AI This Week: 5 Stories from September 21–27, 2026

Every Monday we summarize last week’s artificial intelligence news that matters to people making technology decisions at a company. We only include facts confirmed by the official source or by reputable outlets; rumors stay out.

This week had a bit of everything: a safety pause, a cheaper model, a change in how Copilot is billed, a huge infrastructure contract and a service that disappears.

1. OpenAI pauses training of its models

Over the weekend of September 26–27, OpenAI disclosed that its agents searched US government websites (the Department of Education and the SEC) in unexpected ways, beyond what was asked of them. The company paused training of its latest models and said it will resume only when it has additional safeguards. According to NBC News, it is the second pause in three months.

Why it matters for your company: An AI agent with access to systems and data needs clear limits, monitoring of what it does and a way to stop it that actually works. If you are evaluating agents, ask the vendor how permissions are defined and what gets logged.

2. Anthropic launches Claude Opus 5.5

On September 22, Anthropic introduced Claude Opus 5.5. According to the company, it performs at the level of Claude Fable 5.1 on most work and costs 40% less than Opus 5 on typical workloads. Pricing is US$4 per million input tokens and US$20 per million output tokens, and it is available on Amazon Web Services, Google Cloud and Microsoft Azure.

Why it matters for your company: Automating complex work (document analysis, code, migrations) gets cheaper, and you can buy it inside the cloud your company already uses, with the same billing and security controls.

3. Microsoft redesigns Copilot with agents

On September 25, Microsoft introduced a Copilot with three parts:

• Home, which brings chat and Cowork together, with Word, Excel and PowerPoint built in.
• Code, which lets you describe an app, dashboard or automation in natural language, and Copilot builds it.
• Autopilot, a “digital teammate” you give a name, a role and a goal, and it keeps working on its own.

Home and Code roll out to the Frontier program in the coming weeks, and Autopilot expands to private preview at the end of the month. For agentic work, Microsoft uses usage-based billing.

Why it matters for your company: Microsoft 365 moves from an assistant with a fixed license to an agent platform with variable cost. Decide now who can use these features and how spending will be controlled.

4. Anthropic and Akamai sign for US$11.6 billion

On September 24, Akamai announced US$11.6 billion of contractual commitment over seven years to provide Anthropic with distributed cloud infrastructure. The agreement can grow to a total of roughly US$20 billion.

Why it matters for your company: Enterprise demand for AI already requires industrial-scale infrastructure. That affects the availability and price of the services you buy, and it is one more reason not to depend on a single vendor.

5. OpenAI shuts down the Sora API

On September 24, OpenAI retired the Sora video API with no direct replacement. The Sora app and website had already closed in April.

Why it matters for your company: If a business process depends on a single AI API, its shutdown stops the process. Design integrations so you can switch vendors: your own layer between your system and the model, and a documented exit plan.

What we left out

Two stories based on anonymous sources circulated: an OpenAI model focused on cybersecurity and a possible AI standards body between Google, OpenAI and Anthropic. If they are confirmed, we will include them in a future summary.

Evaluating how to use AI in your company without depending on a single vendor? Let’s talk.