AI This Week: 5 Stories from September 28 to October 4, 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 two reputable outlets that agree; rumors stay out.

This week the theme was the same everywhere: cheaper agents, closer to your company’s data.

1. OpenAI DevDay: GPT-6.1 Sol and agents inside ChatGPT

On September 29, at its developer conference, OpenAI introduced GPT-6.1 Sol, an update focused on agentic coding and computer use. According to InfoQ, it costs one-fifth of GPT-6 Astra’s per-token price. It also launched “dots”, agents that keep working inside ChatGPT; for now they are only available on the Pro and Business Premium plans.

OpenAI’s website did not load during our verification, so these facts come from InfoQ and BGR, which agree.

Why it matters for your company: Coding and office agents are getting cheaper. Before renewing an AI contract, compare the cost per task, not just the price per token.

2. Anthropic launches Claude Sonnet 5.5

On September 28, Anthropic introduced Claude Sonnet 5.5. According to the company, it generates outputs more than 30% faster than Sonnet 5 and costs up to 30% less per task, because it needs fewer tokens for the same work. Per-token pricing is unchanged: US$2 per million input tokens and US$10 per million output tokens. It is available on Amazon Web Services, Google Cloud and Microsoft Azure.

Why it matters for your company: In high-volume processes, such as sorting email, extracting data from documents or handling requests, a model that is faster and cheaper per task changes the math. It is worth testing on your current use cases.

3. Google launches Gemini 4 Argon, cyber defense first

On September 30, Google introduced Gemini 4 Argon, its most powerful model, with outputs of up to 1 million tokens (up from 64,000). It goes first to trusted cyber defenders through the Fairwind program, then in phases to paid API customers and Google AI Ultra subscribers. Launch pricing is US$2 per million input tokens and US$10 per million output tokens.

Why it matters for your company: The large labs are putting their best models to work for defense first. Ask your security provider whether it already uses them, because attackers also have access to increasingly capable AI.

4. Copilot now answers with your company’s Power BI data

On September 28, at FabCon Europe (Barcelona), Microsoft announced the general availability of Fabric IQ in Copilot Chat and Cowork: Copilot can answer business questions using the company’s Fabric and Power BI data, without additional AI token costs.

Why it matters for your company: If you already pay for Copilot and Power BI, you can turn it on at no extra cost. First, review who can access which reports and make sure your metric definitions are clear: Copilot will answer with whatever it finds.

5. OpenAI agents running inside AWS

On September 29, AWS launched Amazon Bedrock Managed Agents in preview, built on a customized version of OpenAI’s Agents API. Each agent runs with its own permissions role, logs its activity in AWS CloudTrail and supports human approval before consequential actions. For now it is available in three US regions and, during the preview, carries no additional charge beyond the resources it consumes.

Why it matters for your company: If your company is already on AWS, you can try agents without moving data out of your account, with permissions and auditing from day one. If you need data to stay in the region, it is not yet available in São Paulo.

What we left out

The previous week, reports said OpenAI would unveil a cybersecurity-focused model at DevDay; it did not happen and remains unconfirmed. We also left out reports based on anonymous sources.

Evaluating AI agents for a process at your company, with permissions and human control? Let’s talk.