TL;DR
- Google introduced an agent for companies that, in Google's own words, plans the work, uses the company's tools and "brings back something finished" inside the documents and inbox people already use.
- Anthropic shipped Claude Haiku 5.5, its small model. The company says it costs around 75% less to run than the previous one, on average.
- OpenAI announced image ads inside ChatGPT. The test starts this month in the U.S., only for people on the free plan or the Go plan, while they're creating an image.
- Google opened its watermark detector to everyone: upload a photo, video or audio file and it tells you if AI from Google or one of its partners made it.
- OpenAI fired three safety researchers. The company says they broke rules on sensitive information; one of them says it was for putting safety first.
- This week's move: send your own chatbot the hardest question a customer asked you this month, and read the reply the way that customer would.
Almost everything announced between October 5 and 9 carries the same fine print: the company selling it is the one saying so.
What you can check is closer to home: inside your own business.
1. Google introduced an agent that hands back finished work
What happened. On October 8, at its Gemini at Work event, Google Cloud introduced the Gemini agent, which it calls "a universal agent for work." According to Google's announcement, the agent has all of an organization's business context, plans the work, uses skills and tools, connects to the company's business systems, and "brings back something finished" inside the documents and inbox people already work in. Google also says it picks the best model for each job and comes with built-in cost controls.
The announcement is aimed at companies that are already Google Cloud customers. It gives no date and no price for a small business.
What it means for you. Nothing you can turn on today. What matters is the phrase Google chose to sell it: something finished. Those are Google's words; the announcement doesn't show a finished job or say who checks it. And by Google's account, the job gets closed out with whatever the company has given it: its documents, its email, its systems.
What to do. Nothing this week. We're flagging it because it shows where the tools in your office are headed next year.
2. Claude's new small model is built for repeat work
What happened. On October 7, Anthropic introduced Claude Haiku 5.5, its fastest and cheapest model. The company says it's built for work that repeats many times a day: summaries, sorting messages, live customer support. Anthropic says it costs around 75% less to run than Haiku 4.5, on average. In its price table, the charge for every million input tokens (the small chunks of text a model reads) dropped from $1 to 10 cents for requests up to 100,000 tokens, which the company says made up 9 in 10 requests on the previous model.
What it means for you. It doesn't change what you see in the Claude app. It changes what people pay when they build with Claude: the chatbot on a website, the assistant in a WhatsApp line, the program that summarizes email. When running a task a thousand times a day gets that cheap, expect more assistants answering in more places, maybe including yours.
What to do, five minutes. What to ask your own chatbot, and what to look for in the reply, is in the exercise further down.
3. ChatGPT is about to test image ads
What happened. On October 5, OpenAI announced a new ad format in ChatGPT: images that show a product or service in use. Testing starts later this month in the U.S. with an initial group of advertisers, and the ads appear while someone is creating an image in ChatGPT, only for people on the free plan or the Go plan (the cheapest paid one). OpenAI says the ads are clearly labeled, kept separate from the image being created, and that "advertising does not influence the answers ChatGPT provides." By OpenAI's count, ChatGPT reaches 1.2 billion people a week.
What it means for you. Three weeks ago we covered the ad that opens a conversation with the brand. Now OpenAI is testing another format, the image. ChatGPT keeps adding ways to pay to show up, and each one starts in the U.S. with a small group of advertisers.
What to do. Nothing to buy. What we said then still holds: before you pay to show up, your page has to answer what the customer is asking.
4. Anyone can now check whether Google's AI made an image
What happened. On October 7, Google DeepMind opened its SynthID Detector to everyone, in English. Until now it was an early version built for media professionals. You upload an image, video or audio file at synthid.com, and it tells you whether it carries the invisible watermark used by Google's AI and its partners: OpenAI, NVIDIA, Kakao and, soon, Apple. Google says it has already watermarked more than 180 billion images and videos.
What it means for you. A month ago we covered the opposite approach, as reported at the time: sealing the real photo at the moment it's taken. This one looks for the mark a machine left behind, and it has a hard limit: it only recognizes what Google and its partners made. A clean result doesn't prove a person made it.
What to do, two minutes. When an image looks off and you're about to act on it (damage someone is claiming, a product someone is offering you), run it through the detector first. If it comes back marked, you know. If it comes back clean, you decide the way you always have: a call to the number you already had. Don't upload photos that carry customer or patient information.
5. OpenAI fired three safety researchers
What happened. On Friday, October 9, OpenAI confirmed it fired three safety researchers. The company says "they violated clear policies on handling sensitive information," and that the firings "were not about safety concerns or speaking out." The three posted a letter asking OpenAI to keep its promise to let outside safety monitors in, and one of them said they were fired for putting safety first.
What it means for you. Two versions, and nobody outside has verified either. It's in the roundup because this is the company behind ChatGPT, not because it changes anything in your office.
What to do. Nothing this week.
What we saw this week
Sorted by who's saying it, the week looks like this. That Google's agent hands back finished work is Google's claim, with no date or price for a small business. That Anthropic's new model costs about 75% less on average is Anthropic's claim. That ads don't touch the answers is OpenAI's claim. Google's detector is the one thing this week you can use yourself, and it only recognizes what Google and its partners made. And the firings come in two versions, with nobody outside comparing them.
This week, almost everything reached you as the seller's word.
Distrusting all of it doesn't help either. What helps is sorting what you can only believe from what you can check yourself. You'll never get to inspect a big company's agent. What the assistant already talking to your customers says, you can, in five minutes.
Last week we wrote that you don't give the first answer about your business anymore. As models get cheaper, expect more chats answering, yours included. And of all the answers that talk about your business, that's the one that depends most on you: you set it up, or someone you pay did.
The exercise that doesn't expire
Five minutes. If you wrote down the three decisions before you switch on an assistant last month (among them, when it hands the conversation to a person and who reads what it says), today you check whether they actually got set up, with a question you pick, before a customer asks it. If you never wrote them, this test shows you where to start. Repeat it any time someone changes the settings.
Today, send the assistant that answers for your business (the chat on your website, on WhatsApp or on Instagram) the hardest question a customer asked you this month, in the customer's words and without their name. Read the reply the way the customer who asked it would, and check three things: did it answer what was asked, did it give out anything that isn't true for your business (a price, hours, a promise), and did it offer to hand the conversation to someone on your team.
If your assistant books appointments or takes payments, don't confirm anything, and tell your team first that it's a test. And if the customer's question mentions their health or their money, change the details: whatever you type in that chat gets stored by whoever sells it to you.
If your business doesn't have an assistant, write that same question in a note, and under it, how you answered it yourself. That's the first thing any program you put on the front line would need to know.
Screenshot the reply and keep it next to the question.
Keep reading
- You don't give the first answer about your business anymore - last week's roundup
- Give a program a job it can't finish - the three decisions before you switch on an assistant
- One night between can't and can - the ad that opens a conversation, and why it isn't one to buy this year
- If AI gets your price wrong, ask it where it got the number - what to do when an assistant describes your business with the wrong facts
- SEO vs. GEO vs. AEO: which one does your business need in 2026? - the full map
- Google Cloud introduces the Gemini agent, Google, October 8, 2026 - the announcement
Frequently asked questions
What is the Gemini agent, and can my business use it? It's an assistant for companies that, by Google's account, handles multi-step work and delivers it finished inside your documents and email. For now it's aimed at Google Cloud customers, and the announcement gives no date or price for a small business.
I use Claude on the free plan. Does Haiku 5.5 change anything for me? Not in what you see. Anthropic's announcement is mostly about people who build with Claude: the per-use price dropped by about 75% on average. If you notice it at all, it'll be in assistants other people build with it.
Do ChatGPT ads change what it tells me? According to OpenAI, no: ads are labeled, kept separate, and don't influence answers. The image test starts this month in the U.S. and only shows up while someone is creating an image, for people on the free plan or the Go plan.
Does SynthID tell me if a photo is fake? It tells you whether the file carries the watermark from Google's AI or its partners (OpenAI, NVIDIA, Kakao and, soon, Apple). A clean result doesn't prove it's real: it may come from another tool. Treat it as one more clue, not the only one.
What if my chatbot gets the test wrong? Don't fix it blind. Keep the question and the reply, and ask whoever set it up to correct it. What it should be allowed to do without asking you, and when it should hand the conversation to a person, is in the three decisions before you switch on an assistant.
Wrapping up
What you can't check, the seller tells you. What your assistant says, you can check yourself.
Message your own chatbot today. You'll be done before your coffee gets cold.
Did it say something you wouldn't? That screenshot explains more than any description could. Grab a free call and we'll start there: Book a free call
One last thing: if your chatbot's reply makes you wince a little, good. You saw it before a customer did. Thanks for your time.
Sources
- Google Cloud introduces the Gemini agent, Google, October 8, 2026 - https://blog.google/innovation-and-ai/infrastructure-and-cloud/google-cloud/gemini-at-work/
- Introducing Claude Haiku 5.5, Anthropic, October 7, 2026 - https://www.anthropic.com/claude-haiku-5-5
- Building advertising for the way people use AI, OpenAI, October 5, 2026 - https://openai.com/index/new-chatgpt-ads-format-and-measurement/
- We're making it easier to identify AI-generated content globally, Google, October 7, 2026 - https://blog.google/innovation-and-ai/models-and-research/google-deepmind/synth-id-ai-content/
- SynthID Detector - https://synthid.com/
- OpenAI fires 3 safety researchers in dispute over AI risks, The Associated Press via ABC News, October 9, 2026 - https://abcnews.com/Technology/wireStory/openai-fires-3-safety-researchers-dispute-ai-risks-137123277
