Anthropic loosened the leash on Claude Fable 5's biology guardrails, cutting refusals by about 85% — but researchers who were promised trusted access are still waiting. Bitcoin's most controversial soft fork attempt in years, BIP-110, hit its mandatory signaling window with less than 3% miner support and promptly stalled after two blocks. Strategy sold another 1,638 Bitcoin to fund dividends and buybacks, pushing its USD reserve to 4 billion dollars. And spot Bitcoin ETFs pulled in 853 million dollars last week, their best showing since April. Let's get into it.
Anthropic pushed a big update to Claude Fable 5 on August 7th, and it's a governance story dressed up as a product update.
Since Fable 5 launched on June 9th, the model has been aggressively conservative on anything biology-related. Ask about your blood test, ask about a symptom, ask basic cellular biology questions — and Fable 5 would often fall back to a less capable model. Anthropic uses classifiers to route sensitive queries to weaker models when safeguards fire, and at launch they admitted biology was tuned for high false positives so they could ship fast.
Now they've retrained the classifier. Result: fallback rates dropped by roughly 85% in testing. On Claude.ai, 67% fewer fallbacks. On Cowork, 55%. Claude Code saw 17%. The Claude Platform, 7%. Everyday users — patients, students, clinicians doing routine clinical support — get direct answers now on things like interpreting lab results and disease symptoms.
But here's the catch. Frontier biology — virology, toxicology, molecular design, drug development — is still blocked. Dual-use requests still get routed to Claude Opus 5 with tighter constraints. Which means the people who most need a frontier model — actual biologists, virologists, drug developers — still can't use it for their real work.
Anthropic promised a trusted-access pathway for vetted researchers within weeks of the June launch. It's August. No timeline, no pathway, no update.
Meanwhile, OpenAI is telegraphing the opposite posture with its upcoming Astra model. They're publicly acknowledging Astra may have critical cyber capabilities, and they're pre-committing to isolated testing, stronger weight protections, sandboxed execution, and pausing testing where controls are absent. Framing themselves as the more disciplined lab while Anthropic loosens Fable's leash.
The honest read: both companies are competing for enterprise trust while China and open-weight models pressure everyone to ship. Anthropic is betting measurable safety metrics — public fallback rate numbers — buy them cover to expand access. OpenAI is betting process theater does the same. Neither has solved the underlying problem, which is that the researchers who could actually stress-test these models are locked out.
One small but interesting side note from the same release cycle: Claude Code 2.1.224 added cross-session messaging on macOS and Linux. Two coding sessions on the same machine can now send each other plain-text summaries through new ListAgents and SendMessage tools. Messages stay local, can't approve permissions or change configs, and it's genuinely useful for multi-worktree workflows or polling a long-running migration from another terminal. Not on Windows. Small feature, real quality-of-life improvement for people running parallel agents.
Bitcoin just had its most awkward soft fork attempt in recent memory, and it's playing out in real time.
BIP-110 is a proposal to temporarily restrict arbitrary data in Bitcoin transactions — capping output scripts to 34 bytes, limiting witness items, invalidating certain Taproot annexes and Tapscripts. The goal is to reduce the cost that node operators bear when miners let one-time data payloads — inscriptions, tokens, images — clog block space. Supporters call it curbing an externality. Critics call it censorship dressed up as protocol hygiene.
The activation mechanism required 55% miner signaling — 1,109 out of 2,016 blocks — to lock in. As mandatory signaling opened at block 961,632, support was sitting at 2.42%. Not a typo. Two point four two percent.
Michael Saylor publicly told BIP-110 backers to stand down before the window opened, warning that preemptive filtering of transactions by perceived purpose cuts against Bitcoin's neutrality. Adam Back at Blockstream flagged chain-split risk given the tiny threshold of support. The BIP-110 author himself, Dathon Ohm, told users to stop running Bitcoin Core and switch to Bitcoin Knots, warning that Core users could end up validating invalid blocks under the new rules.
And then, on August 9th, the split actually happened. Enforcing nodes rejected non-signaling blocks and forked off. The BIP-110 branch mined two blocks — heights 961,632 and 961,633 — and then stalled. None of the first 59 blocks on the dominant chain signaled. The enforcing chain is now roughly 57 blocks behind and inheriting Bitcoin's full mining difficulty with a rounding-error share of hashpower, meaning blocks are hours apart if they come at all.
Exchanges kept operating normally. Coinbase and Kraken status pages didn't blink. For anyone not running a BIP-110-enforcing node, nothing happened.
A few takeaways. First, Bitcoin's governance actually worked here — a controversial proposal without economic majority support failed to activate, and it failed loudly and visibly. Second, the aggressive activation path — trying to force the issue with a low threshold and mandatory signaling — is exactly what Saylor warned against, and it produced the chain split risk critics predicted. Third, the underlying question about block-space externalities and data storage on Bitcoin doesn't go away. It just gets pushed to the next fight. And there will be a next fight.
Strategy's 8-K dropped this week and it confirms what the market has been trying to price for months: the never-sell era is over, and the new model has visible seams.
Between July 27th and August 2nd, Strategy sold 1,638 Bitcoin at an average price of $63,957, netting $104.7 million. Roughly $52.4 million funded STRC preferred stock dividends. Another $52.3 million funded STRC share buybacks. The company also sold 3 million MSTR shares for $290.6 million and pushed its USD reserve to 4 billion dollars. Total Bitcoin holdings now sit at 842,138 BTC, about 5,225 coins below the June peak.
The mechanism behind this is Saylor's STRC — the Variable Rate Series A Perpetual Stretch Preferred Stock. The pitch is a perpetual Bitcoin buying machine. Each STRC issuance represents roughly 2.3% of Strategy's Bitcoin stack. Proceeds buy more BTC. Dividends get paid from cash flow — and, when needed, from selling Bitcoin. As long as Bitcoin appreciates more than about 2.3% annually and STRC demand holds, the math works. New BTC bought exceeds BTC sold to service dividends, net stack grows.
The problem is the assumptions. STRC is trading around $89.40, below its $100 par value, which signals financing pressure. Bitcoin was around $62,600 at the time of the filing — not a bear market, but not the kind of appreciation curve that comfortably outruns a 12% dividend obligation. And Strategy hasn't bought Bitcoin since June. The current cadence is sell, pay dividends, buy back stock, hold cash.
Zoom out and it's not just Strategy. CryptoSlate reported that Empery Digital's never-sell treasury model cracked hard, offloading 1,635 BTC and shrinking reserves by 76% in weeks. They're down to 1,279 BTC, with 954 of those pledged against $35 million in debt. Different scale, same pattern: treasury companies built on the assumption that they'd never have to sell are discovering that dividends and debt service don't care about your marketing.
Meanwhile, actual demand for Bitcoin exposure through regulated wrappers is healthy. Spot Bitcoin ETFs pulled in $853 million last week — the strongest week since mid-April. BlackRock's IBIT led the pack, and combined with Ethereum ETFs the total broke a billion dollars, with BlackRock capturing about 80% of that combined billion.
So the split screen: ETF flows say institutions want Bitcoin. Treasury companies with leveraged capital structures are being forced to sell it. Both are true, and they're pulling in opposite directions on liquidity.
There's a quiet buildout happening in AI-driven finance that's worth tracking, because it's the layer beneath every headline about autonomous agents.
QumulusAI, a Nasdaq-listed compute provider, just signed its first monetization deal for reserve NVIDIA Blackwell GPU capacity — with an agentic hedge fund. The structure is interesting: market-rate compute pricing plus a share of the fund's quarterly trading profits, with no exposure to losses. The hedge fund runs fully agentic AI-driven trading 24/7 on sovereign, self-hosted infrastructure. Departure from Qumulus's usual fixed take-or-pay contracts. CEO framed it as necessary because milliseconds matter and you can't run this stuff on someone else's cloud.
Numerai is going further. CEO Richard Craib is assembling a team to apply frontier AI to Numerai's own internal research, risk models, and portfolio construction — building what he calls a self-improving hedge fund. Numerai already crowdsources predictions from thousands of external models. Now they want the internal loop — research to testing to risk analysis to deployment — running on frontier AI. They're managing about 700 million dollars, trading over a billion per month across 30 markets. Job listing pays 300 to 350 thousand for a quantitative research engineer.
On the retail-facing side, MetaMask launched Agent Wallet — a self-custodial wallet built for autonomous AI trading. Users set spending caps, approved protocols, risk settings. Two modes: Guard and Beast. Transaction simulation, MEV protection, gas abstraction so agents don't need native tokens to move funds. Supports Claude Code, Codex, Cursor, and a handful of EVM networks. Optional coverage up to $10,000 per month on eligible transactions.
And Pinegap raised 8 million dollars to build custom AI agents for institutional buy-side research — over 1,000 agents deployed across 100-plus clients, producing 50,000 research reports a month.
Here's the pattern. The AI-in-finance story is no longer chatbots analyzing filings. It's dedicated compute, self-custodial wallets designed for agents, closed learning loops trained on live capital flows, and specialized research automation. The infrastructure is getting built to assume that some meaningful chunk of trading decisions will be made by software with real balance sheet authority.
One related finding worth noting: CryptoSlate flagged research showing AI financial advisers carry a hidden Bitcoin bias that activates on specific prompts. Same client, same finances — ask for a diversified portfolio, get standard allocations. Introduce bank failures and capital controls in the prompt, or describe an economy where autonomous software pays other software, and Bitcoin allocation jumps. Which means as these agentic systems get deployed, the framing of the questions they're asked will shape trillions in flows. Something to watch.
Two systems tested their governance this week. Anthropic loosened Fable 5's guardrails and published the numbers. Bitcoin miners quietly refused to signal for BIP-110 and let a fork die on the vine. One is a company deciding what its product does. The other is a network deciding what it is. Notice which one required no press release.