Three Stories This Week, All Asking the Same Question: Who Does the Work
Busy week, one thread holding it together. Anthropic shipped a small model that runs repetitive work for roughly four times less than its predecessor. Meta started testing an ad assistant that no longer just advises: it changes your account. And Google gave its agents an email address, a calendar and a line in the company directory.
The thread is that all three answer the same question from a different angle. Who does the repetitive work, what it costs, and under whose authority it happens.
High-Volume Work Just Changed Price
On October 7, Anthropic released Claude Haiku 5.5, the smallest and fastest model in its range. A small model is one that knows less than the big ones but is built to handle enormous volumes of simple requests without wrecking your bill. Anthropic says an average workload costs around four times less to run than on the previous version, and the model is available on Amazon, Google and Microsoft. Second piece of news, quieter and more important: this model holds up far better against hidden instructions. A hidden instruction is a sentence slipped into a web page, an email or a document that gives an order to the agent reading it, on your behalf. The previous small model got hijacked more often than it held. This one almost always holds.
Why this matters to you. Everything you gave up automating because the cost at volume did not add up becomes viable again: qualifying inbound requests, triaging tickets, pulling information out of documents, summarizing calls. On that kind of work, what decides is not writing finesse, it is the price per thousand calls. And the resistance to hidden instructions unlocks the one category that has been stalled for two years, agents that read your inbox or browse the web for you: the risk of a hijacked agent rarely justified the upside.
What I would do. I would take one high-volume task you already run on a big model, move it to the small one, and compare both outputs on two hundred real cases before switching for good. At Woosmap, the question was the same at a different scale: scanning hundreds of thousands of domains only worked because we refused expensive work everywhere cheap work was enough. What I would not do is move the authorization checks into the model. A model that almost always holds, across ten thousand calls a day, still gets caught. Your code decides what an agent may send or change, never the model itself.
Meta's Ad Assistant Stops Advising and Starts Acting
On October 6, at Advertising Week in New York, Meta announced that its assistant inside Ads Manager is moving to action. Until now it recommended and you applied. Now it executes: building an ad set, swapping the target audience, producing copy and visuals, adjusting a budget, all from a conversation. Meta also lets you hand it your brand and strategy documents to frame what it produces. More importantly, a system of previews and approvals lets the advertiser pick what gets delegated and what stays human. Account-level actions remain in test for now, approvals included.
Why this matters to you. An assistant that advises, you can ignore for six months. An assistant that changes a budget has to be fenced in before it arrives. The real question is not whether it works well. It is who, on your side, is allowed to let it touch the account, and how you will see afterwards what it changed. In most teams I work with, that answer does not exist, simply because nobody had to ask.
What I would do. I would write the rule this week, in three lines, before the test lands in the account: what it does on its own, creative variants for instance, what goes through a human approval, targeting and audiences, and what it never touches, the budget. I would also check that the account's change history is readable by someone other than its author. At Spendesk, I took over a four-market paid setup built by someone who had already left: the hard part was never the campaigns, it was reconstructing who decided what and why. With an agent writing into the account continuously, that archaeology becomes a weekly job unless you organize it now.
Google Gives Its Agents an Email Address
On October 8, at its Gemini at Work event, Google showed something that looks less like a product than like a hire. A manager describes a role, events coordinator for example, and Google creates an agent with its own Workspace account: an email address, a calendar, storage and an entry in the company directory. Colleagues add it to a group chat or mention it by name, like a teammate. The agent works directly inside Gmail, Drive, Docs, Sheets and Calendar, and draws on Google's models or Anthropic's depending on the task. It is a closed preview for now, and Google has published no price.
Why this matters to you. An agent with a mailbox and a directory entry is no longer a tool, it is an account. So it raises exactly the questions you already settled for employees and contractors: who creates it, what data it can reach, who reviews what it sends outside the company, and who switches it off when the project ends. You answered those a long time ago for humans, never for a colleague who writes a hundred emails a day and never takes a holiday.
What I would do. Nothing to deploy: the preview is closed and the price is unknown. I would do exactly one thing, and it takes an hour. Decide who, in your company, owns the creation of agent accounts, and write it down. If the default answer is "any manager", you will discover fifteen live agents one morning and nobody will know what they can access. If you administer Workspace or Microsoft 365, I would also find where the setting that blocks those accounts lives. Better to switch it on the day you decided to than the day Google opens it.
Everything Else, in Three Lines
GPT-5.5 disappears from ChatGPT, ChatGPT Work and Codex on October 14, while access through the programming interface stays intact: if you pinned it in a custom assistant, a scheduled task or a Codex configuration, change it this week. Meta and Sierra announced, with Shopify, Stripe and Walmart, a common standard for the personal agents that buy on your customers' behalf, but nothing has shipped yet: the specification is promised for the end of October, so there is nothing to do. And OpenAI published several hundred previously unsolved mathematical results produced by one of its models, which is remarkable and changes nothing about your Monday morning.
See You Next Friday
This week's three announcements say the same thing in three different languages. Repetitive work costs four times less than it did, the agent has its hands inside your tools, and it is starting to carry a name in your directory. None of it calls for a big programme. All of it calls for a written rule, and one named person to hold it.
If one of the three is live for you, moving a volume task onto a small model, the scope you grant an agent inside your ad account, or the rule for creating agent accounts on your side, let's talk. See you next Friday.