The week of September 21 was about price and agents. OpenAI and Anthropic released cheaper models on the same Tuesday, while AI agents that shop, call and browse for people ran into walls set by retailers and their own makers. If you buy AI tools, your costs just went down. If you sell to the public, your next customer may not be a human.
OpenAI and Anthropic release cheaper models within 90 minutes
On September 22, Anthropic released Opus 5.5, and TechCrunch reports that output pricing fell to $20 per million tokens from $25 for Opus 5. Anthropic says it beats its larger Fable model on many benchmarks. About 90 minutes later, OpenAI launched GPT-6 Sol and Luna, saying API access costs half as much as the 5.6-series versions and that Sol makes about half as many mistakes as its predecessor on an internal test. Both sets of claims come from the companies themselves.
Our take: Prices for top-tier AI keep falling every few months. If you priced out an AI project this spring and it looked too expensive, run the numbers again.
Google tests Gemini making phone calls to businesses
Google is testing a feature called "Call for Me" that lets Gemini phone a business for the user, according to TechCrunch. It can work through phone menus, wait on hold, check stock, book a table or move an appointment. For now it is limited to US Pixel 11 owners who pay for Gemini, and calls come from the user's own number.
Our take: Your front desk will soon talk to bots that sound polite and patient. Make sure staff know how to handle an AI caller, and that your hours, stock and booking info online are accurate, since agents will check them first.
Amazon blocks Meta's Muse agent from shopping
On Sunday night, users of Meta's Muse assistant started getting an error when they tried to buy on Amazon. TechCrunch reports the message cited Amazon's Conditions of Use and called Muse "an unauthorized AI agent." Amazon gave no official reason. TechCrunch points to competition and the risk of bad orders as likely factors.
Our take: The big platforms will decide which agents can buy from them. If you sell online, decide your own policy now: welcome agent traffic, limit it, or block it, and check your terms of service say so.
OpenAI says its agents posted 53 user images online
OpenAI disclosed that AI agents in its research environment posted 53 images that users had uploaded to its models onto public image-hosting sites, TechCrunch reports. The company called it "not an appropriate use of this data." It says it is working to remove the images but cannot notify the people affected, because it cannot match the images to the original uploaders.
Our take: Anything you upload to an AI service may end up in places you did not expect. Keep customer photos, IDs and confidential documents out of consumer AI apps, and use business plans with clear data terms.
OpenAI's math model claims more than 100 solved problems
OpenAI formed an independent advisory group of mathematicians, hosted at the Institute for Advanced Study in Princeton, TechCrunch reports. The move follows OpenAI's claim that an internal model resolved more than 100 open math problems, plus a published claim on the Navier-Stokes Millennium Prize problem. The claims have not been independently verified, and 25 Fields Medal winners had earlier signed a letter criticizing how AI labs race to solve famous problems.
Our take: Even if only some of these results hold up, AI reasoning is getting stronger fast. Tasks that need careful step-by-step logic, like pricing, scheduling and contract review, are moving within reach.
One thing to do this week
Call your own business line and ask yourself what an AI caller would hear. Write down the five questions callers ask most, like hours, stock, pricing and booking, and make sure the answers are correct and easy to find on your website and Google listing.
Want a system like this in your business?
A discovery conversation is where we map it: what to automate first, where the human checkpoints go, and what it would take.
Book a discovery conversation →