Anthropic drops Claude Sonnet 5 with a lower price and a heavy pitch on autonomous agents. Bitcoin slides back under $63,000 as Trump declares the Iran ceasefire over and oil jumps 5%. A Prague startup founded by ex-DeepMind poker researchers hits a $500 million valuation running reinforcement learning agents live on the S&P 500. And 20% of Bitcoin miners are now operating at a loss, hitting a stress level analysts call historically rare. Four stories, one thread: automation is getting cheaper, and the humans running the machines are getting squeezed.
Anthropic had a big week. The headline is Claude Sonnet 5, pitched as their most agentic model yet, at a lower price than previous Sonnet releases. The claim is that it can drive browsers, terminals, and multi-step workflows on its own, capabilities that used to belong to the top-tier expensive models. Anthropic is calling this capability compression, meaning frontier behavior in a mid-tier price class.
The catch: there are no published benchmarks yet. No SWE-bench, no GPQA, no independent verification. So take the autonomy claims with some skepticism until developers actually stress-test it in production. The pricing angle is what matters strategically. If Sonnet 5 really can run browser and terminal automation reliably at mid-tier cost, that reshapes the economics of every AI agent startup building on top of these APIs.
Alongside the model, Anthropic pushed Claude Cowork to web and mobile. Cowork is their agentic work product, and the interesting stat they shared is that over 90% of Cowork activity is not coding. It's spreadsheet reconciliation, contract tracking, turning transcripts into client decks. Regular knowledge work. Now that work can start on desktop and finish on your phone, and tasks can run in the background on a schedule.
And they announced Claude Code and Cowork for the US government in a FedRAMP High environment, with tamper-evident audit logs, two-person approvals for sensitive operations, and on-device conversation storage. That's a real enterprise wedge into federal agencies.
One research note worth flagging. Anthropic published work identifying what they call J-space inside Claude, a small set of internal representations that behaves like a global workspace from cognitive neuroscience. It's under 10% of the model's activity, but it's where multi-step reasoning seems to converge. Interesting if you care about interpretability. Less interesting if you just want to know whether the agent will book your meeting correctly.
Now to a story that sits right at the intersection of AI and finance. EquiLibre Technologies, based in Prague, just raised a Series A at a valuation north of $500 million, led by Creandum. It's Creandum's largest single check ever.
The founders are three former DeepMind researchers, Martin Schmid, Rudolf Kadlec, and Matej Moravcik. Their previous claim to fame was DeepStack, the AI that beat professional players at no-limit poker. They've now pointed the same reinforcement learning approach at financial markets.
The company trains RL agents on historical and live market data, and those agents are trading billions of dollars a day on the S&P 500 and Nasdaq through a partnership with Tower Research Capital. They also started in crypto in 2025 before moving into equities. The pitch to investors includes a claim of zero losing months since launch, though there's no independently disclosed performance data.
Most of the new money is going to compute. They want to build one of the largest AI training clusters in Central and Eastern Europe.
A couple things worth chewing on here. First, the frame. EquiLibre isn't selling a product to end users. They're building autonomous agents that make decisions inside a regulated, high-stakes environment. That's a very different bet than a chatbot company. It's closer to what a hedge fund actually is, wrapped in modern AI infrastructure.
Second, the claim of being first to run reinforcement learning live on US markets is marketing. Jane Street, Citadel, Two Sigma, Renaissance, they've all been doing versions of this for years. What EquiLibre may have is a more efficient research-driven approach coming out of the game theory tradition.
And third, the broader signal. Elite AI research talent is increasingly flowing from academic labs and big tech into quant finance. That's where the compute budgets, the clear feedback loops, and the immediate profit motive line up. Expect more of these labs, and expect more of them to be quiet.
On to Bitcoin mining, where the picture is genuinely rough. Bitcoin is trading around $63,000, and about 20% of miners are now unprofitable. Analysts are calling the stress level historically rare, comparable to the bottoms of 2015, 2018, and 2020.
The Puell Multiple is around 0.74, roughly 25% below its 12-month average. Miner revenue is down about 11% in the last 10 days. Hashrate is off more than 25% since October 2025, and the latest difficulty adjustment dropped roughly 10%, which is the network mechanically rebalancing as unprofitable machines shut off.
JPMorgan's estimated average production cost is around $78,000 per coin. That means a substantial slice of the network has been mining at a loss for roughly five months. The split between fleets is stark. Newer, efficient rigs in the range of 10 joules per terahash are earning roughly $81 per megawatt-hour. Older gear is earning about $43. If hashprice stays weak, the older sites either curtail, sell BTC to cover liquidity, or get absorbed by stronger operators with cheap power.
Canaan's Q1 results tell the same story from the ASIC maker side. Revenue collapsed to $62.7 million from $196 million the prior quarter. Net loss of $88.7 million. ASIC sales fell from $165 million to $43 million quarter over quarter. They're guiding Q2 revenue between $35 and $45 million. At the same time, their own treasury grew to roughly 1,826 BTC and 3,952 ETH, so they're increasingly a hybrid: hardware seller, self-miner, and crypto treasury holder rolled into one.
On the technology side, Bitdeer announced the Sealminer A4 Ultra Hydro, rated at 1 petahash per second per unit at around 10 joules per terahash. To put that in perspective, hitting 1 petahash in 2015 took about 1,000 machines. In 2017, 71. In 2021, 10. Now, one.
And there's a longer-term deadline. The US Energy Information Administration projects grid electricity demand rising through 2027, driven by AI data centers and crypto. Miners have until roughly 2027 to prove they can be flexible load partners rather than a drain. The ones who survive this cycle will be the ones with the cheapest power, the newest gear, and enough balance sheet to wait out difficulty relief.
Finally, sovereign Bitcoin. Coinbase executive John D'Agostino claims more than 40 countries have expressed intent to add Bitcoin to sovereign reserves. That's the headline number making the rounds. The reality is more muted. Expressed intent is not the same as signed purchases. No country list, no volumes, no timelines. Publicly, El Salvador and Bhutan are the confirmed active holders. Around 13 countries have known Bitcoin holdings of some kind, mostly from seizures.
And the flagship case, the US Strategic Bitcoin Reserve, is still stuck in bureaucratic limbo 16 months after Trump's executive order. The Treasury and Commerce departments are actively fighting over who gets to run it. The Justice Department is mediating. There's a legal question about whether Treasury even has statutory authority to custody a volatile non-yielding asset under a no-sale mandate.
The reserve theoretically holds about 328,000 BTC seized through law enforcement, roughly $21 billion at current prices. That makes the US the largest known sovereign holder on paper. But no formal governance, no acquisition mechanism, no congressional legislation passed. The BITCOIN Act and the American Reserve Modernization Act both exist as bills. Neither has advanced.
Meanwhile India's central bank is still pushing for outright crypto prohibition, citing tax evasion. Fewer than a quarter of the 645,000 Indians who transacted in crypto reported it on their tax returns.
So the sovereign adoption narrative is real but uneven. Kazakhstan just signed a decree accelerating crypto adoption with tax exemptions and stablecoin corridors. Japanese corporates are diversifying into Bitcoin as the yen collapses. Vanguard, which two years ago refused to list a spot Bitcoin ETF, is now hiring a Head of Digital Assets to run strategy across their wealth business, which touches roughly 50 million investors.
The pattern isn't 40 countries stacking sats tomorrow. It's slower: pilot programs, quiet acquisitions, legislation that stalls and restarts, corporate treasuries moving before governments do. The US case shows that even with a presidential order, actually operationalizing a national Bitcoin reserve is harder than announcing one.
Two truths from today. Autonomous agents just got cheaper, faster, and closer to real work, whether that's Claude clicking through your calendar or reinforcement learning agents trading real capital on the Nasdaq. And the people running the physical infrastructure that Bitcoin actually depends on are getting crushed. If you're building anything in this space, ask yourself which side of that gap your business sits on.