The claim: OpenAI’s new always-on Dots are the biggest agent story of the week.
The counter-fact: the quieter filing, Cloudflare’s, decides who gets paid when agents show up.
Four things crossed my desk on September 30 and October 1, and I read each one down to the paper trail before deciding what it was. One matters, one is a pricing trick, one is a build decision you will make this quarter, and one is a report I refuse to treat as fact yet. Here is the operator’s read, in that order.
Matters: Cloudflare will bill agents per request. Cloudflare’s own beta announcement, dated September 30, says domain owners can charge AI agents for access to websites, APIs, MCP tools and datasets using the HTTP 402 Payment Required status code. Payment settles inline with the request, in USDC on the Base network through Coinbase’s x402 Facilitator, so there is no checkout page and no subscription for the agent to sign. The beta is closed and limited to U.S. sellers and buyers, and the named early users are Cloudflare’s own AI Gateway, Ceramic.ai, Stocktwits and API2PDF. I run a content engine that makes thousands of automated requests, so I read this as a price list that is about to be pinned to every door I walk through. If you sell data or an API, the request-level price is the first new revenue model for your endpoint in years, and if you buy, your agent budget just grew a line item that does not appear on any seat license.
A 402 response is a quote, not an error. Your agent needs a spending cap before it needs a smarter model.
Marketing, with a real feature inside: OpenAI Dots. The Next Web’s writeup describes Dots as always-on agents, each with its own cloud computer and browser, kept separate from your laptop unless you connect it. They reach more than 4,000 apps through plugins and take messages in ChatGPT, Slack and Teams. One Dot ships with the Pro tier and with Business Premium, and Enterprise, Edu and Healthcare workspaces get a beta, so most small teams are waiting. The isolated cloud machine is the part worth copying, because an agent that cannot touch your laptop cannot break it. The “works while you sleep” pitch is the marketing, since nobody has published a failure rate for unattended runs.
Marketing, with a clock on it: Gemini 4 Argon. VentureBeat reports introductory pricing of $2 per million input tokens and $10 per million output, with cached input at $0.10, and a stated move to $4 and $20 afterward. Google says Argon leads or ties on more of 18 benchmarks than GPT-6 Astra or Claude Opus 5.5, while Astra still wins on advanced software engineering. Access starts with a trusted cyber-defender program, then paid API customers and Google AI Ultra subscribers. Do not build a cost model on the introductory rate, because the standard price is exactly double and the benchmarks are Google’s own.
What I Am Not Calling A Story Yet
Al Jazeera says the FTC opened a probe into OpenAI and Anthropic over rogue agents, and it also says it could not independently verify the reporting, which came from other outlets. No filing, no deadline, no named legal basis. I am treating it as unconfirmed until the agency says something on the record. If it is real, the first thing it breaks is any pipeline that tells customers an agent run is safe without logs to prove it.