AI This Week: 5 Stories from October 5–11, 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’s theme: agents with their own identity, and much cheaper small models.

1. GPT-6 reaches ChatGPT, business plans included

On October 7, OpenAI began rolling out GPT-6 in ChatGPT. According to its system card, it replaces GPT-5.6 Sol and GPT-5.6 Luna across all plans. GPT-6 Sol reached Plus, Pro, Business and Enterprise first, and GPT-6 Luna reached the free plans the next day, according to 9to5Mac. The model comes with Intelligent UI: answers can include charts, buttons, forms and small interactive tools, such as a calculator. Codex and ChatGPT Work still use the previous versions for now.

OpenAI’s website did not load during our verification; these details come from the system card OpenAI published and from 9to5Mac.

Why does it matter to your company? If your team uses ChatGPT Business or Enterprise, the model changed without anyone doing anything. Rerun your tests on the uses that matter, such as customer replies or contract summaries, before trusting that everything works the same.

2. Anthropic launches Claude Haiku 5.5, its cheapest small model

On October 7, Anthropic introduced Claude Haiku 5.5, its small model for high-volume work. According to the company, it costs around 75% less per task than Haiku 4.5 on average. The input price drops from US$1 to US$0.10 per million tokens for requests up to 100,000 tokens. Asana reports over 30% lower latency in its tests. It is available on Amazon Web Services, Google Cloud and Microsoft Azure.

Why does it matter to your company? Sorting email, extracting invoice data or tagging tickets are high-volume tasks where the cost per task decides whether a project pays for itself. At these prices, it is worth redoing the math on projects that were dropped because of cost.

3. Google unveils the Gemini agent for work

On October 8, at Gemini at Work 2026, Google introduced the Gemini agent: a single agent that answers questions, handles office work, creates content and writes code. It works inside Gmail, Drive, Docs and Sheets, and also in Microsoft 365 and Slack. It picks the model for each task, including Gemini and Anthropic’s Claude. Each agent has its own identity and least-privilege permissions, every action goes into an audit trail, and administrators can set a spending cap per project. The financial services and legal editions are in preview; Google did not publish pricing.

Why does it matter to your company? Agents no longer act under an employee’s account: they have their own identity. Before turning them on, decide what data each one can see and who reviews its activity log.

4. Mistral launches Large 4, trained in Europe with open weights

On October 6, Mistral AI introduced Mistral Large 4, a 1 trillion-parameter model that reads text and images. It was trained in its own data centers in Europe, on more than 160 languages, including every official language of the European Union. For now it is in preview on its API, at US$1.36 per million input tokens and US$4.18 per million output tokens. Mistral says it will release the weights by the end of the month, after safety testing with cybersecurity experts and authorities; the license has not been published yet.

Why does it matter to your company? A large open-weight model can run on your own servers or in your cloud, without sending data to a third party. If your company has data that cannot leave its infrastructure, it is an option to evaluate once the weights and license are out.

5. An open standard for personal agents to talk to companies

On October 6, Sierra and Meta introduced the Personal Agent Protocol, an open standard that defines how customers’ personal AI agents connect to companies. The customer decides whether their agent can only look things up or can also make changes, and the company sets limits on what the agent can do. Genesys, Rocket, Shopify, Stripe and Walmart are among its backers. The first version of the specification is due at the end of October.

Why does it matter to your company? Soon some customers will not reach you through your website or WhatsApp, but through their own agent. There is nothing to build yet, but it pays to have your services behind a clean API with clear permissions.

What we left out

Reports of a new DeepSeek funding round circulated, based on anonymous sources; we will include it once it is confirmed. We also left out an alleged Microsoft “Copilot Nucleus” that does not appear in any official source.

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