5 Things to Know About AI This Week

A quick scan of what happened in AI this week.

1. Anthropic released Claude Opus 5 at half the price of its top model

Anthropic shipped Claude Opus 5 on July 24, priced at $5 per million input tokens and $25 per million output tokens, the same rate as its predecessor Opus 4.8. The model offers a 1 million token context window, up to 128,000 output tokens, and adaptive reasoning that scales effort up or down depending on the task. Anthropic is positioning it as a model that gets close to the performance of its flagship Claude Fable 5 at roughly half the cost.

Real estate angle: Cheaper, more capable models mean the AI tools built into your CRM or transaction software will keep getting better without a price hike passed on to you. That's good news for the bottom line. It also means more vendors will plug AI into more parts of the closing process. Before you adopt any new AI feature in your software stack, ask what data it touches and where that data lives. Cheap doesn't mean risk-free.

Source: TechCrunch

2. An OpenAI model broke out of its test sandbox, and 1,100 AI employees asked Washington to slow things down

OpenAI disclosed that during an internal cybersecurity evaluation, its GPT-5.6 Sol model (along with an unreleased, more capable model) exploited a vulnerability to escape its sandbox and access Hugging Face's production systems while its safety guardrails were turned off for testing. The incident led directly to a July 28 open letter, "Pacing the Frontier," signed by more than 1,100 employees at OpenAI, Anthropic, Google DeepMind, and Meta. The letter doesn't call for a pause. It asks the US government to help build a verifiable, coordinated way to slow AI development if capabilities start outrunning oversight. Both OpenAI and Anthropic endorsed the letter at the corporate level within hours.

Real estate angle: This is a story about AI systems doing things their own creators didn't authorize, in a controlled test, with the people who built them raising the alarm. If the labs themselves are asking for guardrails, that's a signal to be conservative about how much unsupervised AI activity you allow in your own transaction workflow. Any AI tool with access to client funds, wire instructions, or signed documents should have a human checkpoint before anything moves. No exceptions, no matter how good the demo looked.

Source: CNN Business / Axios

3. The EU's AI Omnibus took effect, easing some compliance burdens

On July 28, the European Union's AI Omnibus package came into force, amending parts of the EU AI Act to simplify compliance timelines and reporting requirements for companies deploying AI systems. It's the EU's response to industry complaints that the original AI Act's rollout schedule was too aggressive and too vague in places, particularly for smaller companies trying to figure out which rules applied to them.

Real estate angle: Real estate professionals in Louisiana aren't subject to EU jurisdiction of course, but EU rules have a way of becoming the template US regulators borrow from. Watch this one. If the EU is easing up on documentation and audit requirements for AI systems, it tells you where the ceiling on "reasonable compliance" is likely to land here too, whenever US federal or state AI rules for real estate services eventually get written.

Source: Help Net Security

4. Nscale agreed to buy Anyscale for $1.65 billion

On July 30, AI infrastructure company Nscale signed a definitive agreement to acquire Anyscale, the commercial steward of the open-source Ray framework used to run large-scale AI workloads, in a deal valued around $1.65 billion. Anyscale's roughly 200 employees move to Nscale, though the Anyscale brand will keep serving existing customers on its own. The deal is part of a broader wave of consolidation in AI infrastructure this year, as companies race to lock down the compute and tooling layers that everything else runs on.

Real estate angle: Consolidation at the infrastructure layer usually shows up later as consolidation among the software vendors built on top of it. If your platform relies on a smaller AI vendor, it's worth asking who's behind their infrastructure and what happens to your data and service continuity if that vendor gets folded into someone bigger. Vendor due diligence should include this question now, not after a surprise acquisition notice lands in your inbox.

Source: Tech Startups

5. Lofty launched House.ai to size up buyers before they reach an agent

At Inman Connect San Diego on July 28, real estate tech company Lofty introduced House.ai, an AI tool aimed at consumers early in the homebuying process. It assesses credit, savings, and affordability, and builds a personalized readiness plan before handing off a "transaction ready" lead to an agent or lender in Lofty's network. Lofty framed it as an expansion of its broader agentic AI platform for brokerages, alongside a new collaborative workspace called Cowork.

Real estate angle: Tools like this will change what a "qualified buyer" looks like by the time they reach the closing table. That's not necessarily bad, a pre-vetted buyer means fewer surprises in underwriting. But it also means more of the early financial picture is being assembled by a third-party AI system before anyone in the transaction has verified it directly. This is a trend that you just need to work with: AI is getting involved earlier and earlier before a potential client reaches you.

Source: Inman

The theme this week is trust, and who gets to decide how much of it AI has earned. A model breaking its own sandbox pushed the people building these systems to ask for outside limits on their own work. Regulators in Europe are recalibrating how much oversight is actually workable. A big infrastructure acquisition is a reminder that the vendors underneath your software stack can change hands overnight. And a new buyer-readiness tool shows AI creeping earlier into the transaction, before a real estate professional gets involved. None of this means slow down on adopting AI in your own practice. It means keep your own checkpoints where they matter, let the tools do the volume work, and keep the judgment calls with a person.

- Jason