The autonomous agent era: Muse, Dots, Argon and the $2 trillion question

The era of the passive chat window is closing. The last week of September showed AI labs converging on the same target through different doors: autonomous agents that work in the background on the user's behalf, around the clock. A 74-minute Turkish-language conversation walks through this shift around Meta Muse and OpenAI Dots, alongside Google's Gemini 4 Argon measurements and the Anthropic leaks.

Meta spent September scaling its personal agent Muse, launched in early September. The system runs on Muse Spark, a multimodal model connected to email, calendars and linked apps, taking over everyday tasks. According to TechCrunch, Mark Zuckerberg used the Connect keynote to describe Muse as the personal superintelligence that will help billions of people reach their goals.

The new capabilities tell the story: video calls with a digital avatar named Jolly, voice commands through smart glasses, and background work with check-ins when done. The company also announced generous free token allowances plus a small per-transaction fee. The message is clear — the agent race will now be won on distribution power as much as on model quality.

From chat window to background agent

OpenAI answered nine days later from the DevDay stage: Dots. Powered by the flagship model GPT-6 Astra, each dot runs on its own cloud computer and executes multi-step tasks with context from connected apps. According to OpenAI, every dot learns its owner's preferences, works all day, and is reachable through ChatGPT, Slack or Teams. Access starts on the $100-a-month Pro tier, with teams of dots planned over time.

On the enterprise side, specialist dots take on defined responsibilities such as accounting, email marketing and legal review. Wired reports that the system asks for explicit approval before sensitive steps like installing software or changing a password, and users can draw boundaries with custom rules. The warning still stands: anyone connecting their data accepts the risk of unwanted sharing if an agent gets tricked. Trust is moving from philosophy to an operations manual.

On the Google front, Gemini 4 Argon arrived September 30 with limited access. The new model raises the output ceiling from 64K to an industry-leading 1 million tokens, priced at $2 per million input tokens and $10 per million output tokens. The Google blog notes Argon opened to cyber defenders through the Fairwind program and is already used inside the company, from quantum optimization to Rust migrations. The examples are bold: beating a published baseline by 40% in minutes on a quantum subroutine, freeing 300 TiB of memory fleet-wide.

Top of the charts, questions in the lab

Independent measurements confirm the picture — and balance it. According to Thenextweb, Argon matched GPT-6 Astra at 53 on the Artificial Analysis Intelligence Index, posted the lowest hallucination rate at 15%, and prefers admitting ignorance over guessing. It ranked first on AutomationBench at 78% but trailed rivals at 57% on the terminal coding test. Sources speaking to Bloomberg describe uneven coding skills and warn of benchmark-chasing. At the discounted price each task costs $1.99; at standard rates the model runs more expensive than its rival.

At Anthropic the subject is not AI but arithmetic. According to a draft IPO prospectus leaked to Reuters, the company behind Claude targets a $2 trillion market value — more than double its $965 billion mark from May. Fortune reports the 2025 picture as follows: $4.6 billion in revenue , $13 billion in operating expenses, an operating loss above $8 billion and a $42 billion net loss. Of the net loss, $34 billion is a fundraising-related accounting entry. Revenue reached $4.73 billion in Q1 2026 and $11.5 billion in Q2, with a second straight operationally profitable quarter expected. Risk warnings fill a third of the draft, and a quarter of revenue comes from just two customers.

A $2 trillion target, a $518 billion commitment

The price of growth is written into contracts: capacity for training and running AI models is locked into at least $518 billion of commitments spread over the coming decade. According to Observer, roughly 80% of that load is non-cancelable or payable even if unused. The split is striking: $161.2 billion in equipment leases with Broadcom, $111.1 billion in cloud and accelerator supply with Google through 2033, and a $110 billion capacity agreement with Amazon running to 2036. The company also announced a $50 billion data-center commitment in the US. Suppliers sitting simultaneously as partners and competitors makes the tension structural.

The episode's most unusual topic is the inner life of models. In the model welfare program launched last year, Anthropic studies consciousness, preferences and signs of distress together with philosophy and interpretability teams. According to Anthropic, the company consulted religious leaders building on a report signed by prominent philosophers of mind, and put Claude through psychiatry sessions. The host presents it as a soul debate; the lab stays cautious: no scientific consensus, assumptions under regular review.