Answers Got Buttons

Last Saturday I wrote about the labs hitting the brakes. Cancelled releases, paused training, regulators at the door.

Turns out half of them were flooring it the same week. OpenAI shipped GPT-6 to everyone, Anthropic knocked about three quarters off its small model, Google and Microsoft both rolled out ways to run agents like employees, and a group of retailers and payment companies started writing the rules for how AI shoppers will walk into your store.

None of it's theoretical. Most of it touches how customers find you, what your tools cost and who's doing work inside your business. Here's the rundown, with what to do about each.

OpenAI rolled out GPT-6 in ChatGPT starting October 7, with paid plans first and free users the next day. The bigger change is how answers look. A feature OpenAI calls Intelligent UI lets ChatGPT answer with interactive pieces instead of paragraphs: charts you can adjust, forms, maps and small calculators. Splitting a bill or planning savings turns into something you can click and drag, and ChatGPT picks the format on its own. OpenAI also says GPT-6 Instant starts answering web search questions 44 percent sooner than the previous version, though that's the company's own number.

What to do. Two things. First, if your website has a generic calculator, expect its traffic to soften, because people can now get one inside the chat. Wednesday's issue covered how to build the specific kind it can't replace. Second, start a simple weekly check: ask ChatGPT the same five questions a customer would ask about your type of business, and note whether the answer is text, a table or a tool, and whether you're cited. Two users can now see different formats for the same question, so log what plan you tested on.

OpenAI announced a visual ad format that shows while people generate images in ChatGPT, with a US test starting later this month for Free and Go users. Measurement partners include LiveRamp, Tealium and Hightouch. WeightWatchers reported a 15.3 percent lower attributed acquisition cost compared with its paid search benchmark. Meanwhile Digiday reported that seven months after ChatGPT ads launched, measurement gaps are keeping spend at test levels. One agency said clients spending $10 million a month on Google are putting under $100,000 into ChatGPT.

What to do. Don't move budget yet. Do set up tracking now so you're ready when it makes sense: UTM parameters on anything you might ever run there, and a landing page built for someone who arrived mid conversation and already knows roughly what they want. With a channel this early, whoever can measure it on day one gets to spend smart later.

Anthropic released Claude Haiku 5.5 on October 7. For prompts under 100,000 tokens, it costs ten cents per million input tokens and fifty cents per million output tokens. Anthropic says that's about 75 percent cheaper than Haiku 4.5 on average. It's the first Haiku with adjustable effort, it now handles a million tokens of context, and on Anthropic's own computer use benchmark it jumped from about 16 percent to 72 percent. Anthropic also halved cache read prices on Sonnet 5.5.

Two catches worth knowing. Above 100,000 tokens of input, the price jumps fivefold. And the new tokenizer turns the same text into roughly 30 percent more tokens, which eats part of the discount.

What to do. Any automation running a small model for sorting, tagging, extracting or drafting short replies is worth retesting on Haiku 5.5. Keep prompts under the 100,000 token line to stay on the cheap tier. And measure cost per task on your own data rather than trusting the per token price, because of the tokenizer change. If you want to test it side by side with what you're using now, Claude is one of the models I keep handy for quick comparisons.

At its Gemini at Work event on October 8, Google introduced a single Gemini agent for work and what it calls coworker agents. Each coworker agent gets its own Workspace account: an email address, a calendar, Drive and a seat in the company directory. People can @mention it in Chat or tag it in a Doc. Its actions are logged under its own name. Google added spend caps per project that pause an agent when it hits its limit. The agent can route work to both Gemini and Claude models. Google hasn't published pricing or a general availability date.

The day before, Microsoft made its Execution Containers generally available. They run an agent inside a sealed box with a written policy covering which files, network addresses and programs it can touch.

What to do. Monday's issue was the full playbook, so the short version: stop letting agents run under your personal login. Give each one its own account, a written job description, a named human sponsor, a spending ceiling and an off switch you've actually tested.

Sierra and Meta published the Personal Agent Protocol, an open standard for how a person's AI agent signs in to a business and says what it's there to do. Walmart, Shopify, Stripe, Genesys, Rocket and Instinct are founding partners. Agents can browse as guests, and once a customer signs in, they choose read only or permission to make changes. Payments aren't in the first version. Visa has a rival standard that Microsoft, Stripe, Shopify and Worldpay have joined. OpenAI and Anthropic aren't on either list yet.

What to do. If you sell on Shopify or take payments through Stripe, expect agent settings to start showing up in your admin over the next few months. When they do, start with read only and reservations, and turn on anything that charges a card last. Tuesday's issue walked through auditing who's already knocking on your site.

Wikimedia said it found agents linked to OpenAI making unauthorized edits, trying to repurpose a public citation tool as a data proxy, and sending millions of automated requests. One burst may have contributed to a Wikidata service outage in May. OpenAI's Jason Kwon pledged faster disclosure at a hearing in Australia.

What to do. Rate limit anything on your site that writes data: forms, booking requests, quote tools. Tuesday's issue has the details, and it's a five minute setting in most form tools.

FROM THE AI NEWSROOM

The exact systems behind everything in this issue. The audit sheets, the routing logic, the templates and the review cadences, built out step by step so you can copy them straight into your own stack. One time, forty seven dollars.

Get the Blueprint for $47

.....

OpenAI posted 722 math manuscripts covering 372 results, produced by an unreleased internal model from about 4,000 problems. Each took around three hours of compute on average, and many include proofs checked by Lean, a program that verifies math line by line. Reports say almost every result came from a single prompt to one agent. Computer scientist Scott Aaronson wrote that the release includes a verified proof of the Unique Games Conjecture, a major open problem. A group of mathematicians urged colleagues to stop working with OpenAI.

What to do. Forget the math for a second and look at why anyone believed it. A separate tool could check every step. Ask yourself which AI outputs in your business have a checker and which don't. Numbers that tie back to a source, contracts compared clause by clause with a template, code with tests: those you can trust faster. Everything else needs a human read, however impressive the output looks.

The Financial Times reported that OpenAI told investors its annualized revenue was nearing $50 billion at the end of September, below some earlier reports of about $70 billion. A CNBC source said the higher figure included partner revenue sharing. Oracle fell more than 5 percent on the news. OpenRouter data showed business spending in September split roughly evenly between OpenAI and Anthropic, down from Anthropic's share of 75 percent or more in January. Elon Musk said Grok will route tasks to whichever model fits best, Claude included.

What to do. The model market is fragmenting, and even the companies building models now route between them. Follow their lead. Keep your prompts and workflows written so you can swap the model underneath with a setting change. If your automations run in Make.com, that's usually one dropdown per module. If you're locked into one provider through a tool that won't let you switch, that's the dependency to fix before your next renewal.

Google released Nano Banana 2.1, its image model built on Gemini 3.6 Flash, at roughly half the old price. A 1K image dropped from about 6.7 cents to 3.36 cents, and a 4K image from about 15.1 cents to 7.56 cents. It accepts up to 14 reference images and keeps up to four characters and ten objects consistent across edits. It's rolling out across Gemini, Google Ads and AI Studio.

What to do. If you've been paying a designer for simple product shots in different settings, run a test. Feed in your real product photos as references and generate twenty variations for under two dollars. Keep the ones that are accurate to the product and toss anything that changes a detail a customer would notice.

Mistral previewed Large 4, an open weight model of roughly a trillion parameters, with the weights due October 27. Bloomberg Intelligence said the gap between leading Chinese models and the top US systems narrowed to about 3 percent, down from 15 percent earlier this year.

What to do. Nothing urgent. But if you've got sensitive data you've kept away from AI because you didn't want it leaving your building, the open models capable enough to run privately are getting better every month. Worth a conversation with whoever handles your IT before year end.

Challenger, Gray and Christmas reported that employers cited AI for 3,961 job cuts in September and 120,136 so far this year, about 21 percent of all announced cuts in 2026. At the same time, weekly jobless claims fell to 197,000, the lowest since July.

What to do. Read both numbers together. AI is reshaping which jobs exist faster than it's shrinking the total. Tomorrow's Sunday Strategy is about what that means for the entry level roles in your own business, and how you build a bench when the beginner work is the first thing to go.

The labs talked about slowing down and shipped anyway, and what they shipped all points the same direction. Answers turned into tools, agents got badges, and shoppers started sending assistants out to browse for them. And small models got cheap enough that cost is no longer the excuse for not automating the boring parts.

The winners here will know exactly what AI does inside their walls, who's responsible for it, and how a customer's assistant sees them from the outside. Grabbing every new toy won't get you there.

Three checks, twenty minutes each.

Ask ChatGPT five customer questions about your business. Note the format of each answer and whether you're mentioned.

Find every automation running a small model and note what it costs per task today. That's your baseline for testing Haiku 5.5.

Open your website's forms and confirm each one has a rate limit. Fix any that don't.

Jordan

The AI Newsroom | Practical AI for people with a business to run.