1. The gap between using AI and delegating to it is widening
Andrej Karpathy, a founding member of OpenAI who later led AI at Tesla, wrote on X on October 11 that the gap in how people understand AI capability is widening. By his rough estimate, about 6 billion people have barely touched AI, and 1 to 2 billion use free chat products as a better search engine. Meanwhile, around 20 million professionals who use frontier models for math and code now see projects that once took weeks or months completed by AI agents from a single prompt. He added that executives and many other roles spend much of their day talking to people, so they experience the change only second-hand (original post).
Ethan Mollick, an associate professor at the Wharton School, wrote on October 1 that he had expected the work of dividing tasks among AI agents to need human design for some time, but that organizing work turned out to be one more thing AI can learn. He described a personal agent that contacted an airline on his behalf to ask for an extension of an expiring credit, and expects company customer service teams to be flooded with requests from such agents (original article).
Our view Both observations point the same way. The gap is not about whether a company can use AI, but whether it hands over multi-step work as a whole, such as drafting quotations or giving first replies to inquiries. Retailers, online sellers and trading companies will also need a policy for inquiries sent by their customers' AI agents. We think the right starting point for small businesses is not a company-wide rollout but one weekly task, handed to an agent together with its procedure manual, with a person checking the result. If the major Japanese online marketplaces have not published a policy on orders and inquiries placed through agents by March 2027, we will lower the priority of customer-desk preparation and revise this view.
2. Small models got cheaper, with conditions attached
On October 7, Anthropic released Claude Haiku 5.5, a small model aimed at high-volume work such as summaries, classification and database queries, and at speed-sensitive uses such as live customer support. The company says it costs around 75% less to run on average than Haiku 4.5. At the same time it halved the price of cache reads on Sonnet 5.5 (announcement).
Developer Simon Willison added the fine print on his blog the same day. Haiku 5.5 costs $0.10 per million input tokens and $0.50 per million output tokens, but once a single input passes 100,000 tokens the price rises fivefold to $0.50 and $2.50. He also found that the same long prompt counts as about 1.25 times as many tokens as on Haiku 4.5, which he calls a hidden price increase (original post).
On October 13, Google made its top model, Gemini 4 Argon, generally available through its API. The introductory price is $2 per million input tokens and $10 per million output tokens, and a footnote states that after the introductory period ends on November 30, 2026 the price becomes $4 and $20 (announcement).
Our view Headline prices fell, but the bill depends on three conditions: input length, how text is counted, and time-limited pricing. When a long document such as a contract or a product catalog is read in one go, Haiku 5.5 becomes five times more expensive per token past the 100,000-token line, which can cancel out the price cut. We think small businesses should not compare published rates, but should run about ten of their own real documents through each model and estimate a month of costs before choosing. If Google extends its introductory price beyond November 30, we will revise our view on what top-tier models will cost.
3. Set a spending cap that actually stops the service
In an October 3 blog post, Simon Willison argued that because AI agents make it easy for anyone to build programs that spend money, metered services should make hard budget caps the default: when the limit is reached, the service stops instead of sending a warning email. His reasoning is that most businesses would rather see errors than a surprise bill of more than $10,000. He also noted that on September 16 AWS began offering, to a limited number of customers, a monthly spend limit that pauses a project for the rest of the month once it is reached (original post).
Our view This applies to small businesses even more than to individual developers, because a company without a dedicated IT person may not notice an automated job running all night until the next morning. We think any business using AI APIs or cloud services should decide a monthly limit at setup and switch on a setting that stops processing at that limit before anything else. It should also decide how staff take over by hand when the service stops. If the major cloud providers make hard caps the default by March 2027, we will drop this from our required setup checks.
4. Japan's communications ministry opens a review of data protection
On October 16, Japan's Ministry of Internal Affairs and Communications announced a study group on data protection and security in telecommunications in the age of frontier AI, set up after a large data leak at a telecom operator. It will look at technical and organizational measures and at how regulation should work, given frontier AI and the shift of communications services to the cloud. The first meeting is on October 22, and conclusions are expected around summer 2027 (press release, in Japanese).
Our view The review targets telecom operators, but because it is a debate about regulation in a world of cloud services and AI, we see it reaching the data of small businesses as users through the contract terms of telecom and cloud providers. Companies that process customer lists or transaction records with outside AI services should list now which data goes to which service. With that list, they will know at once what to review when contract terms change. If the summer 2027 conclusions stay limited to operators' internal measures and put no new obligations on business users, we will revise this view.
Our outlook: when customers' AI places orders with companies' AI
According to Mollick, Meta's personal agent Muse was the number one app in the App Store in early October, and OpenAI has released a rival called dots (original article). We expect that within one to two years, personal agents like these will deal directly with companies' agents, handling inquiries, quotes, bookings and orders without a person on either side. Call it AI-to-AI consumer-to-business commerce: the customer's AI compares and chooses, and the company's AI answers on the spot. The companies chosen first will be those that publish prices, stock and delivery times in a form machines can read, and that answer an agent's question in seconds. Companies that can only respond by phone and email will not even make the shortlist. Preparing product data and automating inquiry handling each take months. A company that starts only after customer-side AI is commonplace will lose business to competitors during those months. We judge that now is the time to lay the groundwork.
Our overall view
Based on the first half of October, we judge that AI has entered a stage where unit prices are falling and differences in how companies use it turn directly into differences in their operations. The main drivers are agents that can take over multi-step work as a whole, and price competition from small models up to top-tier ones. For the next 2–4 weeks, we think the most effective step for a small business is not to chase new models but to pick one recurring task to delegate, and to decide in advance a monthly spending cap and the scope of data sent outside. One condition would change this view: if other major providers follow the December 1 price change for Gemini 4 Argon with price increases of their own, the cost assumptions change, and we would narrow our advice to delegating only high-frequency tasks.
Questions this article answers
Is Claude Haiku 5.5 really cheaper?
Anthropic says it costs about 75% less to run on average than the previous model. But once a single input passes 100,000 tokens the price rises fivefold, and the same text counts as about 1.25 times as many tokens. Companies that handle long documents should estimate a month of costs on about ten of their own documents before choosing.
How can we stop AI usage fees from running out of control?
Decide a monthly limit at setup and switch on a setting that stops processing when the limit is reached. Since September 16, 2026, AWS has offered some customers a monthly limit that pauses a project for the rest of the month. Also decide how staff take over by hand when processing stops.
What should a business prepare for inquiries and orders sent by customers' AI agents?
Publish prices, stock and delivery times in a form machines can read, and set up a way to answer agents' questions instantly. Companies that only respond by phone and email drop off the shortlist at the comparison stage. Preparing product data and automating inquiry handling each take months, so we think the work should start now.
Sources
- Andrej Karpathy, post on X (2026-10-11)x.com
- Ethan Mollick, The Dot and the Swarm, One Useful Thing (2026-10-01)oneusefulthing.org
- Anthropic, Introducing Claude Haiku 5.5 (2026-10-07)anthropic.com
- Simon Willison, Claude Haiku 5.5 (2026-10-07)simonwillison.net
- Google, Gemini 4 Argon is now available for everyone (2026-10-13)blog.google
- Simon Willison, We’re going to need default hard budget caps on pretty much everything (2026-10-03)simonwillison.net
- 総務省 報道資料（2026-10-16）soumu.go.jp
Related services
If you want to try this in your own business
To see how the topics in this article would apply to your own operations, and what it would cost, send us a message through the contact form.