OpenAI Operator, Microsoft Autopilot, and Google Live Avatar: What’s Driving the AI Industry in Late September 2026

OpenAI accidentally revealed Operator — the agent that never sleeps

On September 25, user Jake Box posted a screenshot from the ChatGPT subscription interface. Among the benefits of the $100 Pro plan, it showed the line “Your always-on assistant” — named Operator. The post received 620 thousand views in a few hours, and OpenAI made no official comment.

Here’s what is known so far: a configuration named “o1” with an email suffix bearing the same name was found in a testing catalog. If the agent has an email identifier, it’s no longer just a chatbot but something closer to an employee within the system. In theory, Operator could monitor events, perform scheduled tasks, and return to work without the user having to reopen the conversation.

There are also rumors of a ring-shaped device costing more than $300 and a $500 plan, but OpenAI has not officially linked Operator to any of them. In fact, there are two “clues” — the description on the subscription page and the internal configs — but there is still no working product.

On September 29, OpenAI held a developer meetup in San Francisco featuring Sam Altman, but Operator was not demonstrated there.

Microsoft is no longer teasing — Copilot Autopilot is already working

While OpenAI is hinting at what’s to come, Microsoft gave a detailed look at its enterprise Copilot on September 25, with three modes: Home, Code, and Autopilot (formerly “Scout”). The difference is fundamental.

Autopilot is an agent that is assigned a name, role, and goal. It then monitors channels on its own, reads threads, handles repetitive tasks, and returns to a project days later. In the demo, the agent checks suppliers, creates schedules, prepares meetings, and messages participants itself to request updates. It lives in the corporate cloud environment, with its own memory, computer, and workspace. Colleagues can mention it almost like another employee in Teams.

Home combines ordinary chat with long-term tasks. Unfinished items, suggestions, and recent actions are always visible. Code, powered by GitHub Copilot, lets users create trackers, dashboards, and internal apps using natural-language prompts — all code runs in an isolated environment.

As for pricing: the standard subscription covers regular chat and work in Office. Cowork, Autopilot, and long-running agent tasks consume additional tokens. Administrators can set spending limits and restrict access to specific model families.

Google: a live face for AI

Google introduced Gemini 3.8 with the Live Avatar feature — a video avatar that responds in real time, synchronizing speech, lip movements, and facial expressions. It processes video and audio simultaneously, meaning it can see what is happening on the screen and hear the other person at the same time.

It supports 97 languages with synchronization and no noticeable delays. In a hotel demo, the avatar handles guest check-in while simultaneously interacting with external systems — the conversation does not stop while the system waits for a response. Companies can use ready-made avatars or create their own from photos. Custom avatars are currently available only to Enterprise customers. All generated content receives invisible SynthID watermarks.

Tesla Optimus: employees refused to help robots replace them

According to The Information, Tesla encountered an unexpected problem while training Optimus. To collect data on human movements, employees were asked to wear motion-capture suits and perform ordinary tasks. Some refused. The logic was simple: they did not want to be the first to train the robot that would later take their jobs.

After the refusals, Tesla moved data collection to separate dedicated teams. At the same time, there are production problems: the Optimus Gen 2 has more than one hundred screws and small components in its hands alone, and most parts are installed manually. In August 2026, Tesla assembled only a few hundred units. The goal of one thousand robots per week by the end of the year still looks overly ambitious.

Meanwhile, the New York startup Midjourney (not the one that makes images — a different company) announced an archive containing 2 million hours of human behavior: 20 thousand types of tasks, 50 different environments, and 90% video with clearly visible hands. It also includes 3D hand tracking, depth maps, and action descriptions. There are another 50 thousand hours of gameplay and 69 thousand hours of conversations in different languages. All of this is intended to train humanoid robots in simulation before they enter the real world. The company raised 15 million dollars in a seed round for this.

What else happened this week

• The price war continues. Analysts from Artificial Analysis report that budget smartphones priced under $200 will disappear by roughly 40% by 2030, as AI companies buy up the same memory chips needed by manufacturers of low-cost phones.
• OpenAI + Stargate. A partnership involving 750 MW of computing capacity is being discussed, with a gradual rollout through 2028. Testing Catalog found new tiers of platform access, including “Accelerate” for production workloads.
• ChatGPT Ultrafast mode. A hidden Standard / Fast / Ultrafast switch was found in the code. Ultrafast, powered by Cerebras, reaches up to 750 tokens per second — roughly 14 times faster than Standard. For now, it is available only to selected customers.
• Google Photos — a virtual wardrobe. AI automatically collects clothes from your photos into a single “digital wardrobe,” sorted by category. Launched in the US, Brazil, and India.
• Microsoft AI — a draft code of conduct. The “Humanist AI” document states that people are more important than AI, and that an agent cannot resist being shut down, expand its authority, or set goals for itself. Open for public discussion for 6 weeks.
• Suno is back in court. Sony and Universal Music Group filed a new lawsuit — this time against the Suno V6 model. The accusation: “laundering data through AI,” when a new model learns from the output of an older one that was trained on protected recordings.

In conclusion

In a single week, the three largest players all moved in the same direction: an agent with persistent memory and working context — not a one-off chat, but something like a permanent colleague. OpenAI is still at the leak stage, Microsoft is already in beta, and Google is approaching the idea through a video avatar. The question is not whether such agents will exist — it is how quickly they will learn to work for hours without human involvement, without losing context or doing foolish things.

The story of the Tesla employees is a good metaphor for the moment: the people who are best at explaining to a machine “how it’s done” are often the first to realize that they are training their own replacement.

This analytical review was prepared by the editorial team at seo-mind42.ru based on data from specialized research groups, vendor documentation, and industry reports.