Bitcoin just cracked below 66,000 dollars, triggering 1.8 billion in liquidations and sending the fear gauge up nearly 20%. Strategy made its first BTC sale since 2022 — just 32 coins, but enough to rattle every treasury company copying its playbook. Anthropic shipped Claude Opus 4.8 with a coding score that beats GPT-5.5 and a new feature that spawns up to 1,000 parallel subagents in a single session. And the legal AI arms race is heating up — Wordsmith just raised 70 million dollars while Harvey hits an 11 billion dollar valuation. Let's get into it.
Anthropic released Claude Opus 4.8 just 42 days after 4.7 — the fastest Opus update cycle yet. And the numbers are sharper than the timeline suggests. On SWE-Bench Pro, it hits 69.2%, up from 64.3%, and beats GPT-5.5's 58.6%. On the Intelligence Index, 61.4 versus 60.2 for GPT-5.5. The USAMO 2026 math score jumped from 69.3% to 96.7%. That's not incremental.
But the more interesting story is honesty. Anthropic claims 4.8 is roughly 4 times less likely than 4.7 to let its own buggy code pass without flagging it. Overconfidence dropped more than tenfold. Testers say it pushes back on bad plans and asks clarifying questions instead of confidently generating broken output. For anyone who's actually shipped code with an AI agent, that's the failure mode that matters.
The headline feature is Dynamic Workflows in Claude Code. You give Claude one prompt for a large task — say, a codebase migration — and it spawns hundreds or up to a thousand parallel subagents, runs them, verifies the output against your criteria, and reports back. Available on Enterprise, Team, and Max plans. This is the kind of thing that used to be a six-month consulting engagement.
Pricing stays at 5 dollars per million input tokens and 25 dollars per million output. There's a new Fast Mode at 10 and 50 dollars, claiming 2.5x speed. And on claude.ai there's now Effort Control — you pick how hard the model should think per response.
Meanwhile Anthropic teased Mythos, their next-generation architecture. Mythos Preview is already in the hands of about 50 partners — Apple, Google, Microsoft, AWS — and through Project Glasswing they've found over 10,000 high or critical-severity vulnerabilities. The catch: these models can autonomously discover zero-days and develop exploits, which is exactly why the White House reportedly raised concerns about expanding access. Anthropic also just closed a 65 billion dollar Series H at a 965 billion post-money valuation, with revenue tracking toward a 30 billion annualized run rate. The agent race isn't slowing down.
Bitcoin fell more than 6% in a day, hitting a 24-hour low of 65,708 dollars. Ether broke below 1,900. About 1.8 billion in crypto positions got liquidated. The Crypto Fear and Greed Index just posted its biggest jump into fear territory since the February 5th crash. Prediction markets on Kalshi now imply a 66% chance BTC trades below 55,000 before year-end, and a coin-flip chance of going under 50,000.
What's driving it? A few overlapping stories. One is geopolitics — fresh US and Iran strikes resumed this week, and the Treasury just sanctioned four Iranian crypto exchanges. There's even a theory floating around that the sell pressure is Iranian sanctions-related, not Strategy-related. Two, the macro divergence is brutal: the MSCI All Country World Index hit a fresh all-time high on the AI rally, AI tokens are ripping, and crypto is the one asset class getting left behind. Bitwise's CIO put it bluntly — crypto has become a contrarian bet while AI stocks dominate attention.
Citi's analysts argue the real problem isn't Strategy's small sale. It's the lack of fresh buyers. New demand has dried up. And that lines up with Bitcoin retesting the February low for the third time. Each retest weakens the floor.
Now, the contrarian read. The Power Law model — which measures how far BTC trades from its long-term trend — shows Bitcoin at one of its deepest discounts in history, a level last seen during the March 2020 crash and the FTX collapse. Both preceded major rebounds. Bitwise also put out a sovereign default-risk model pegging Bitcoin's fair value at 224,000 dollars if debt fears deepen. So you've got max fear in the short term and a structurally bullish setup if you zoom out. Pick your timeframe.
This is where it gets interesting. Strategy sold 32 BTC at around 77,000 dollars — about 2.5 million total — to fund preferred stock distributions. It's a tiny sale relative to their holdings. But it's the first sale since 2022, and it punctured the "never sell" narrative that every imitator was built on.
Metaplanet's stock dropped 8% on the news, closing around 271 yen. They hold 40,177 BTC at an average cost near 15.5 million yen per coin. They're still buying — they just raised about 255 million dollars in a private placement priced 2% above book value, with fixed-strike warrants at a 10% premium. They're also using a novel moving-strike warrant tied to their MnAV ratio, which only becomes exercisable if the stock trades above 1.01 times NAV. The target: 210,000 BTC. So even with shares down, Metaplanet is doubling down on the acquisition flywheel.
Strive bought 2,500 BTC at an average price of about 74,092 dollars, bringing holdings to 19,000 coins. They have zero debt and 137 million in cash. And they just announced plans to expand both their ATM equity programs by 2.1 billion dollars each — bringing total ATM capacity to over 5 billion. That's a lot of potential share issuance if the premium holds.
And that's the whole question. The Strategy model works because of three things: sustained access to capital markets, a meaningful equity premium over NAV, and execution discipline. Strategy has all three. Imitators don't. As the premium-to-NAV compresses across the sector, capital deployment stops being accretive — you're just diluting shareholders without adding meaningful BTC per share. Spot Bitcoin ETFs become a cleaner alternative.
Meanwhile ARK Invest's Bitcoin "Conviction Buyers" cohort grew their holdings 69% in Q1, from 2.13 million BTC to 3.6 million — the highest since 2020, even with BTC down 22% in the same period. So the long-term accumulators are buying. The leveraged corporate treasuries are getting stress-tested. Those are two different trades.
While everyone watches crypto bleed, the enterprise AI buildout keeps compounding. Wordsmith, an Edinburgh-based legal AI platform for in-house legal teams, raised 70 million dollars in Series B led by Highland Europe and Index Ventures. Total funding now 100 million. They serve over 500 in-house teams including BT, Canva, Starling, and Sage. The pitch: receive, route, resolve, and record every legal request inside a company, automating the routine and escalating only when human judgment is required. Every action logged for audit.
The context is a market growing from about 5.2 billion in 2026 to a projected 41 billion by 2034. Harvey just raised 200 million at an 11 billion dollar valuation, but Harvey targets law firms — Wordsmith targets the corporate legal department. Different buyer, different workflow. Then there's Bayshore in Munich, just out of stealth with 6.9 million euros from Earlybird, building an agentic compliance platform that turns regulations into machine-readable guardrails with deterministic logic and full audit trails. Multiple Global 2000 companies already piloting.
Icertis unveiled Vera, an AI-native contracting system with three modules — Copilot for conversational queries, Engage for drafting and redlining, Analytics for portfolio risk. They claim contracting acceleration of over 80%. And Gavel Exec, which lives inside Microsoft Word and reads counterparty redlines directly, is cutting contract review time by up to 60% for regulated financial firms.
The pattern across all of these: governance is becoming the moat. Zip, the 2.2 billion dollar procurement platform, just launched five AI Superagents — one of them a Legal Superagent that redlines contracts against approved playbooks. Their whole pitch is keeping data inside permissioned, audit-trailed environments so your finance team stops pasting NDAs into personal ChatGPT accounts. Zip is also shipping a Model Context Protocol server so Claude or ChatGPT can query procurement data without it leaving the governance perimeter.
This is the real story under the AI hype: the winning enterprise AI products aren't the smartest models. They're the ones with the cleanest audit trail. Compliance officers are the gatekeepers, and they need to prove how every decision was made three years from now.
One thing to watch: Strategy sold 32 coins to make a coupon payment. If a 9.5% weekly drawdown forces that decision at the top of the food chain, every imitator with thinner equity and a compressed NAV premium has a much harder math problem coming. The treasury model gets tested in the drawdown, not the rally.