Bitcoin is clawing its way back above 63,000 dollars after a brutal weekend that briefly took it under 60,000. Strategy is buying again, picking up 1,550 BTC for about 101 million dollars, while El Salvador quietly added another 8 coins to its national stash. Meta's expensive AI reset is hitting turbulence, with its Muse Spark API delayed for the third time. Anthropic just admitted that 80% of its production code is now written by Claude itself. And a hardware flaw in Trezor's Safe 7 chip is making headlines, though funds appear safe. Let's get into it.
Meta has officially abandoned the open-source playbook that made Llama a household name in AI circles. The company that once championed free, downloadable foundation models is now running a closed, proprietary strategy under a 28-year-old chief AI officer, Alexandr Wang, brought in from Scale AI. The flagship is called Muse Spark, launched in April. It's natively multimodal, embedded inside WhatsApp and Instagram, accessible only through a limited API. And here's the thing — that API has been delayed three times. From April to May, May to June, and now Meta just says "this month" with no firm date. Wang himself called Muse Spark an "appetizer" at the Bloomberg Tech Summit. Not exactly the confident framing you want when your company is burning 145 billion dollars in AI capital expenditure this year. On independent benchmarks, Muse Spark trails GPT-5.4 Pro and Gemini 3.1 Pro. It's competitive on some tasks, behind on others. The earlier Llama 4 Behemoth model? Cancelled outright after failing to hit internal capability targets. The strategic shift matters. Meta's old pitch was: we give away the model, you build on top, the ecosystem grows. The new pitch is: we control access, we charge for the API, we monetize through subscriptions and partnerships. That's a very different bet. It assumes Meta can actually out-execute OpenAI and Google at their own game, with proprietary models behind paid endpoints. And so far, the evidence is mixed. The repeated API delays suggest infrastructure problems, not just polish. When you've told the market you're spending nine figures a quarter to win this race, three slipped deadlines start to look like an execution problem. The next few months are the real test. Either Muse Spark's successor lands hard and closes the gap, or Meta becomes the cautionary tale of how much money you can spend chasing a moving frontier.
Anthropic dropped a number this week that should make every software engineer pay attention: 80% of their new production code is now authored by Claude. Not assisted by Claude. Authored by it. The result is an 8x increase in code shipped per engineer per quarter compared to their 2021 to 2025 baseline. In one internal example from April, Claude autonomously fixed more than 800 API errors and cut the related error rate by 1,000x. This isn't an isolated bet. Gartner's 2026 Magic Quadrant for AI coding agents just put Anthropic, Cursor, GitHub, and OpenAI in the leaders quadrant — and pushed the big cloud providers down to challengers. The category Gartner is naming is "agentic software engineering": autonomous agents that perceive context, plan multi-step actions, execute, verify, and iterate. Plan, act, verify, loop. Dropbox just unveiled Nova, an internal platform for running coding agents at scale inside their monorepo. Their most successful deployments aren't writing new features — they're handling operational debt. A tool called Deflaker automatically investigates and repairs flaky tests, iterating through CI until things pass. LG CNS in Korea launched AIND, an agentic platform that takes natural language requirements and produces working enterprise systems. And Cadence, with NVIDIA, announced what they're calling a Level-5 autonomous chip design engineer that compresses five-week verification loops down to under one day. Gartner expects more than 70% of enterprise software engineers to rely on AI for coding tasks by 2028. The implication for how you build a team is significant. The bottleneck shifts from writing code to reviewing it, verifying it, and governing what gets shipped. Anthropic explicitly warns about "alignment cascades" — small misalignments stacking up across agent sessions if you don't have verification gates. The job isn't going away. It's mutating into something closer to running a factory than crafting a product.
Bitcoin is back above 63,000 dollars after dipping below 60,000, and the custody conversation is getting more interesting in two directions at once. First, the bad news. Ledger Donjon, the security lab inside Ledger, found a hardware flaw in the TROPIC01 secure element chip inside Trezor's brand-new Safe 7 wallet. The chip is made by Tropic Square, a Trezor sister company. The flaw was found via laser fault injection in a controlled lab. It cannot be fixed by firmware update. That sounds alarming, but the reality is more nuanced. The Safe 7 uses three independent security layers — TROPIC01 is one of them, alongside an Infineon OPTIGA Trust M and the STM32U5. Compromising TROPIC01 alone doesn't expose your PIN or your funds. It requires physical access, lab equipment, and disassembly. No real-world exploitation has been reported. What's actually notable here is the disclosure: a direct competitor found the flaw, the chip vendor confirmed it, and Trezor published it openly. That kind of coordinated transparency is rare in hardware security. On the more constructive side, Nunchuk released two open-source tools that let AI agents interact with Bitcoin wallets under what they call "bounded authority." The idea is that as AI agents start moving money — and they already are, MetaMask just launched an AI agent wallet this week — you don't want a model with full key control. Nunchuk uses multisig wallets with spending policies, limits, delays, and human approvals required beyond set thresholds. The agent can execute small routine transactions; humans sign off on anything bigger. It's the same logic behind Bitkey's wrench-attack defenses we've heard discussed lately: time-locked withdrawals, server-key policy enforcement, on-device verification. The throughline across all of this is that single-key, hot-wallet thinking is increasingly obsolete. Whether the threat is a coercive attacker, a curious lab researcher, or an autonomous agent with access to your seed phrase, the answer is the same — distribute authority, add delays, and verify on-device.
While Wall Street panics about Bitcoin dipping below 60,000 dollars, El Salvador just keeps buying. The country added 8 BTC over the past week, bringing total holdings to 7,633 coins, worth roughly 624 million dollars at current prices. Over the last 30 days, they've added 31 coins. The strategy is dead simple — roughly one Bitcoin a day, no market timing, just steady accumulation. What makes this interesting is the context. El Salvador agreed to soften its Bitcoin laws as part of an IMF deal, making acceptance optional for businesses. The IMF has publicly suggested the ongoing purchases might just be rebalancing existing holdings rather than new buys. El Salvador disputes that. Either way, Bukele is using the dip as a political statement: when the price falls, the state buys more. And he's not alone anymore. By some counts, roughly 23 countries are now pursuing some form of sovereign Bitcoin strategy. Brazil, the Czech Republic, Luxembourg, Saudi Arabia, and Taiwan are among the newer names being floated. Whether all of those become formal reserve programs is another question — a lot of this is exploratory or political signaling rather than a treasury allocation. But the direction of travel is clear. Meanwhile in the corporate world, Strategy resumed buying after a controversial 32 BTC sale last week — adding 1,550 coins for about 101 million dollars and bringing their stack to 845,256 BTC. Strive, a smaller Bitcoin treasury company, bought exactly 32 coins — a pointed mirror of what Strategy had just sold. And Morgan Stanley quietly opened a door that matters more than the price action: their wealth management clients can now lend Bitcoin, Ethereum, or Solana to Galaxy Digital and receive shares of spot crypto ETPs in return. That's Bitcoin as institutional lending collateral, inside a major wirehouse, with onboarding timelines that used to take months now compressed dramatically. The price is volatile. The infrastructure being built around it is not.
One observation to close on. Meta is spending 145 billion dollars to chase a closed AI model nobody can download. El Salvador is spending a few hundred thousand a week to buy an open monetary network nobody controls. Decade from now, ask which one looked like the better capital allocation.