Welcome to the first issue of Avolis Field Notes, our weekly newsletter. Every Friday, we will pick out the handful of AI stories from the past week that we think actually matter to people running real businesses, explain them in plain English, point you to the original sources, and tell you what we make of them. We will skip the hype when we can and call it out when we can't.
A theme runs through this week's news, and it is one we care a lot about. The technology got cheaper, faster, and more capable again this week. The hard part, getting people to trust it, learn it, and use it well, stayed exactly where it was. That is the thread running through both of our other pieces this week, too. More on those at the bottom.
Here is what happened.
01
The AI price war just got real
What happened
On Tuesday, Anthropic released Claude Opus 5.5, which it calls "the strongest-performing model we've tested to date," and cut its price to $4 per million input tokens and $20 per million output tokens. Within minutes, OpenAI answered with two new models. GPT-6 Sol, aimed at coding and agent work, costs $2 and $10, which is half of what its predecessor cost. GPT-6 Luna, built for high-volume routine tasks like summarizing and pulling data out of documents, costs just $0.10 and $0.50. OpenAI says the new prices are permanent.
Our take
If "tokens" is new to you, think of them as the unit AI companies bill by, roughly a few characters of text each. The short version is that the cost of running capable AI dropped sharply in a single afternoon. For a business owner, that means price is becoming a weaker excuse every month. The tools are getting cheap enough that the real cost of an AI project is shifting almost entirely to the human side: the time it takes to redesign a workflow, train the team, and support them while they adjust. As one analyst told TechTarget, what matters is whether these models solve real business problems "in a secure, performant, governable and trustworthy manner." We agree, and that part does not come in the box.
02
Anthropic's revenue is growing at an almost unbelievable pace
What happened
The New York Times reported, and Axios and Bloomberg followed, that Anthropic's annualized revenue has passed $100 billion. That is up from roughly $67 billion about two months ago and around $10 billion at the end of 2025. The company is reportedly aiming to go public as early as November.
Our take
Whatever you think about AI valuations, this is a real signal that large companies are spending heavily on AI, especially coding and workplace tools. What the revenue number cannot tell you is how much value those customers are getting back. The research we have been reading suggests a wide gap: MIT researchers found that about 95 percent of organizations saw no measurable profit impact from generative AI in 2025. Spending is the easy part to measure. Adoption is the part that decides whether the spending was worth it.
03
AI is now doing a big share of the work of building AI
What happened
Anthropic said on September 17 that its Claude models now lead about 26 percent of the company's own model research and development, up from essentially none in February, and that roughly 90 percent of its R&D happens in collaboration with Claude. The company also acknowledged the obvious concern, that systems helping to build their own successors could become harder for humans to understand and control.
Our take
The headline here is speed. If AI is helping to build the next generation of AI, the gap between releases is going to keep shrinking, and the tools your team learned six months ago will keep changing underneath them. That is exactly why we push clients away from one-time training. A single workshop gets stale fast. What lasts is a team that has the habits, the shared playbooks, and the support to keep learning as the tools change.
04
Anthropic picks Accenture as its first "embedded evaluator"
What happened
On September 18, Anthropic announced that evaluators from Accenture will work inside the company with access comparable to an employee's, testing its models, assessing its safety decisions, and reporting publicly on what they find. Both companies expect to spend at least $1 billion each over five years. The choice surprised some observers, who had expected AI safety nonprofits like METR to go first. Anthropic says more evaluators will follow.
Our take
The best way to judge the AI labs' safety talk is to watch what they actually do, and this is one of the first concrete actions worth watching. It is a real commitment, and it also raises a fair question about how independent a large consulting firm can be from one of the most important companies in its market. We will be watching for the moment an evaluator publishes something the company would rather they didn't. Notice one more thing, too: when Anthropic wanted to make its accountability credible, it put people inside the building. Even the most advanced AI company on earth reached for embedded humans.
05
Washington and Sacramento are pulling in opposite directions
What happened
On September 18, California Governor Gavin Newsom signed an executive order directing the state to speed up independent audits of the largest AI labs and to advance a "kill switch," meaning an emergency shutdown mechanism, for the most powerful models. Four days later, President Trump told the United Nations General Assembly that the United States "totally rejects any attempt to construct a globalist scheme to control" artificial intelligence, according to Scientific American. OpenAI's Sam Altman and Anthropic's Dario Amodei were set to brief the UN Security Council on AI and international security on Wednesday, Fortune reported.
Our take
Almost none of these rules land directly on a 50-person HVAC company or a regional accounting firm, since they target the handful of labs building frontier models. What they do tell you is that the rulebook is going to stay unsettled for a while, and that it may look different depending on which state you operate in. You do not need to wait for it. Every business using AI should already have its own simple policy: what AI is allowed to do on its own, what needs a person to approve it, and what information should never leave the building. That policy protects you no matter which way the politics go.
06
Meta's personal AI agent is a hit, and it already had a security scare
What happened
Meta's new personal AI agent, Muse, which can work across your apps, files, calendar, notes, and messages, launched on iPhone on September 8 and on Mac on September 17. It was downloaded about 2.5 million times in its first 12 days, outpacing ChatGPT over the same period. Along the way, security researcher Patrick Wardle disclosed a serious flaw in the Mac version that could have let another program on the computer take control of a user's account.
Our take
Here is the part business owners should sit with. Some of those 2.5 million downloads are on your employees' phones and laptops, and some of those laptops belong to your company. People are going to bring powerful AI tools to work whether or not there is a plan, which is why banning them rarely works and ignoring them is worse. A clear policy, a short list of approved tools, and some training on what is safe to share will do far more to protect your business than hoping nobody installs the next hit app.
07
AI agents are learning to check out and pay
What happened
Stripe, the payments company behind millions of online checkouts, published details on September 22 about how it is redesigning its checkout for AI agents. Using an experimental web standard called WebMCP, a Stripe checkout page can hand an AI agent a set of structured tools (read the amount due, choose a payment method, fill in the form) so the agent does not have to guess where to click. Stripe says its internal testing showed agents used 42 percent fewer tokens and finished 39 percent faster this way.
Our take
This one sounds technical, but the implication is simple. Some of your customers will soon send an AI agent to pay an invoice, book a service call, or reorder parts on their behalf. The businesses that are easy for those agents to work with will pick up a quiet advantage. It is worth asking your web and payments providers now how agent-friendly your checkout and invoicing really are.
08
The music labels sue Suno again, and the argument matters beyond music
What happened
On September 18, Universal Music Group and Sony Music filed a second lawsuit against the AI music company Suno, this time over its new v6 model. Suno says v6 was trained from scratch in partnership with Warner Music Group and others. The labels argue that it still learned from earlier models they say were built on their recordings, writing that "training a 'new' model on the outputs of an infringing model does not eliminate the infringement; it launders it."
Our take
Plenty of our readers use AI to generate marketing copy, images, and proposals, even if few of them are making music. This case is testing whether a "clean" new model can still carry legal baggage from the models it learned from. Until courts settle questions like that, it is smart to know which tools your team uses for customer-facing content, to favor vendors that are open about their training data, and to keep a person reviewing anything that goes out with your name on it.
