Hey friends,
Anthropic and OpenAI both shipped cheaper frontier models within a day of each other. That is not a contradiction of the slowdown talk, it is the price of being allowed to pause. Also: a Connecticut deadline lands in eight days.
Anthropic released Claude Opus 5.5 on Tuesday and cut the price of its most capable tier. Output tokens now cost $20 per million, down from $25, and Anthropic says the new model costs 40% less to run than Opus 5 on typical work. OpenAI shipped GPT-6 Sol and Luna the same week at half the cost of its previous 5.6 versions. Reported by TechCrunch and Anthropic.
This is Anthropic's first release since its chief executive published an essay asking the industry to pace the frontier, and the company says so in the announcement. The logic is uncomfortable but plain. A lab cannot pause while it is burning cash, so cheaper and more efficient models are what make a voluntary slowdown survivable. Axios read the industry the same way: the frontier has moved from who has the best model to who has the cheapest one.
What changed inside the model matters almost as much as the price. Anthropic says early testers used Opus 5.5 to finish a 680,000-line code migration in under a day, and that it fixed page load times across a web app in 39 of 40 attempts. TechCrunch notes it is also less likely to use jargon and more likely to put the important part of an answer first, a quiet daily gain for anyone reading AI output for a living.
So what: your AI bill just got structurally cheaper, so an annual contract signed last year is priced against a market that no longer exists. Re-price at renewal, and move heavy repeatable work such as document summarising, data extraction and code review onto the cheaper tiers. Wait for the mid and small tiers the company says are coming before committing to a volume deal.
⚖️ Connecticut's AI hiring law goes live October 1
Connecticut's AI Responsibility and Transparency Act takes effect October 1, 2026. Any employer using AI in hiring, promotion, discipline or termination decisions needs a risk management policy, an impact assessment for each covered system, and notices plus a route to human review for affected employees. Penalties reach $5,000 per violation, enforced by the state attorney general. The part most companies have not absorbed is who carries the duty: the employer deploying the tool, not the vendor who built it. Details at Insurance Business. So what: inventory the AI already inside your HR stack, including resume screening and performance tools, and stop treating a vendor compliance page as your defence. Eight days.
🔒 The largest US cyber loss is a plain email, by a factor of seven
New analysis of FBI data puts business email compromise losses at just over $3 billion in 2025, more than seven times the combined losses from data breaches, ransomware and stolen intellectual property. It produced 24,768 reported incidents. Ransomware, the thing everyone fears, caused $32.3 million. Seven out of every eight dollars US businesses lost to cybercrime came through an email, and reported victims rose 18% to nearly 34,700. Analysis at Insurance Business. So what: check whether your cyber policy caps funds-transfer and social-engineering claims below its breach limits, because many still do. Then enforce one control, that any payment change is verified by phone on a known number, not by replying in the thread.
🛒 Amazon opened its seller tools to outside AI agents
Amazon now lets independent sellers run parts of their business from inside Anthropic's Claude or Amazon's own Quick assistant through a plugin, so inventory checks, price changes and listing updates happen without opening Seller Central. Amazon says the hookup takes about 60 seconds and no coding, and that roughly 90% of sellers already use some outside AI. The stakes are large: independent sellers are more than 60% of units sold on Amazon worldwide, and seller fees brought in $46.8 billion in the second quarter, more than AWS. Reported by GeekWire. So what: if you sell on Amazon, this is a free operational lever, and Amazon blocked Meta's competing shopping agent days earlier, so access is granted, not automatic.
🏗️ One data centre can now cost $40 billion, and 20 sites carry most of the storm risk
Howden's latest research on data centre risk finds a single facility can cost up to $20 billion to build, or up to $40 billion with all the technology installed. Only 20 US locations account for 81% of the square footage hit by severe hail or tornadoes over the past decade. Reported by Digital Insurance. So what: if you operate near one of those 20 sites, expect insurers to tighten data centre terms first and commercial terms around them next. If you share a utility with a large build, expect the cost allocation to reopen at the next rate case.
🌍 China just overtook the US as the top destination for elite AI researchers
Carnegie China's analysis found China took 40.6% of elite AI talent in 2025, up from 27.1% in 2022, while the US fell from 46.4% to 34.2%, as researchers increasingly choose to stay home. Via Techmeme. So what: if your AI hiring plan assumes top researchers will relocate to the US or Europe, re-test that assumption and budget for in-region or remote teams. The supply is moving.
🏛️ UK regulators want a choice screen for AI assistants on every new Android
The UK Competition and Markets Authority proposed rules requiring Google and Android handset makers to show search choice screens at first use and yearly, and offer a choice of AI assistant rather than a preinstalled default. Via Techmeme and Engadget. So what: choice screens have historically reset default-app economics overnight. If your growth plan depends on being the default assistant on a phone, treat that placement as a yearly renewable contract, not a position you own.
🎬 YouTube put an AI in Studio that replaces your thumbnail freelancer
YouTube's Studio app now gives AI feedback on an unpublished video's pacing, structure and storytelling, generates thumbnails and titles matched to a creator's style, and can test three thumbnail options against different audience segments at once. YouTube says creators have run more than 40 million tests on titles and thumbnails. Reported by The Verge and TechCrunch. So what: if you publish video, move testing in house. Generate three thumbnails, let the platform split-test them, and stop paying per asset for a job Studio does at no marginal cost.
Think about a taxi meter. The rate per kilometre can fall while your fare goes up, because you took more trips or sat in traffic. AI pricing works the same way. Every price cut makes it tempting to hand more work to the model, and agent tools make that easy because they retry, check their own output and loop without asking.
There is a second trap in where the cut lands. Anthropic's headline number is about output tokens, the words the model writes back. That still matters, because the meter that moves your bill is the one matching your usage pattern. Long documents going in with short answers coming out react very differently to a cut than a chat-heavy workload does.
So what: before you celebrate a price cut, measure your mix. One week of usage data tells you whether you are input-heavy or output-heavy, and which tier is doing work a cheaper one could handle. That measurement beats any pricing announcement, because it is the only part of the bill you control.
🧠 Trivia
Both big AI labs cut prices this week. Which part of the bill did Anthropic's headline number actually cut?
Output tokens, the words the model writes back, from $25 to $20 per million. The lesson stands: the meter that changes your bill is the one matching your usage pattern, not the headline percentage.
AI for Everyday Business: practical AI, no jargon.
Forwarded this? Subscribe here
Sent by AI for Everyday Business.
8735 Dunwoody Place STE R, Atlanta, GA 30350