Bitcoin is closing out August as the best-performing major asset of the month, up 24%, shrugging off fresh U.S. strikes on Iran that pushed oil higher and equities lower. Strategy is back in buying mode, adding $370 million in BTC last week, its first purchase in about two months. OpenAI dropped GPT-5.5, its first fully retrained base model since GPT-4.5, aimed squarely at agentic work. Google Cloud launched Gemini Enterprise for Legal, and Greenberg Traurig rolled out Thomson Reuters CoCounsel to all 3,200 of its lawyers. And StarkWare pulled off the first quantum-resistant Bitcoin transaction on mainnet, no fork required. Let's get into it.
OpenAI had a busy week, and the through-line is enterprise. Start with GPT-5.5, codenamed Spud. It's the first fully retrained base model since GPT-4.5, rolling out to Plus, Pro, Business, and Enterprise inside ChatGPT and Codex. API access is delayed, coming after more safety and scale testing. The pitch is autonomy. This model is built to plan, use tools, check its own work, and grind through messy multi-step tasks without you holding its hand. The benchmarks back it up: 82.7% on Terminal-Bench 2.0, 84.9% on GDPval, 78.7% on OSWorld-Verified, 98% on Tau2-bench Telecom. And OpenAI claims it uses fewer tokens than GPT-5.4 to get there, so even at a higher per-token price the total cost per task drops for a lot of workflows.
On pricing, OpenAI is also running a three-month promotion on GPT-5.6 Sol. From August 21 through at least November 21, input tokens drop from $5 to $4 per million, a 20% cut, and output tokens drop from $30 to $20 per million, a 33% cut. Cached input sits at $0.40 per million. It's temporary, and post-November rates aren't defined, but it tells you where the competitive pressure is.
The more interesting release, honestly, is the Admin plugin for ChatGPT Work and Codex. Workspace admins can now review usage, manage members and groups, check permissions, and approve or deny spending requests directly inside a ChatGPT conversation. No console navigation. It runs within each admin's existing role, doesn't grant broader access, and in Enterprise and Edu workspaces it's disabled by default until an admin opts in. This is small but significant. It's OpenAI moving the assistant from answering questions to actually executing bounded IT operations with audit trails. That's the real agentic frontier: not writing poems, but changing entitlements with a proper approval chain.
Legal is turning out to be one of the fastest enterprise AI verticals, and two announcements this week make that clear. Google Cloud launched Gemini Enterprise for Legal, a stack of specialized AI agents with skills like citation verification, contract lifecycle management, brief drafting, data subject access request execution, and regulatory monitoring. Early adopters include Cleary Gottlieb, Weil Gotshal, Freshfields, and Williams and Connolly. These are not sleepy firms experimenting on the margins.
At the same time, Greenberg Traurig rolled out the next generation of Thomson Reuters' CoCounsel Legal to its entire 3,200-lawyer global workforce across 51 offices. Not a pilot, not a proof of concept, full deployment. CoCounsel now plans, researches, reasons, and drafts matter-wide work with citations grounded in Westlaw and Practical Law. It integrates with DeepJudge, iManage, NetDocuments, SharePoint, HighQ, Box, and the full Microsoft stack, so it lives inside the tools lawyers already use. Thomson Reuters describes the target capability level as a senior associate: take a plain-language request and produce a researched, cited, polished work product.
The governance piece is worth flagging. AI use is disclosed to clients, outputs are reviewed by matter lawyers and supervising shareholders, and there are hard controls preventing training on client data. That's the pattern you're going to see everywhere: agentic execution paired with mandatory human review and auditable trails. The strategic read is straightforward. Big Law is moving from AI pilots to production, and the moat is no longer just the model. It's the integration with authoritative content like Westlaw plus the internal knowledge graph. If your product is a wrapper around a general model with no grounding, that's a shrinking niche.
Two big developments on Bitcoin's post-quantum roadmap this week, coming from very different directions. First, Blockstream researchers proposed SHRINCS, short for Shrunken SPHINCS. It's a post-quantum signature scheme built on SHA-256, designed specifically to not blow up Bitcoin's block space. Public keys are 48 bytes, the smallest stateful signature is 548 bytes, and the stateless fallback is 5,776 bytes. In theory it preserves roughly 3 transactions per second, versus about 0.36 tps if Bitcoin adopted the standard NIST hash-based scheme. That's an order of magnitude difference in throughput, which matters a lot when you're talking about a network-wide migration. The catch: each signature spends a one-time key, so users have to track which keys have been used across devices and backups. The security proof is pending, the reference software is unaudited, and adoption would require a soft fork. Blockstream has demonstrated SHRINCS-signed transactions on Liquid.
The second story is more provocative. StarkWare researcher Avihu Levy executed the first known quantum-resistant Bitcoin transaction on mainnet, 3.1 BTC, without any protocol change. The technique is called Quantum Safe Bitcoin, and it uses signature grinding plus a proof-of-work puzzle wrapping to close the mempool exposure window between when a transaction is signed and when it's confirmed. That's the vulnerability window a quantum attacker would exploit. Because the transaction uses a non-standard format, regular Bitcoin nodes won't relay it, so it went through MARA's private Slipstream mempool. It builds on Binohash from BitVM and fits within Bitcoin's script limits. Isabel Foxen Duke at the MARA Foundation was clear: private mempools are not a long-term solution, they're a break-glass option. Levy himself supports a soft fork. But it proves the point that a stopgap exists today, which shifts the debate. The hard-fork versus soft-fork camp just got a third option: a user-elected, on-chain workaround while the ecosystem figures out consensus.
Corporate Bitcoin treasuries had a mixed and slightly weird week. Strategy is the headline. After a two-month pause, Michael Saylor's company added roughly $370 million in BTC last week, its first buy since late June. Saylor's We're Back post confirmed what the market suspected. Bitcoin dominance climbed above 60% and Strategy's valuation expanded on the news. Bitcoin is trading near $78,000.
On the smaller end, Capital B, the Paris-listed firm formerly known as The Blockchain Group, raised 21 million euros with warrant structures explicitly earmarked to expand its BTC treasury. Genius Group, a Singapore-based AI education company on NYSE American, unveiled a five-year plan to accumulate $827 million in Bitcoin by 2031, funded through perpetual preferred stock under a $1.2 billion shelf registration. Bitcoin deployment starts Q4 2026. And Alpha Modus announced a share-for-Bitcoin swap that could add 3,170 BTC, over $200 million, to its balance sheet. The market reaction was brutal, stock down about 25% on dilution fears.
Then there's the concerning stuff. Metaplanet moved 4,800 BTC worth $377 million to Coinbase, part of a 10,270 BTC transfer this week, which is over 29% of its reported holdings. That doesn't mean they're selling, but it's a large operational movement worth watching. And CryptoSlate flagged one mid-tier Bitcoin treasury that has essentially staked its entire BTC reserve on a single 30-day reset price with a September 24 test date and a $75,000 ceiling. Separately, a Bitcoin treasury company registered 93% of its shares for resale and put a third of its crypto into options.
The lesson here is that treasury company is not a single strategy. Strategy is buying spot with a straightforward playbook. Others are layering options, preferred stock, warrant structures, and share-for-BTC swaps that make the risk profile very different from just holding Bitcoin. If you're evaluating these names, read the filings, not the ticker.
Every enterprise AI announcement this week came wrapped in the same three words: audit, permission, approval. The agentic era isn't going to be won by the flashiest model. It's going to be won by whoever makes the audit log boring enough for a general counsel to sign off on.