We may earn money or products from the companies mentioned in this post.
AI business news for September 28–October 5, 2026.
I’m interested in AI for a pretty ordinary reason: I want to build useful things without needing a huge team or spending a fortune getting started.
I’d like a win. You probably would, too.
But there’s a problem with following this stuff. Every announcement sounds like the one you can’t afford to miss. Spend enough time reading them and you can end the week with six new subscriptions and the same unfinished business idea.
There’s another problem. Plenty of the people we’d like to sell to are tired of hearing about AI. They’ve seen the fake pictures, read the lazy articles, or fought with a chatbot when they needed an actual person.
I don’t think a customer owes us enthusiasm for the technology. We have to give them a reason to care about the thing we’re selling.
That’s how I’m looking at this past week. A few announcements could help us get something built for less. Others are a reminder to check whether somebody still needs what we’re planning to sell.
The week started with a reason to check the bill
On September 28, Anthropic released Claude Sonnet 5.5. The company reports more than 30% faster output and up to 30% lower cost per task than Sonnet 5. The token prices stayed the same. Its explanation is that the newer model generally needs fewer tokens to get the work done.
Those are Anthropic’s results. I’d want to see what happens on the particular job I need it to do before building a budget around them.
Still, this is the kind of announcement that gets my attention.
Suppose you’re thinking about a small tool that turns a contractor’s job notes into an estimate they can review. The contractor has the notes and photos. They’re just tired of doing paperwork at night.
If the model does that job well for less money, you have more room to price the product sensibly. You might also have enough room to include the checks that make it dependable.
But if it leaves a detail out and the contractor has to go through everything again, the cheap run wasn’t much of a bargain.
I’d gather ten sample jobs, with permission and private information removed, and compare the output against what a good finished estimate should contain. Count the retries and corrections along with the model bill.
What would change my mind is a lower cost for work I can actually use.
Then OpenAI made the same week cheaper and more complicated
At its September 29 DevDay event, OpenAI introduced GPT-6.1 Sol at one-fifth of Astra’s standard input and output token prices. OpenAI describes it as approaching Astra’s performance.
That gives builders another model to test. I wouldn’t move a working product over just because the announcement looks good. I would put it through the same sample jobs and see whether the saving holds up.
At the same event, OpenAI introduced Dots, agents designed to handle ongoing responsibilities, initially available through selected paid plans.
That’s where I’d stop and look at the other side of the business.
If I’m planning to sell a general assistant that researches things, writes emails, and organizes work, what happens when the customer’s existing subscription can do enough of that already?
I’d rather answer that question before spending three months building.
For the estimate example, I’d try the job in the general tools first. See what they get right. See where the contractor still has to explain things, copy information around, or catch missing details.
Maybe there’s a useful business in handling those awkward parts. Maybe the general tool is already good enough for that customer.
Either answer is valuable. Finding out an idea is weak before you spend much on it is a win, even if it doesn’t feel like one that afternoon.
A plugin might let you try something smaller
OpenAI also announced interactive plugin extensions and improvements to plugin creation, submission, and discovery.
This interests me because a standalone app can become a fairly large project before anybody has agreed to pay for it.
You start with one useful job. Soon you’re working on accounts, billing, dashboards, passwords, and a settings page. The customer still hasn’t seen the part they came for.
A plugin might let you test a smaller version inside a tool people already use. You’d need to check what the platform supports, its rules, and how you could charge. Being in a directory doesn’t guarantee customers.
But I like the possibility of getting to the actual test sooner.
Can someone use this on their own work? Do they want to use it again? What would they pay to keep it?
Those answers are worth more to me than having every screen finished.
Shopify gave store builders something to think about
On October 1, Shopify announced Canvas, a workspace for designing store pages with help from Sidekick.
If I were considering an AI business that builds stores, I’d take a close look at that before choosing my offer.
It doesn’t tell us every merchant can now build exactly what they want. It does give us another reason to check what the customer can already do themselves.
And I’d spend some time asking what happens after the store opens.
Where do returns get messy? Which customer questions keep coming back? What supplier information needs fixing before a product can go live?
I wouldn’t assume any of those is a business just because I can imagine an AI tool for it. I’d ask a store owner to show me a recent example and how they dealt with it.
A person walking you through a real problem can save you a lot of guessing.
I wouldn’t lead the sales page with AI
For that contractor tool, I’d be more interested in a promise like: “Turn your job notes into an estimate you can review before dinner.”
You’d have to prove it. You’d also need to explain what the tool handles, what information it uses, and what the contractor must check.
But at least that gives the customer something concrete to consider.
I wouldn’t hide the AI. I also wouldn’t expect someone to buy because I’d put it in the headline.
If they’re skeptical, I’d want them to try it on a real job. Let them judge whether it helps. If it makes a mess, their skepticism was justified and I have work to do.
What I’d take into this week
If you already have a product, the new models are worth testing against the work it actually does.
If you’re choosing an idea, check the tools your customer already pays for. Then find someone willing to show you the problem you think you can solve.
If you’ve been building for weeks, get the useful part in front of a prospective customer before adding another feature.
That’s my read on the week: better options for getting work done, and more reason to be specific about whose work you’re helping with.
I’d be happy to finish the next seven days with a rough version that somebody wants to use again. That would tell me a lot more than another folder full of ideas.
If you’re still deciding where to start, our guide to AI businesses you could build gives you some options to investigate. For choosing the tools, see our practical small-business shortlist.
Sources: Anthropic’s Sonnet 5.5 announcement, September 28; OpenAI’s DevDay recap, September 29; Shopify’s changelog, October 1. Performance claims above are attributed to the companies announcing them. The business examples are illustrative; they are not reports of products we tested.