The $5 Model Revolution, DeepSeek's $71B Pause, and Google's Autonomous Agents

Yesterday in AI  |  27 July 2026

The $5 Model Revolution, DeepSeek's $71B Pause, and Google's Autonomous Agents

The artificial intelligence industry experienced a dramatic shift in economic fundamentals this week as the cost of top-tier intelligence plummeted and specialized hardware bets surged. This episode breaks down Anthropic's launch of Claude Opus 5—delivering near-frontier performance and 1M context at half the cost—and what it means for SaaS software moats and subscription renewals.

We explore Etched's massive $300 million funding round and $1 billion in pre-orders for low-voltage inference chips designed to challenge traditional GPU architectures. We analyze the fallout behind DeepSeek's paused $71 billion fundraising round following a leaked founder transcript, dissect South Korea's "San Francisco AI Declaration" and high-bandwidth memory strategy, look at Google's proactive "Gemini Intelligence" agent layer across 40+ apps on Samsung devices, and cover Wispr's new research lab aiming to build a contextual "Jarvis" voice assistant.

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, this is Yesterday in AI, your daily digest of everything happening in the world of AI in roughly 10 minutes. I'm Mike Robinson. It's Monday, July 27th, and the past few days had one theme: price. The cost of a smart answer fell off a cliff. The cost of building the machines that serve those answers got a fresh billion in bets, and the people who used to have the cheapest labor in the room got nervous about a $5 model doing their job. Let's get into it. Start with the headline because Anthropic didn't bury it. On Friday, they shipped Claude Opus 5 and the pitch is blunt. This is close to the frontier intelligence of their top model, Fable 5, at half the price. Same OPIC as before, $5 per million tokens in, $25 out. Tokens are the word chunks these models read and build by, plus a million token context window, big enough to hold a small library in its head at once. Here's why the price matters more than the benchmark. For two years the rule was simple. The smartest model was also the most expensive, so you rationed it. You used the cheap model for grunt work and saved the good one for the hard problems. Opus V sands that trade-off down. On the agentic tests, the ones where the model has to actually do a multi-step job instead of just answer a question, it's landing near the top of the table while costing a fraction of what that tier used to. On one computer use benchmark, it beats Fable V's best result at roughly a third of the cost. So what breaks? The Neuron put it bluntly. A lot of software companies are cooked. If your product was a thin wrapper around we'll organize your text or we'll draft your emails, and the model underneath just got twice as capable for the same money, your moat was rented. You didn't own it. That's not doomed for everybody. But if you're a business looking at your SaaS renewals, this is the quarter you start asking which vendors are selling you a feature, the base model now does for free. Cheaper intelligence only works if something can actually serve it, fast and at scale, and that's where the money went next. A startup called Etch raised $300 million on Thursday and says it's sitting on about a billion in pre-orders at a $10.3 billion valuation. Sequoia led it with S.K. Heinex and Jane Street in the mix. Etch builds chips for running models, not training them, which is a different animal. Training is the one-time cost of teaching the model. Inference is the forever cost of answering every question every day for every user, and as agents run in the background for hours, that bill never stops. Their trick is low voltage inference, running the math engines at under half the voltage of a normal AI chip. Less heat means the chip doesn't slow down to cool off, so more of it stays working, more answers per dollar. Think of it like a race car versus a delivery fleet. Nvidia sells the race car, the thing that wins the training benchmark. Etched is betting the real money is in the fleet, thousands of boxes quietly serving tokens at 3 in the morning. Whether they can ship racks at volume is the open question, but a billion in orders showing up before the product's fully out tells you where the smart hardware money thinks the bottleneck moved. Cheap intelligence and cheap serving are the story China's been telling for a year, and this weekend it bit one of China's own. Deep Seek, the lab that spooked everyone in early 2025 with a cut rate model, abruptly paused its second fundraising round. This one was targeting something like $71 billion. Then a leak blew it up. Somebody leaked a transcript from a roughly four-hour meeting between founder Liang Wenfang and his investors, and it went viral on Chinese social media. What did he say that was so radioactive? He was honest. He argued the gap between China and the US comes down to computing power, not talent, and admitted China's heavy dependence on Nvidia chips. Those remarks sounded like American conventional wisdom, not the confident we've already caught up line the Chinese AI conversation usually runs on. Cander got him in trouble and the round is on ice while the company manages the fallout. That tension between national pride and national strategy showed up in a friendlier form a few thousand miles east. South Korea used an AI summit in San Francisco on Friday to plant a flag. President Li J. Myung unveiled what he's calling the San Francisco AI Declaration, and he brought the whole room with him. Sam Altman from OpenAI, Dario Amade from Anthropic, Jensen Huang from Nvidia, and Broadcom's CEO were all on stage. The play is smart, and it's specific to what Korea already has. They make memory chips, the fast storage every AI system needs, and lots of it. So instead of out-researching OpenAI, which is a losing game, Korea's pitch is to be the country you can't build AI without. Chips, data centers, deployment, the whole chain, and they paired it with a domestic push called AI for all, aiming to reach 50 million people by year-end across public services, welfare, and schools. South Korea's play is a masterclass and focus. Defend and scale the specific high-value link in the chain that you actually own, rather than wasting capital chasing a glamorous race someone else is already winning. While nation states secure supply chains, consumer tech giants are scrambling to put autonomous capabilities into the phone in your pocket. At Samsung's Galaxy Unpacked event in London on Wednesday, Google rolled out the first pieces of what it's branding Gemini Intelligence. Be careful with that name, it's slippery. This is not a new Gemini model. It's a proactive assistant layer. Software that goes and does things for you instead of waiting to be asked. Think less chatbot, more errand runner. The headline capability is task automation, and the number that matters is 40. In February, this touched a handful of apps. Now Gemini reaches into more than 40 of them to do your quote, life admin. Order the ride share, book the dinner reservation, the flight, the concert tickets. Show it a photo or a shopping list and it reads the image as the instruction, then works in the background while you go do something else. You step in, tweak one detail, hand it back. If that sounds familiar, it should. Remember those agentic tests I mentioned on Opus V? Models doing multi-step jobs? That's this, just wearing a consumer coat. The Frontier Labs are teaching models to run multi-step jobs, and Google is racing to put that behavior on a phone your mom already owns. They also pre-installed Gemini Notebook, formerly Notebook LM, on the foltables, which turns your documents into slide decks, quizzes, even a podcast. And Gemini's spreading to the Galaxy Watch and Glasses this fall. Now, my grounded take. Phone assistants have promised to run my life since the first Siri demo, and mostly couldn't. Google's own footnote tells the story. Supervise closely, interrupt when needed. That's the language of a product that mostly works. And mostly works as a different animal when your credit card is on the line. The capability is real and improving fast, but the gap between a launch demo and letting an agent book a $300 flight while you look away is exactly what everyone should measure before trusting one with money. And Google isn't the only one chasing this, which brings me to my favorite story of last week. Whisper, the dictation tool quietly running inside 450 of the Fortune 500, launched its own research lab with a goal that is not shy about. Build Jarvis. Yes, just like Iron Man. The CEO has wanted that since he watched the movie as a 10-year-old in India. He poached a chief scientist off Amazon's original Alexa team and has committed hundreds of millions to hire 50 researchers, all to build the voice assistant that watches your screen and acts before you ask. Same dream Google is selling, but coming from a $2 billion startup instead of a trillion dollar giant. When a dictation app and Google are sprinting to the exact same finish line, the keyboard's days are numbered. And that's the show. If you have any feedback for me, email Mike at yesterdayandai.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. Thanks for tuning in today. Stay curious, and I'll see you tomorrow.