable, and called it "techno-rentierism" the three pricing models, simply - usage-based: pay per token or API call, including attempts that don't work - subscription: pay a fixed amount per user or tier, usually with usage limits - outcome-based: pay when the task is completed, while the provider absorbs failed attempts and extra compute the amended S-1: sec.gov/Archives/edgar/data/…

got a new toy for my desk 🎙️ since the scrawl + grok bot workflow I'm basically talking to my computer all day, circling stuff on the screen and just saying what I want done already apologized in advance on our office slack. pretty sure the team is about to think I'm the most obnoxious person in the building

got a new toy for my desk 🎙️ since the scrawl + grok bot workflow I'm basically talking to my computer all day, circling stuff on the screen and just saying what I want done already apologized in advance on our office slack. pretty sure the team is about to think I'm the most obnoxious person in the building

been saying this for a while, AI is supply constrained and most of what we use today is subsidized by the labs. amjad's right that cheap subscriptions won't hold replit moving to usage based makes sense, but I think it's a halfway step. compute keeps getting more expensive, so charging per token just passes that pain to the customer. and that only really works for technical people who get what a token is, the average consumer has no idea what they're paying for and what nobody talks about is who pays when the AI gets it wrong. with usage based the customer pays for the tokens even when the output is useless, they take on all that risk from the start. at some point someone else has to carry it, and I don't think it'll be the customer forever

openai was about to cut chatgpt pro usage in half at the same $200, people got angry, and a day before it kicked in they doubled limits instead feels like a win but I don't think it changes much. everyone talks like AI demand is infinite, but the real constraint is supply. compute is expensive and it's not getting cheaper as fast as people are using it a16z shared that only about 2% of US households pay for any AI subscription, and a lot of what we all use today is being subsidized by the big labs if compute keeps getting more expensive, charging for usage gets harder and harder. you either cut limits and make people angry, or eat the cost and usage never captured the value anyway. the agent that closes a deal or fixes a bug burns the same tokens as one that does nothing useful that's why I keep coming back to outcome based billing. charge for what actually got done, not for access honestly not sure how long cheap AI lasts, but I don't think it's forever

greg's right that dashboards are going free, but I think the bigger shift is that people stop paying for the tool and start paying for the thing getting done a couple months ago I said outcome based billing will eat the AI industry, still think so for example, an SDR agent charging $29/mo feels disconnected when it books a meeting that turns into a big deal. it should charge when a qualified meeting actually happens, like $500 per meeting but then who decides the meeting was qualified? the agent, the customer, the CRM? what if they no-show, or the deal closes 3 months later? who has the proof, and when does the charge actually happen that's the real hard part and almost nobody is solving it yet. the pricing idea is easy, agreeing what counts as done is not we're working on it at creem, there's a lot to figure out still seats had a good run

saw kit's scrawl demo last week and had to try it, it's great for engineering but I'm doing so much stuff nowadays that are not related to engineering at all, and it's working amazingly well with @grok bot for example, if I'm talking with a customer on email, I now open the email, hit record, circle the bits i'm answering and just talk. grok bot gets the transcript and my drawings, finds the thread and drafts the reply for me to confirm or sends it directly this one took a 13 second recording. on Tuesday i cleared 11 of these in about 3 minutes. zero code involved.