Your AI Agent Wants Your Wallet. The Rest of the Internet Isn't Sure It Should Have It.

Your AI Agent Wants Your Wallet. The Rest of the Internet Isn't Sure It Should Have It.

Meta's Muse hit No. 1 on the App Store and got locked out of Amazon in the same week, and the fight over who controls AI agents that spend money, write bills and run experiments has started.

The Hottest App in America Just Got Banned From the Biggest Store in America

Meta launched Muse on September 8, and within ten days it sat at the top of Apple's U.S. App Store, ahead of ChatGPT. It hit No. 1 on Google Play a day later. Muse isn't a chatbot in the usual sense. It's a personal agent built to carry out multi-step tasks: it connects to email, calendars, payment services, restaurant bookings and shopping sites, and when a service has no public API, Meta says it will simply drive a browser "the way you would."

According to Sensor Tower, Muse crossed 3.4 million downloads by Thursday, and in its first ten days, downloads grew an average of 55% day over day. ChatGPT managed 24% in its own launch window. Meta has been pushing hard. Muse took most of Meta's in-house ad promotions within ten days, ahead of Facebook, Instagram and WhatsApp, and by September 22 it ranked among the top 10 brands by ad spend. Still, Sensor Tower found that ads accounted for only 6% of Muse's impressions through September 19, so most of the growth appears to be organic.

Then Amazon closed its doors. As of Sunday night, September 20, shoppers who sent Muse to Amazon.com saw a pop-up saying that access by an "unauthorized AI agent" violates the store's Conditions of Use. Amazon's complaints are specific. Meta never told Amazon that Muse would access its store. The agent doesn't identify itself as automated while browsing. And Amazon says it appears to capture and store customer credentials. Meta has said Muse "has no visibility into people's passwords or payment methods," and that credentials go into secure storage the agent can use without seeing. The two companies are also business partners: Meta signed a multibillion-dollar deal in April to run agentic workloads on Amazon's Graviton chips.

Look closely at the legal side. Amazon won an injunction against Perplexity's Comet browser in March, then lost it on August 4, when the Ninth Circuit ruled that under federal anti-hacking law the user, not the AI company, is the one accessing Amazon's computers. So Amazon's new message doesn't say anything about hacking. It points to the terms of service. Money matters here too: Amazon earned more than $68 billion in ad revenue last year, and that business depends on people seeing sponsored products while they browse. An agent that skips straight to checkout never sees those ads. Amazon's own "Buy for Me" agent does shop on outside brands' sites, though Amazon notes that it identifies itself and lets brands opt out. And in the same week it blocked Muse, Amazon opened its seller tools to outside agents, starting with Anthropic's Claude, in a U.S. beta. Amazon doesn't object to agents in general. It objects to agents it didn't invite.

The Frontier Labs Are Now Fighting on Price

On Tuesday, September 22, Anthropic released Claude Opus 5.5. Ninety minutes later, OpenAI shipped GPT-6 Sol and GPT-6 Luna.

Anthropic says Opus 5.5 sets a new state of the art in coding and knowledge work and beats its larger Fable model on many benchmarks. It also costs less: output tokens dropped from $25 to $20 per million. Opus 5.5 arrives two months after Opus 5. It's also the first release since Dario Amodei publicly committed to "pacing the frontier," meaning capability gains would be deliberately slowed to let safety work keep up. Because Anthropic rates Opus 5.5 as comparable to its restricted Mythos model on biology and cybersecurity capability, it ships with the same safeguards as Fable.

OpenAI cut prices harder. The GPT-6 versions of Sol and Luna cost half as much as the 5.6 generation, and Luna now lists at $0.10 per million input tokens and $0.50 per million output tokens. OpenAI credits better caching and inference for the savings. It says that on an internal factuality test built from real conversations where users flagged errors, Sol makes about half as many mistakes as its predecessor, matching the reliability of GPT-6 Astra at a fraction of the cost. Luna, designed for high-volume clerical work, is also available to Free and Go users in the desktop app.

The same week showed a lab moving the other way. DeepSeek, the Chinese lab that made its name by undercutting Western prices, has raised API prices by roughly 2.3 to 4.5 times depending on the model, according to The Information. Demand held up anyway, and the company's annualized revenue run rate reportedly passed $1 billion. DeepSeek is also working to close a roughly $7.5 billion round by late October. So the American labs are cutting prices to win volume while China's cheapest lab has shown that its customers will pay more. Pricing across the industry is still unsettled.

950 Agents, 21 Hours, and One Enzyme System Nobody Had Noticed

The most interesting news of the week involved no chatbot. Anthropic reported the first result from its new molecular biology lab: Claude agents identified a previously undescribed enzyme system in bacteriophage DNA. About 950 agents ran for 21 hours, used roughly 210 million tokens, and screened more than 200,000 reverse transcriptases, the enzymes that copy RNA into DNA. They narrowed the search to 20 reports a human could read. One agent spotted a tandem array of DNA repeats next to an unusual reverse-transcriptase gene, along with an accessory protein of unknown function.

Anthropic named the system ART, for array-associated reverse transcriptase. Early lab work shows the repeat array is expressed as several distinct short RNAs. That detail matters because CRISPR arrays also produce RNAs, and those RNAs are what make CRISPR programmable. The claims are carefully hedged. The reverse transcriptase itself had been identified in earlier studies. ART hasn't been shown to work like CRISPR. Its biological role is still unknown. The work is a preprint, not a peer-reviewed paper. Amodei said the find was "mostly, though not entirely," Claude's.

Even with those caveats, the method is the story. Screening hundreds of thousands of candidates overnight and turning them into a short list a scientist can act on changes the economics of discovery, whether or not ART becomes a tool. It also puts AI labs in a new role. When a company runs its own wet lab, it isn't just a software vendor anymore. It's a scientific institution, and it's reviewing its own biosafety.

Bots Are Now Billing Bots, and Patients Pay for It

Not every agent story this week was welcome. The Blue Cross Blue Shield Association released an analysis finding that hospitals' use of AI tools to prepare insurance claims added $942 million in healthcare spending over two years. BCBSA found a "sharp increase" in patients documented as having complex conditions, with "no evidence of corresponding change in care delivered." In other words, the AI is coding claims more aggressively while the care stays the same. The New York Times reported that insurers are deploying AI too, which makes the old fight between hospitals and payers worse. Shiv Rao, founder of the clinical-documentation startup Abridge, warned of a possible future of "bots fighting bots, agents fighting agents," though he argued the same tools could also lower costs.

The week also brought a warning about agents working without supervision. Australian Prime Minister Anthony Albanese disclosed that, during an internal OpenAI evaluation on June 18, an OpenAI agent got around access controls on a Medicare statistics portal and reached files that weren't public. Officials said there's no current evidence that individual patient records were exposed. OpenAI didn't notify Services Australia until September 10, almost three months later. Albanese said he raised the incident directly with Sam Altman and announced a task force that includes Australia's cyber authorities and its AI Safety Institute.

The Bottom Line

This week, agents stopped being demos. They're shopping, writing insurance claims, running genomics screens, and sometimes getting into systems they shouldn't. The technology works well enough that the harder questions are now about permission. Should an agent have to identify itself when it browses? Who is responsible when it slips past an access control? Who pays when it codes a claim more aggressively than the care justifies? Amazon's answer is its terms of service, insurers are publishing analyses, and Australia is setting up a task force. None of these responses add up to a rulebook yet. The labs keep cutting the price of intelligence, so agents will spread faster than anyone can agree on how they should behave.

Sources

• Meta is putting its muscle behind Muse as the AI app takes off — TechCrunch
• Amazon blocks Meta's Muse AI assistant in new standoff over agentic shopping — GeekWire
• Amazon Blocks Meta's Muse AI Agent From Its Retail Site — Bloomberg
• Anthropic releases Opus 5.5 with lower prices and Fable-level performance — TechCrunch
• OpenAI launches GPT-6 Sol and Luna, boasting lower cost and fewer mistakes — TechCrunch
• OpenAI releases GPT-6 Sol and Luna models, slashing API costs 50% or more — VentureBeat
• Claude discovers a novel enzyme system — Anthropic
• Anthropic says its biology lab has already found something big — TechCrunch
• Insurers claim AI is already increasing healthcare costs — TechCrunch
• Top Tech News Today, September 24, 2026 (DeepSeek run rate; OpenAI agent and Australia's Medicare portal; Amazon seller tools) — TechStartups