Grok 4.6 is out, and SpaceXAI is now openly saying it wants to train future versions on the sum total of everything its employees do. Bitcoin is stuck near $63,500 after an in-line July CPI and a softer PPI print, with traders already looking past the data to Jackson Hole. BlackRock quietly slashed the in-kind creation threshold on its IBIT Bitcoin ETF from $25 million to $1 million, opening the door to family offices and mid-sized RIAs. Goldman Sachs is spending $2.25 billion to buy NEOS, jumping straight into the Bitcoin income ETF business. And Metaplanet had to publicly deny selling any Bitcoin after moving 5,014 coins between its own custody addresses for a grand total of about $8 in fees. Let's get into it.
SpaceXAI dropped Grok 4.6 on August 12, and the headline number is a 61 on the Artificial Analysis Intelligence Index. That ties it with OpenAI's GPT-5.6 Sol, beats Moonshot's Kimi K3, and puts it in the fourth spot overall. It still trails Claude Opus 5 and Fable 5, but the gap is narrower than it was six months ago. The pitch is long-running agents. Pricing starts at $2 per million input tokens and $6 per million output tokens, with a 500,000 token context window. It's live in Cursor, Grok Build, and through partners like OpenRouter, Vercel, and Cloudflare. First week gives you 2x usage on Build and Cursor.
The technically interesting part is how they trained it. Instead of just rewarding correct code on the first try, they trained it to catch its own mistakes, pause, and recheck without losing the plot on long tasks. They used Grok 4.5 to regenerate supervised fine-tuning trajectories, filtered out bad traces with model-based checks, and ran agentic reinforcement learning across coding, kernel optimization, web dev, and CAD. The New Stack put it well, SpaceXAI is training on the stuff most labs throw away, the failures.
But the bigger story is what Musk said at the all-hands. He wants to train future Grok models on the sum total of all SpaceX information, including how the roughly 14,000 to 15,000 SpaceX employees actually do their work. And here's why that matters. For agent models, the scarce data isn't text on the internet, it's behavioral data. How people navigate tools, how they solve problems, how they coordinate. That's exactly what internal workforce data captures.
Meta tried something similar with its Model Capability Initiative and had to pause it after private employee conversations became accessible internally. A SEV 2 incident. SpaceXAI hasn't said what data is in scope or whether employees can opt out. Grok 4.7 is reportedly 3 to 4 weeks away, Grok 5 by year end. If this data pipeline actually works without a Meta-style blowup, it's a serious moat. If it doesn't, it's going to be an ugly headline.
Medical AI is quietly moving from demo to deployment, and the news this week shows what that actually looks like on the ground. Stanford Medicine published lessons from ChatEHR, their large language model embedded in the electronic health record. The headline finding is blunt. Benchmarks don't tell you how the thing behaves in a real clinic. They caught the model making unsupported claims, and they had to build ongoing clinician-facing monitoring to detect and log problematic outputs in real time. Governance, audit trails, escalation paths. The boring stuff that actually keeps patients safe.
Meanwhile, 12 major US health systems just formed a Diagnostic AI Consortium. Advocate, Cedars-Sinai, Mount Sinai, Northwell, Northwestern, and others, covering roughly 20 million patient visits a year. Aidoc provides the infrastructure. The point isn't to test one algorithm. It's to test how AI behaves across different scanners, different populations, different workflows. First findings expected in 2027.
Atlantic Health and K Health rolled out PatientGPT inside Epic MyChart. It uses the patient's actual EHR data to triage symptoms and hand off to a virtual primary care visit without making the patient repeat their history. Vanderbilt got a $600,000 grant from Eli Lilly to build an AI triage agent that speeds up referrals for Alzheimer's patients who need anti-amyloid infusions. And in South Korea, DGIST is building an autonomous ultrasound robot for breast cancer screening. Eight AI agents coordinating in real time, guiding the robot's movements, with a $7.4 million budget through 2028.
What ties these together is the shift from AI as a chatbot to AI as an agent doing sustained, consequential work inside a workflow. That's the same story as Grok 4.6, just with much higher stakes. If the model hallucinates about your code, you rerun it. If it hallucinates about your medication list, someone gets hurt. Stanford's lesson is the one everyone else is about to learn.
BlackRock just cut the in-kind creation threshold on IBIT from $25 million down to $1 million. That's a 96% reduction, announced by Robbie Mitchnick on Bloomberg TV. And Mitchnick said the goal is to keep pushing it toward zero.
Here's why this matters. In-kind means an authorized participant delivers actual Bitcoin to BlackRock and gets IBIT shares back, or the reverse for redemptions. No cash sale, no immediate taxable event. Cost basis and holding period carry over. The SEC only approved this mechanism for spot Bitcoin ETFs in July 2025. Until now, the $25 million floor meant only the biggest Wall Street desks could use it. At $1 million, family offices, mid-sized RIAs, and large individual holders can convert appreciated Bitcoin into IBIT shares without triggering capital gains. IBIT already holds around 777,000 BTC, roughly 81% of weekly category inflows. This move locks in the moat.
At the same time, Goldman Sachs is buying ETF manager NEOS for $2.25 billion. NEOS runs a $30 billion ETF business including Bitcoin and Ether income funds. Goldman's derivative ETF platform goes to $130 billion in total assets. The direct target is BlackRock's BITA income fund, which Goldman is leapfrogging by roughly 19x on the Bitcoin yield side. Bitcoin income products, the covered-call style ETFs pitching distribution yields north of 20%, are becoming a real category, and Goldman just bought its way to the front.
And zoom out on custody. Coinbase still holds 80 to 84% of US spot Bitcoin ETF assets. ICE is pitching its Digital Trust arm as a secondary custodian. Copper's US arm just became a FINRA member and SEC-registered broker-dealer. Zerohash, Morgan Stanley's infrastructure partner, got rebuffed on its trust bank application and is refiling. The plumbing layer for institutional Bitcoin is being built out fast, and Coinbase's dominance is finally getting real competition. The line from Grayscale this week captures it. Millions of everyday savers will own Bitcoin without ever downloading a crypto app. Through their advisor, their brokerage, their retirement account. That future runs on this infrastructure.
Bitcoin treasury companies had a rough day. Metaplanet moved 5,014 BTC, about $322 million, between its own custodial addresses. Total network fees, $8. Lookonchain lit up, rumors flew about a sale, and CEO Simon Gerovich had to come out and publicly confirm that no Bitcoin was sold and the company still holds 43,000 BTC. This is a good reminder that on-chain transparency cuts both ways. Every move gets seen, and every move gets misread.
More substantively, both Metaplanet and Strategy are sitting on serious unrealized losses given Bitcoin near $63,500. CoinDesk flagged this as the risk of concentrating on just one token. Strategy actually sold $100 million to defend the $100 price on its STRC preferred stock. STRC trades near $95, and DeFi is now packaging that risk into a tranched product where the senior tranche only starts absorbing losses if STRC falls below $47.66. So a 7% quote-unquote safer yield built on a leveraged Bitcoin treasury vehicle. Financialization arrives quickly.
On the fundraising side, Metaplanet unveiled BitBonds, a $1.3 million private debt sale, unsecured, up to 4.3% annual interest. Small, but it's a template. Brazil's largest Bitcoin treasury firm is planning an ETF with a 95% allocation to Strategy's STRC. Bullish, the NYSE-listed exchange, still holds 19,990 BTC worth $1.28 billion heading into today's earnings. BitGo added 74 BTC in Q2, bringing its corporate stack to 2,523.
The market context is unforgiving. Glassnode says short-term holders are pinning Bitcoin below $68,700 as they try to break even on underwater positions. CryptoSlate flagged a wall of 1.79 million BTC sitting between the current price and $65,000. Options traders are still paying for downside protection to $60,000 even while betting on $70,000. Arthur Hayes says the real catalyst he's waiting for is a $60 billion Fed cap on the FIMA repo facility. Until then, the treasury companies are stuck. The strategy of buying Bitcoin with borrowed money works beautifully in a bull market. In a range, it's a stress test. And we're in the middle of finding out who can hold on.
The pattern this week is agents everywhere. Grok training on employee behavior, AI triaging patients in the EHR, and BlackRock building the pipes so your brokerage account can hold Bitcoin without you thinking about it. The interfaces are getting quieter. The infrastructure underneath is getting louder. Watch who owns the plumbing.