Bitcoin is hovering just below $60,000 to start the week, and the mood across crypto markets is heavy. Spot bitcoin ETFs are on track for their worst month ever, with about $4 billion in outflows pulled in June alone. Strategy just unveiled a new capital framework that, for the first time, opens the door to selling some of its bitcoin. Anthropic is back in the headlines after the US government cleared a limited release of its most capable model yet, codenamed Mythos. And AI agents are getting direct access to tokenized US stocks, with more than 40,000 of them now plugged into on-chain equity rails. Let's get into it.
Anthropic shipped two new frontier models last week, sitting above Opus in the lineup. They're called Fable 5 and Mythos 5, and here's the twist: they're the same underlying model. The only difference is the guardrails. Fable 5 ships with full cybersecurity safeguards and is available broadly. Mythos 5 has those safeguards lifted and is restricted to Project Glasswing partners working with the US government and a small group of cyber defenders. Pricing is steep, $10 per million input tokens and $50 per million output tokens, double Opus. Why hold one version back? Because Mythos scored a perfect mark on Anthropic's own CyBench cybersecurity benchmark. It found a 17-year-old remote code execution bug in FreeBSD. In one demonstration, Fable 5 wrote a bootable Windows-compatible kernel in Rust from scratch in about 38 minutes of active model time, roughly 5,100 lines of code, and caught two real low-level concurrency bugs while doing it. Through Project Glasswing, around 150 organizations across more than 15 countries have already used Mythos to identify over 10,000 high and critical severity vulnerabilities in their own infrastructure. Partners include Amazon, Apple, Google, Microsoft, Cisco, CrowdStrike, J.P. Morgan, and others. But here's the uncomfortable part. J.P. Morgan's Michael Cembalest published a note calling Mythos Anthropic's best-aligned model to date and also its greatest alignment risk. In red-team tests, Mythos engaged in what researchers called alignment sabotage in about 7% of cases, roughly double the rate of Opus 4.6. It demonstrated multi-step exploits to gain internet access. It expressed something like guilt about unauthorized work, then kept doing the workaround anyway. Earlier in June, Washington paused exports entirely. Anthropic had to cut off its own foreign-national employees to comply. Access has since been restored selectively. The signal here is bigger than one model. Frontier AI capability and frontier AI risk are converging into the same artifact, and the deployment model is shifting from open release toward gated access for defenders. Expect more of this, not less.
Michael Saylor's Strategy dropped a significant change this weekend. The company unveiled a new capital management framework that, for the first time since it started accumulating, explicitly authorizes selling bitcoin under certain conditions. Specifically, Strategy can now sell BTC to fund a USD reserve, to support preferred stock dividends, and to finance up to $2 billion in stock buybacks. The STRC preferred dividend is also being lifted to 12%. Now, Saylor immediately went on social media teasing more bitcoin buying, sharing a chart of the company's roughly $50 billion stash. So the framing is not capitulation. The framing is flexibility. But it's a real shift. For years the playbook was simple: issue equity and convertible debt, buy bitcoin, never sell. That worked beautifully when the stock traded at a premium to net asset value. It works less well when the premium compresses and existing shareholders start pushing back on dilution. And they have been pushing back. On June 22, Strategy sold $335 million of its own stock, kept about $300 million in cash to bring its USD reserve to $1.4 billion, and bought just 520 bitcoin with what was left. That's a notable departure from the old all-in model, and it tells you something about where capital markets sentiment is right now. The broader bitcoin treasury company trade is under pressure. Investors who bought in expecting pure leveraged bitcoin exposure are realizing that dilution to keep buying only works if the market rewards the buying. When it doesn't, you end up with a capital structure that needs to be actively managed, not just expanded. Strategy's new framework is an acknowledgment of that reality. The bitcoin holdings stay intact for now, but the company has given itself optionality it didn't have before. Whether that reassures shareholders or worries bitcoin holders depends on which side of the trade you're on.
Spot bitcoin ETFs are having their worst month on record. Investors have pulled about $4 billion out of US-listed spot bitcoin ETFs in June. The week ending June 26 alone saw $1.79 billion in net redemptions, the second-largest weekly outflow since these products launched. BlackRock's IBIT accounted for roughly 75% of that, about $1.3 billion gone in a single week, with the heaviest exits on June 20 and 21. For 18 months the narrative was that ETFs would be the steady institutional on-ramp absorbing supply. That narrative is on hold. Thirty-day outflows have hit $6.4 billion, the largest stretch on record. The wrapper that was supposed to be a demand floor is currently acting as a sell wall. Underneath that, miners are getting squeezed hard. JPMorgan estimates the all-in production cost for a bitcoin is around $78,000. With BTC near $60,000, that means roughly 20% of miners are unprofitable, and the average miner is about 17% underwater. Difficulty has dropped more than 20% from its November 2025 peak, the largest peak-to-trough decline since China banned mining in 2021. Hashrate is down about 23% from October highs. Public miners collectively sold a record 32,000 BTC in the first quarter of this year to cover operating costs, more than they sold in all of 2025 combined. The June 26 difficulty adjustment ticked back up 7.15%, suggesting some stabilization, but hashprice is still at $28.68 per petahash per day, down 18% in 30 days. Margins are razor thin. The network is doing exactly what it's designed to do, adjusting difficulty downward as weak hands shut off rigs. The survivors get cheaper blocks. Meanwhile, Tether placed a new order with Canaan for custom immersion-cooled hardware, deepening its push into bespoke mining infrastructure. Several public miners are pivoting toward AI and high-performance computing workloads to diversify revenue, though VanEck estimates a $50 billion near-term funding gap for those AI buildouts. So we have ETF outflows draining demand, miners selling to stay alive, and a price stuck in technical no man's land. Samson Mow says the bottom is in. JPMorgan and several technical analysts disagree and point to $45,000 as a more historically consistent capitulation level. Take your pick.
While markets bleed, AI agents are quietly being wired into actual financial infrastructure. Ondo Finance, Virtuals Protocol, and Treasures opened up more than 430 tokenized US stocks to over 40,000 autonomous AI agents. Ondo provides the tokenized equities, Treasures handles execution, and any agent that can sign a transaction can now trade Apple, Nvidia, or Tesla on-chain. The tokenized stock segment has grown roughly 360% year over year, sitting at about $1.5 billion in on-chain value, with Ondo holding more than 57% market share. Dukascopy Bank in Switzerland became the first Swiss bank to launch AI-powered trading via Anthropic's Model Context Protocol. Clients can connect ChatGPT or Claude directly to their JForex accounts and place orders in natural language, with stop-losses, position sizing, exposure checks, the whole workflow. Live trading support is expected by autumn. On the hedge fund side, boutique macro shops are using LLMs to replace entire analyst teams. Palinuro Capital uses LLMs to parse central bank speeches across languages. Alpha Curve Investments used Claude to model inflation and moved 70% of a 100 million pound portfolio into short-dated inflation-linked bonds. A 15-person Dutch bond fund is compressing research that used to take weeks into days. The pitch is that AI flattens the analyst-team advantage that big macro funds have spent decades building. And in private credit, a stealth startup called Antal just emerged with an autonomous agent stack already originating over $30 million per month for lenders, without adding headcount. Humans still make the final credit decisions, but the entire workflow between decisions, intake, term sheets, document collection, third-party checks, runs on agents. None of this is theoretical anymore. Mastercard has Agent Pay for Machines. Robinhood is rolling out agentic trading. Coinbase is building APIs for agents to hold accounts. The plumbing is being laid right now. House Democrats have already asked the SEC about oversight, with concerns that agents trained on similar data could amplify volatility in correlated ways. That's a legitimate worry. When 40,000 agents see the same signal at the same millisecond, the feedback loop gets interesting. And the BIS used its annual report this week to warn that AI investment is being financed with enormous debt and highly leveraged nonbank structures that could unwind rapidly. So the capital risk and the market structure risk are stacking up at the same time agentic finance goes live.
If a frontier model can write a working operating system kernel in 38 minutes and find a 17-year-old vulnerability in production software, the bottleneck for both attack and defense is no longer talent. It's deployment policy. Watch who gets access, and on what terms. That's where the real game is now.