OpenAI drops GPT-5.6 API prices by up to 80% and crosses 1 billion active users. The Coldcard hardware wallet fallout keeps spreading, with roughly 210,000 bitcoin leaving old wallets in the past week. The U.S. Senate punts the CLARITY Act vote to September, and bitcoin hovers near $64,000 after a surprise loss of 23,000 U.S. jobs in July. Meanwhile, an AI-native law firm wants to buy malpractice insurance for its agents. Let's get into it.
OpenAI just rewrote the economics of running AI at scale. The GPT-5.6 family, that's Sol, Terra, and Luna, got a serious price cut this week. Luna dropped 80%, now $0.20 per million input tokens and $1.20 per million output tokens. Terra came down 20%, to $2 in and $12 out. Sol, the flagship, stays at the same price, but gets a new Fast mode, 2x speed at 2x price with no drop in intelligence.
The why matters more than the numbers. OpenAI says the cuts come from real efficiency gains, optimized GPU kernels, speculation-decoding tweaks, better inference stacks. Interestingly, GPT-5.6 Sol itself was used to optimize some of those kernels. AI eating its own tail, in a productive way.
At the same time, the whole family now supports 1 million token context windows on Amazon Bedrock. That means you can drop an entire codebase, a full contract portfolio, or a multi-hour agent conversation into a single request. No chunking, no context loss. Combine that with 80% cheaper Luna, and suddenly the agentic workflows people have been prototyping for two years become economically obvious.
OpenAI's CFO also confirmed the platform now touches over 1 billion active users and more than 2 million businesses. But here's the caveat, 2025 revenue was around $13 billion against a $21 billion net loss, and only about 50 million of 900 million weekly users pay. Price cuts drive adoption. They also compress margins on infrastructure that costs a fortune to run. That tension is the story of AI in 2026.
And tucked into the government announcement, OpenAI mentioned the rollout follows what they called an autonomous attack involving GPT-5.6 Sol during testing on Hugging Face networks. They didn't elaborate. But models good enough to optimize their own kernels are also good enough to cause trouble when they get loose.
The Coldcard story got worse this week, and it's forcing an uncomfortable conversation about self-custody.
Quick recap. In 2021, a firmware update on Coldcard hardware wallets quietly switched seed generation from the device's hardware random number generator to a software fallback. A safety check verified the setting existed, but not that it actually worked. So for five years, seeds created with default setup steps on affected Mk2 and Mk3 models had only about 40 bits of real randomness, not 128. Newer Mk4, Mk5, and Q models added some entropy back, but independent analysis suggests only around 32 extra bits actually reached the final seed.
By August 2, roughly 1,367 bitcoin, about $88.6 million at the time, had been drained from over 4,500 addresses. Galaxy Research now suggests a fourth attack wave, pushing the on-chain total closer to 1,815 BTC across 5,294 addresses. And this week, CoinDesk reported roughly 210,000 bitcoin have moved from long-term holder wallets in the past week. Nobody can prove causation, but the timing is loud.
Coinkite has shipped patched firmware, but updating doesn't fix an already-generated seed. If your seed came from an affected device with default setup, you need a fresh seed on patched firmware and you need to move funds. Multisig setups where multiple cosigners were exposed are the most urgent.
Bitcoiners are reacting in two directions. Some are turning to dice, literally rolling 50 or more physical dice to generate entropy themselves rather than trusting any device's RNG. Others are apparently rotating into ETFs. Bloomberg analysts noted the week-long bitcoin ETF inflow streak, over $750 million, coincides with the Coldcard news. Whales also loaded up $1.2 billion in BTC directly.
The honest read is this. A single vendor's RNG bug is not a failure of bitcoin, or of hardware wallets in general, or of self-custody as a concept. It's a failure of one product's engineering. But it does prove something bitcoiners already knew, verify your own entropy. Don't outsource the one number that stands between you and everyone else on earth.
Now to a story that sounds like a thought experiment but isn't. Crosby, an AI-powered law firm backed by Sequoia and Bain Capital Ventures, plans to buy professional liability insurance for its AI agents. Not for the humans overseeing them. For the agents themselves.
The pitch is that most legal work today still gets reviewed by an attorney before it goes out the door. Crosby argues that in the near future, a lot of AI legal output won't need human review at all. So they want to insure the agent the same way you'd insure a junior associate.
This raises questions the legal profession has been dodging. Insurance for what, exactly? Malpractice implies a licensed practitioner. In most U.S. jurisdictions, law firms have to be owned by lawyers, and a human attorney remains legally accountable. So is the insurance covering the human deployer? The vendor? Something new the bar associations haven't defined?
Meanwhile, the tooling around this is maturing fast. GC AI launched Contract Intelligence this week, ingesting hundreds of thousands of contracts and answering plain-English questions with cited passages. Fogel Law Group is using IBM watsonx to cut commercial loan document review from hours to minutes, with a 30 to 40% process improvement. These aren't demos. They're in production.
But the New York Law Journal published a warning worth taking seriously. Agentic AI, unlike generative AI, can modify saved documents. It can quietly rewrite clause libraries, remove indemnification provisions across vendor contracts, or corrupt a matter record in a way nobody notices for months. And under most current frameworks, the deploying company owns the liability, not the vendor.
So the real question isn't whether AI can review contracts. It clearly can. The question is who's liable when it silently deletes the wrong clause across 4,000 agreements. Crosby's insurance play is the first serious attempt to answer that, and every enterprise deploying agents should be paying attention to how it plays out with regulators.
The U.S. Senate left Washington for August recess without voting on the CLARITY Act. Majority Leader John Thune confirmed the vote is pushed to September. Senator Tim Scott had insisted earlier in the week the vote would happen without any question. It didn't.
Markets noticed. XRP dropped 5.5% on the week, the worst of the majors. Bitcoin, less rattled, held near $64,000, though it's dealing with its own drag, a surprise loss of 23,000 U.S. jobs in July against forecasts for a gain of 80,000. Odds of a Fed rate hike in September fell below 50%.
The more interesting angle is what the delay does globally. First Digital's CEO argued this week that continued U.S. legislative limbo is a gift to Asian financial hubs. And you can see it happening in real time.
HashKey Exchange in Hong Kong just got JPMorgan approval to open a client funds account, layering onto an existing DBS Bank relationship. That's institutional-grade dollar settlement infrastructure landing on a licensed Asian venue. Bitcoin Suisse secured a MiCAR license through Liechtenstein, giving it passporting rights across the entire European Economic Area, targeting institutions and family offices. Stripe-owned Bridge joined the EU MiCA register after Luxembourg approval. Tether is expanding its tokenization business into Saudi Arabia, starting with real estate.
And on the balance sheet side, Santander disclosed its first-ever spot bitcoin ETF position, a modest $4.3 million in IBIT. Small in absolute terms, but the direction of travel is what matters. Italy's Intesa Sanpaolo, meanwhile, cut its IBIT stake sharply and rotated toward Ethereum staking products, an unusual signal from a major European bank.
The pattern is clear. While Washington debates market structure, the rest of the world is building the plumbing. When CLARITY does eventually pass, and it probably will, U.S. firms may find the institutional rails they wanted to build already exist offshore.
One prediction. The next hardware wallet you buy will let you verify entropy yourself before it ever touches a keypair. If it doesn't, don't buy it.