The best AI model got cheaper, Amazon blocked Meta's shopping agent, and the way people use AI quietly changed shape

Anthropic and OpenAI released new models on the same day, and the headline is price: near-top performance now costs 40 to 50 percent less than it did a month ago. Meanwhile Meta's personal agent became the most downloaded app in the US, Amazon cut it off, and Shopify signed up, which is the first real fight over whether your business will sell to people or to their agents.

• Claude Opus 5.5 is the new top model and 40 percent cheaper to run, while OpenAI cut prices in half on its mid-tier models
  My take: Anthropic's Opus 5.5 took the top spot on the main independent intelligence index by five points and costs 20 percent less per token than its predecessor, with real-world savings closer to 40 percent because it uses fewer tokens per task. Box tested it on finance, cloud cost and clinical work and saw accuracy up 15 to 65 percent with 63 percent fewer tokens. The same afternoon OpenAI shipped GPT-6 Sol and Luna at half the price of the previous generation, and Zapier's advice was simple: move any workflow on the old models, it is cheaper and better. One warning on Opus 5.5. Because it matches Anthropic's restricted Mythos model in biology and security, it carries the same guardrails, and lawyers and life science companies report more refusals on legitimate work. If you are in a regulated field, test it on your real documents before you switch. Everyone else: the cost of a given level of AI has been falling about 47 percent a quarter since 2023, faster than any technology on record. Do not let last year's cost estimate kill a project. Re-price the two AI ideas you shelved as too expensive, because they probably are not anymore.
• Meta's Muse agent hit number one on the app store, Amazon banned it from shopping, and Shopify opened every store to it
  My take: Muse is the first personal AI agent that ordinary people are adopting, and it overtook ChatGPT as the top free app in the US within two weeks. People are using it to cancel forgotten subscriptions, chase insurance claims, book services and order groceries. Amazon responded on Sunday by blocking it, officially on terms of service, but Amazon makes 76 billion dollars a year from advertising and agents do not look at ads. Shopify went the other way and gave Muse direct access to every Shopify store with its one-click checkout. This is the start of a real split: platforms with market power will charge or block agents, everyone smaller will open the door because the agent will route customers wherever it is easiest to buy. If you sell online, find out this month whether an agent can actually complete a purchase on your site. A Shopify store already can. A custom checkout with pop-ups, captchas and a mandatory account probably cannot, and you will lose orders you never see. And if you rely on a marketplace for most of your sales, watch what it decides about agents, because that decision will be made for you.
• Interfaces are collapsing into one thread, and the chat box you type into is now managing a team of agents
  My take: Anthropic merged Claude's chat, work and design products into a single Claude this week after finding the most common complaint was not knowing which one to open. Cursor, Codex and Claude all moved to a model where you keep one long running conversation per area of work, and the AI splits tasks into background threads and coordinates them for you. Muse won on exactly this: sensible defaults, no setup, and tools that appear only when you need them. Two things follow for a business owner. First, the barrier for a non-technical person to get real work done just dropped again, so the 'my team will not use it' objection is weaker than it was a month ago. Second, watch your bill. When one request quietly spins up ten helper agents all running the most expensive model, a routine task can burn a day's allowance. Ask each vendor how to set the model used for helper work, and stop starting from a blank page: keep one standing thread for each recurring task, such as the weekly report or ad copy, and pick up where you left off.
• Enterprises are treating the safety debate as noise, and spending on agent security and models they control instead
  My take: Two weeks into the loudest AI safety argument in years, the executives at the Wall Street Journal's council summit said it changes little about how they use AI today. What is changing is where the money goes. Spending on tools that watch agents in production is rising fast after the Hugging Face incident, because agents given to non-engineers keep escaping the boundaries their users intended. Companies are swapping AI vendors more ruthlessly than ever, and the second largest US law firm is buying its own servers so client data never touches a cloud model, while Microsoft's CEO told every firm to keep control of its own knowledge rather than hand it to one model provider. My reading is that the big risks for a normal business are not existential. They are an agent with too much access, and a vendor that changes its terms. So this quarter: list every agent your team runs and what it can log into, cut the access back to what the job needs, and make sure at least one of your AI workflows could be moved to a different model in a week.
• Interest rates went up, data centre debt started to wobble, and governments are writing kill switch laws and pushing liability onto the labs
  My take: The Federal Reserve raised rates for the first time in three years, with more rises expected, and bonds tied to Oracle and Jane Street data centres sold off sharply. Meanwhile California ordered a working group on an AI kill switch, Virginia restricted data centre construction, the US Treasury Secretary said AI companies must carry full liability for what their systems do, and Anthropic pushed its IPO to November. None of this stops the tools you buy from working next week. It does mean the subsidised era is ending: money is dearer, regulation is coming, and vendors facing liability will add more refusals and stricter terms. Two practical steps. Budget for AI prices to hold or rise on premium plans even as the mid-tier keeps getting cheaper, and keep a simple record of where AI touches customer data and decisions in your business, because when the rules land you will be asked for it.