AI tools are becoming more numerous, cheaper and more controlled by you

This week's updates point in several directions at once. The AI tools can do more, they are coming into more programs, and at the same time clearer lines are being drawn for who gets what.

From answers to doing something

OpenAI has launched Dots, agents that keep working between conversations, while Google's Gemini has gained agent features in Gmail, Calendar and Android apps. An agent is simply an AI that carries out tasks on its own rather than only answering questions. The Gemini app has also gained a maps tool, and Google is testing gathering commands under @ instead of slash. At OpenAI, Plus and Pro subscribers can now log in to 16 external tools with their ChatGPT account. And Anthropic is testing AI code review in Claude Code, where several agents go through code for errors.

It is convenient, but it also requires that you dare give AI access to your mail and your calendar. That is a choice you should make consciously.

Creative AI is broadening

Adobe has made music, voices and sound effects available to all Firefly users at no extra charge, and Suno can create speech with background music for, say, bedtime stories. Black Forest Labs has launched FLUX 3 Image with images up to 4K. Tavus says its model Griffin was mistaken for a real human by 26 out of 54 test subjects in video calls. That is the company's own test, and the model is not yet open to customers, so the figure should be taken with caution. But it is a reminder that you cannot always trust who or what you are talking to.

Data, voices and open models

Anthropic is now asking whether your voice may be used for training. The choice is voluntary and turned off by default. At the same time, companies in India can have Claude requests processed in India. Both are about where your data ends up.

On the open side, DeepSeek is releasing a model that takes up less memory, and Germany's Aleph Alpha has released Kolibri with 1 million tokens of context, aimed at authorities and regulated industries that want to run AI on their own hardware. A token is a small piece of text, and context is how much the model can take in at once. Reflection AI is reportedly set to follow soon, but there is still no name, date or test.

Not everything is getting better for everyone

From October 9, 2026, free users in Gemini will be moved down to the lightest model, AI Plus loses the Pro model, and AI Pro gets Deep Think. The best models are thus moving up in price. That pulls in a different direction from the free, open models.

A quick bright spot for professionals: a Harvard physicist had Claude help with 36 drafts of scientific papers in three months, but the work often only became valuable when experts stepped in.

What you can do

Check your voice sharing settings in Claude, and review which accounts agents may see. If you use free Gemini, be ready for a weaker model from October 9. Finally, keep an eye on Gemini 4, which Google hopes to show in an early version before the end of 2026.