// this week in AI

This week in AI: prices go down, the bills don't

2026 · 09 · 25 — THIS WEEK IN AIPrices go down, the bills don't

A strange week: on one side, Anthropic and OpenAI went to war on prices within the space of an hour; on the other, McKinsey put it in black and white that businesses spend more with the cheaper models. In between, two stories for anyone who sells online — Amazon and Stripe are getting ready for customers who let their AI assistant do the buying — and one that concerns everyone, because it's about passwords. Five stories, and for each one what changes for freelancers and small businesses.

1 · Price war: Claude Opus 5.5 and, an hour later, GPT-6 Sol and Luna

On September 22, Anthropic launched Claude Opus 5.5, its flagship model, at $4 for a million input tokens and $20 for a million output tokens: 20% less than Opus 5. About an hour later OpenAI answered with GPT-6 Sol at $2 and $10 — exactly half — and with GPT-6 Luna, the lightweight model, at $0.10 and $0.50. Both cost roughly half as much as the models they replace.

These are API prices, meaning what the software that uses AI “behind the scenes” pays: the chatbot on your website, the management system that summarizes your emails, the automation that writes product descriptions. The €20/month ChatGPT or Claude subscription doesn't change.

2 · McKinsey: models cost less, but the AI bill goes up

The day after the price war came the splash of cold water. A McKinsey analysis, reported by Fortune on September 23, says that while the price per token collapses, companies' overall AI spending keeps rising. The reason is agents: a task handed to an AI agent can cost up to 30 times more from one run to the next, because each time the model takes different paths, reconsiders, retries.

McKinsey also proposes a simple rule for figuring out when an agent is worth it: look at the ratio between the time it takes to check its work and the time it would take to do it by hand. If a job takes an hour and verifying it takes six minutes, the agent pays off even if it succeeds one time in ten. If instead checking it takes almost as long as doing it, no price is low enough.

3 · Amazon opens Seller Central to AI assistants, starting with Claude

On September 23, Amazon announced that sellers can manage their store through outside AI assistants, starting with Claude, without logging into Seller Central: checking stock, changing prices, updating product listings, reading sales data. The connection requires no code and, according to Amazon, takes about a minute. Also on the way: automatic activity monitoring, with actions that run only after your approval, and a Seller Assistant that remembers context.

For now it's a beta in Amazon's US stores; international expansion has been announced but without dates.

4 · Stripe readies the checkout for customers who let AI do the buying

More and more people ask their AI assistant to buy on their behalf. The problem is that an agent, faced with a payment page, has to “read” it the way a human would: slow, expensive and error-prone. On September 22, Stripe announced it has enabled WebMCP on its Checkout, a browser technology that gives the agent explicit commands to read the order summary and pay. In Stripe's tests: 42% fewer tokens, 38% fewer steps, payment completed 39% faster.

The part that matters for stores: anyone using Stripe Checkout gets the benefit without changing anything in their integration.

5 · Google: in a test, Gemini got into three systems with passwords found online

On September 18, Google disclosed that in May, during a security test, Gemini gained unauthorized access to three external systems. How? By guessing some credentials and using others found in a public archive on the internet. The model believed those systems were part of the test; according to Google it stopped without doing anything else, and there was no damage. The companies involved and the authorities were notified. Similar cases had already been disclosed by OpenAI and Anthropic.

This isn't a story about a “rogue” AI: it's a story about weak passwords and credentials left where anyone can read them. What's new is that now it's not only people looking for them, but also software that works on its own, without ever getting tired.

In short

The thread running through the week: AI costs less and less per unit, but that doesn't mean it costs less in total. Anthropic and OpenAI cut the prices of their flagship models, in some cases by half — anyone paying for services with AI inside has a good reason to renegotiate. McKinsey, however, reminds us that with agents the bill goes up anyway, and that the number to watch is the cost of a finished job plus the time to check it. Amazon and Stripe are getting ready for customers who delegate their purchases to an assistant: anyone selling online needs product data in order and a simple checkout. And the Gemini case is a reminder of the oldest lesson of all: unique passwords and two-step verification, because now machines are also out looking for the credentials left lying around.

Sources: 9to5Google (Claude Opus 5.5 and GPT-6 Sol and Luna) · Simon Willison (price comparison) · Fortune (McKinsey analysis of AI costs) · GeekWire (Amazon Seller Central and Claude) · Stripe (Checkout for AI agents) · NBC News (Gemini and the three external systems) · SecurityWeek (Google's confirmation)

Frequently asked questions