Anthropic dropped Claude Opus 5, cheaper and sharper on coding benchmarks, while a Coldcard bug quietly drained more than 1,000 Bitcoin from nearly 1,200 wallets. The Fed held rates at 3.5 to 3.75%, Bitcoin is clinging to the low 63,000s, and Strategy just booked an 8.2 billion dollar paper loss on its BTC stack. AI agents are moving onto trading desks, and Wall Street has an August 14 deadline that will expose exactly which ETF holders are quietly heading for the exits. Let's get into it.
Anthropic released Claude Opus 5 this week, and the pitch is not about topping a leaderboard for a single afternoon. It's about being very good every day at roughly half the price of the flagship. Same token pricing as Opus 4.8, 5 dollars in and 25 dollars out per million tokens, but a jump on Frontier-Bench from 18.7% to 43.3%, plus state-of-the-art numbers on ARC-AGI 3 and OSWorld 2.0. It ships with a 1 million token context window and adjustable effort settings, from low all the way up to max, so you can dial reasoning depth against latency and cost.
The interesting bit isn't the benchmark. It's the economic argument. Anthropic is basically saying: the real cost of enterprise AI isn't tokens, it's the human review layer sitting on top. If Opus 5 fails less often on multi-step work, large migrations, long evidence packs, agentic coding, then you can automate workloads that were previously borderline. That's the wedge.
Alongside the model, Anthropic shipped MCP 2026-07-28, the fifth spec of the Model Context Protocol. The headline change: the core is now stateless. It moves from a bidirectional stateful design to request-response, which means MCP servers can now live on serverless and edge infrastructure and scale horizontally like any other HTTP service. Authorization was hardened to line up with OAuth 2.0 and OIDC, so plugging Claude into Entra or Okta is now a real enterprise workflow instead of a hack. Claude's connectors directory is already over 950 MCP servers, and they added MCP tunnels in preview, which lets Claude reach internal servers with no public endpoints and no inbound firewall rules.
And then there's Cognizant, which just became a Global Premier Partner in the Claude Partner Network. Over 30,000 of their associates have completed Claude training, and they're pointing to concrete numbers: a contract-intelligence agent cutting review time by up to 40% in life sciences, an underwriting tool saving about 8 hours a week per underwriter in insurance. This is the boring, unsexy part of AI adoption, the enterprise plumbing, but it's where the actual revenue lives.
AI agents are quietly moving from demo to production on Wall Street, and this week gave us a good cross-section of what that actually looks like.
KelAI just raised 5 million dollars in a seed led by Frst, with Y Combinator and Robinhood Ventures alongside. Their pitch: an autonomous research platform that generates signals, backtests them, validates them, and monitors them with minimal human input. Cohesion, another YC company, is already being used by more than ten long-short and long-only funds managing roughly 10 billion dollars. Their agents chew through earnings calls, filings, podcasts, and X chatter, and surface context when a stock moves. They're explicit that they're not picking stocks, they're being a teammate.
AIBNKO came out of stealth as a fully autonomous multi-agentic sell-side research platform, generating institutional-grade equity reports on demand for any stock in any market. And Opetek launched ARIUS, an AI reasoning platform now in live pilot at a Tier One bank for FX options, with more rollouts in New York, London, and Singapore. Their angle is auditability. Every output comes with the calculations, assumptions, code, and market data it used.
CNBC has a good piece on where this is heading. Robinhood and Public are already letting AI agents connect to accounts. Some estimates suggest the majority of trades on certain retail platforms could be executed by agents within a year. The hard problem isn't the trading. It's alignment. Tell an agent to grow your portfolio aggressively and you may not like the answer. Every serious player is layering approval gates and guardrails on top.
Here's the Bitcoin-adjacent read. If agents dominate order flow, funding rates and derivatives markets get more mechanical and more reflexive. And on that note, Bybit just started accepting tokenized Nvidia, Apple, and Tesla shares as collateral for margin loans. Traditional equities becoming collateral on a crypto exchange while AI agents run trades on both sides. The lines between these markets are getting thinner every week.
This is the story of the week if you care about Bitcoin security. Galaxy Research now estimates the Coldcard incident took roughly 1,082 BTC from 1,196 addresses, worth about 70 million dollars. And here's the sharp part: the attack never touched the devices. No physical access, no malware, no phishing.
What happened is weak seed generation. On certain Coldcard firmware, the entropy that produced private keys was predictable enough that an attacker could recreate likely keys offline, sweep the funds, and just keep searching. The whole draining event happened inside a 41-minute window. That's the terrifying part. If your seed was in that vulnerable set, you didn't get phished, you didn't get hacked, you just got mathematically enumerated.
CZ came out and said the obvious: hardware wallets can still have bugs, diversify across multiple wallets. Fine advice, but it dodges the deeper point. Self-custody assumes your device generated real randomness. When it doesn't, and you have no way to check, the whole trust model collapses in silence.
CoinDesk's angle is that this may push some investors toward ETFs. And that's a legitimate reaction, but it swaps one set of risks, key management, for another, counterparty and custodial. Neither is free.
The practical takeaway if you self-custody: multi-sig with hardware from different vendors is not paranoia anymore, it's the baseline. A single vendor, single seed setup is one firmware bug away from being empty. If you generated a seed on an affected device, move funds now, and generate the new seed with verifiable dice-based entropy or a multi-sig scheme where no single device's randomness is a point of failure.
And worth noting: IBM's CEO said this week that quantum computing will start generating meaningful revenue before the end of the decade. That clock is running too. Bitcoin's cryptography isn't broken today, but seed-generation and signature schemes are exactly the layer that needs to keep evolving.
The Fed held rates at 3.5 to 3.75% for the fifth straight meeting, 9 to 3 vote, and this time three dissenters actually pushed for a hike. First hawkish dissent like that since 2016. Powell called it watchful thinking, not a pause, which is Fed-speak for we're not cutting anytime soon.
Bitcoin briefly popped above 64,000, then drifted back to around 63,000, and slipped to two-week lows as US stocks failed to follow Asia's bounce. The 30-year Treasury yield is above 5.20%, TIPS data is showing rising real yields rather than inflation, and that combination is quietly brutal for non-yielding assets like Bitcoin. Deribit just settled around 9.6 billion dollars in monthly options. If BTC breaks 62,000 over the weekend, there's a 1.1 billion dollar short overhang pointing toward 60,000.
Strategy reported Q2 results and the number is eye-catching: an 8.2 billion dollar net loss, driven almost entirely by an 8.32 billion dollar unrealized markdown on Bitcoin under fair-value accounting. But look past the headline. Holdings are up to about 843,775 BTC, an 11% increase in the quarter, average cost around 75,476 dollars per coin. They raised 17.06 billion dollars this year through at-the-market equity, repurchased 1.5 billion of convertible notes at an 8% discount, and built a 3.75 billion dollar USD reserve, enough for more than two years of preferred dividends and interest.
The pivot worth watching: Strategy sold about 218 million dollars of BTC through its new Bitcoin Monetization Program to fund preferred dividends. That's a shift from pure buy-and-hold. They also bumped the STRC dividend to 12% to keep that preferred trading near 100 dollars, and quietly bought back about 25 million of STRC at a 13.47% discount to par.
And then there's August 14. Bloomberg Intelligence puts the average cost basis of US spot Bitcoin ETF capital around 82,249 dollars, meaning the average position is about 22% underwater, roughly 16.33 billion in unrealized losses. On August 14, large managers must disclose their June 30 ETF positions. That's the day we find out who's actually holding, and who quietly walked out the back door. July ETF flows netted positive at 172 million dollars, but year to date they're still 5.3 billion in the red. The tell is coming.
One prediction to close on: the August 14 ETF disclosures are going to matter more than the next CPI print. Positioning tells you who believes their own thesis when their P&L is red.