The Collapsing Cost of Intelligence, Twitch Training Opt-Outs, and $13B Vibe Coding

Yesterday in AI  |  14 August 2026

The Collapsing Cost of Intelligence, Twitch Training Opt-Outs, and $13B Vibe Coding

Frontier intelligence costs collapsed into pennies this week as rapid model release cycles, browser agent integration, and AI-enabled employment scams made headlines. This episode breaks down xAI releasing Grok 4.6 at 60% below standard frontier pricing while competing with top-tier models on benchmark performance and agent efficiency.

We examine DeepSeek rolling out V4-Pro-0813 for under 90 cents per million tokens alongside Google dropping Gemini 3.7 Flash on a lightning 3-week release cadence. We dissect The Wall Street Journal's investigation revealing how North Korean operatives use AI tools and stolen identities to secure remote corporate jobs and funnel $800M annually. We contrast Amazon training AI on Twitch streamer content by default against Apple's nine-figure publisher deals, look at Anthropic embedding Claude Cowork sessions directly into Chrome, and explore the soaring valuations of vibe-coding platforms Lovable and Cognition.

Feedback? Email mike@yesterdayinai.news or connect on LinkedIn, X, or Bluesky. If you like the show, please take a minute to rate and review it so others can find it!

Transcript

SPEAKER_00 0:05

Hi folks and welcome back to another edition of Yesterday in AI, your daily digest of everything happening in the world of AI in roughly 10 minutes. I'm Mike Robinson. It's Friday, August 14th, and the price of Frontier AI just fell off a cliff. Which is great news for you and slightly less great news for a couple of people who really shouldn't have cheap, capable AI in their hands. Let's get into it. Let's start with the model that spent last year as a punchline. A year ago, if you said Grok and Frontier model in the same sentence, people politely changed the subject. On Wednesday, SpaceX AI shipped Grok 4.6 and nobody's changing the subject anymore. Here's why. On the Artificial Analysis Intelligence Index, which is basically a report card that scores how smart a model is across a big pile of tasks, GROK 4.6 landed at 61. That puts it behind only Anthropex Opus V at 63 and Fable V at 62. And it sailed past OpenAI's GPT-5.6 SOL. It edges Fable on real professional and legal work, and it beats Sol on a couple of coding tests the older Grok had been losing. But the number that actually matters is the price. Grok 4.6 runs about $2 for a million words in and $6 for a million words out, roughly 60% under what the Frontier charges. And it's efficient in a way that compounds. On one long agent test, Grok finished the job in about 53 turns, where Claude Opus V Max took about 103. Think of an AI agent like a contractor. Two of them can build the same deck, but one keeps running back to the hardware store and the other shows up with everything in the truck. Over an eight-hour job, the guy with the truck costs you half as much. Some of this traces back to the cursor partnership, where Grok got baked into one of the most popular coding tools and got a lot of real-world reps. Whatever the reason, the model people used to mock is now a serious pick for anyone running agents all day. Because on a long job you care less about who wins one clever answer and more about who finishes the whole task for the least money. Elon Musk called it objectively number one when you consider intelligence, speed, and cost, and teased Grok 4.7 as three to four weeks away and better than everything out there. I'll believe the 4.7 promise when I can run it myself, and I'd love to see independent testers confirm these scores before we crown anybody. But the punchline clearly grew up. That's real. And here's the thing about Grok's big weapon. Price isn't a secret, and it isn't the only one swinging it. On Thursday, DeepSeek quietly rolled out V4 Pro 813. The price? About 43 cents for a million words in and 87 cents out. Put that next to Grok's $2 and next to the Frontier's dollars per million, and you can watch the whole industry slide from dollars into pennies. And this is no toy. V4 Pro beats Anthropic's Opus 4.8 on a handful of coding and cybersecurity benchmarks. This is also the Chinese lab that in July ranks second only to Anthropic in total worldwide usage, so the cheap pricing is clearly pulling real volume, not just curiosity clicks. The same day Google shipped Gemini 3.7 Flash. It came out three weeks after Gemini 3.6 Flash. Three weeks. At half the intro price of the model it's replacing, with big jumps on the coding and agent tests. Tokens, by the way, are just the meter these things run on, like a taxi that charges you by the word. And right now, every taxi in town is slashing the fare at the same time. The cadence is really comical. Blink and there's a new model. The version numbers are moving faster than I can update the show notes. Now the honest trade-off? Cheaper and faster is wonderful for anyone building with this stuff, but a three-week release cycle is also how you ship a model before anyone's really kicked the tires on it. To Google's credit, 3.7 Flash added new safeguards for chemical, biological, and cyber misuse. So at least somebody's treating raw capability as something to fence in, rather than just a bigger number to brag about. Cheap capable AI in everyone's hands is mostly a gift. The catch is that everyone includes people you'd rather not hand power tools to. Which brings us to a story that landed Wednesday from the Wall Street Journal after a full year of reporting. North Korea has built a secret workforce inside American companies. The method, stolen identities, AI tools, and a few U.S. accomplices, all aimed at landing remote jobs. The journal says thousands of operatives have used fake or stolen IDs, and one single team got hired at at least eight U.S. companies in just a few months. The whole scheme is estimated to pull in around $800 million a year, with hundreds of millions of it funneled straight back to the regime. The journal built the story from leaked data, interviews, and previously unseen videos showing exactly how these workers get hired, and AI is what makes it scale. It writes the polished resume, coaches the fake candidate through the interview in real time, and lets one person juggle a dozen employers at once without tripping over their own lies. You know that quiet remote hire who's great at the work but whose camera is mysteriously always broken. Yeah. And the stolen paychecks are only half of it. When you hire one of these operatives, you've quietly placed a state-linked person inside your real systems, with real logins to your code, your customer data, and your internal chat. So a payroll problem turns into a national security problem the moment they decide to use that access for something other than collecting a salary. The practical takeaway hits anyone who's ever been hired for a remote job or hired someone for one. Identity verification, checking where a device actually is, least privilege access, meaning you only get the keys to the rooms you need. That stuff used to be HR paperwork. Now it's a security control. And North Korea wasn't alone this week. Taiwan said AI agents helped attackers hit its government systems back in July, with one operator delegating the reconnaissance and password theft across a whole team of bots. The same tools that let a hobbyist build an app and an afternoon let a sanctioned government run payroll fraud at industrial scale. All of this runs on data, your data, and this week two big companies gave you two opposite answers on whether they'll bother to ask for it. Door number one is Amazon. On Wednesday, Amazon said it'll train generative AI on Twitch streamers' content by default. Every clip, every stream, fair game, unless the creator digs into the settings and opts out. Door number two is Apple, which is reportedly working on nine-figure multi-year deals to pay news publishers when Siri uses their reporting. Actual money, up front, for the content. Two doors, same building. One takes your stuff and makes you find the off switch. The other pays you at the entrance. An opt-out by default is the industry's favorite move for one very cynical reason. Almost nobody opens the settings menu. If your work is good enough to train a billion-dollar model, it's good enough to ask you first. That goes for a Twitch streamer, a newsletter writer, or anybody who's ever posted a photo they were proud of. Once a company has the data and a cheap model to run it through, the next move is obvious. Put the AI right where you already work. That's exactly what Anthropic did this week. It upgraded Claude in Chrome so the little side panel in your browser now runs a full co-work session. In plain terms, your conversation history saves to your account, your existing skills and connectors just work with no setup, and a multi-step task you start in Chrome picks up later on your laptop or your phone. It's like a coworker who's suddenly sitting in the corner of every browser tab you open. Picture the boring version of useful here. You're staring at a messy travel booking page, and instead of copying details into a separate chatbot in another tab, you just ask Claude through the panel right there to compare the options and fill in the form. It already knows your calendar because you connected it once. That's the pitch. And it's a good one. Handy? Absolutely. But this is also the version of Claude that can read and act on the page you're looking at, so it's worth knowing exactly what you've handed it the keys to before you let it start clicking around your bank tab. This is where the whole industry is heading. AI stops being a website you visit and becomes a layer sitting on top of everything you already do. Convenient and worth one deep breath about permissions. And the flip side of all this, the cheap models, the AI writing shotgun in your browser, is honestly the most fun part. You don't have to wait for some company to build the thing you want anymore. You can just build it. Case in point, Lovable, one of the vibe coding startups, where you describe an app in plain English and it writes the software, raised $400 million on Wednesday at a $13.3 billion valuation. That's more than double what it was worth back in December, on a revenue pace near $600 million a year, and Cognition, another AI coding shop, is reportedly in talks to raise at a $40 billion valuation after crossing a billion dollars in annualized revenue. Big dizzying numbers, and they tell you investors think this just describe it and get software thing is only getting started. But the number I actually loved this week came from a reader, Erica, who told the rundown her story. Her family kept missing each other's schedules. They wanted one of those skylight wall calendars for years, but it runs $300. So instead she wrote up what she wanted in Claude, handed that to Lovable, and it built her a family calendar from scratch. It pulls in everyone's Google Calendars and their to-do lists, rotates family photos in the background, and lives on the kitchen iPad. It cost her less than the Skylight, and it fits her family exactly because she built it for her family. And that's the whole point. And that's the show. But just a quick reminder, I'm doing a special interview later today with my guest Bianca Baumann, VP of Learning Solutions and Innovation at Ardent Learning, for an extended episode that will come out on Monday. We're going to discuss why so many AI deployments stall out after the initial launch and possible solutions on how to fix that. I hope you'll enjoy it. Now, if you have feedback for me, email Mike at yesterdayNai.news or connect with me on LinkedIn, X, or Blue Sky. If you enjoy Yesterday and AI, please take a minute to rate and review the podcast wherever you listen. Or share it with a friend. Thanks for tuning in today. Stay curious, and I'll see you tomorrow.