Google Unveils Gemini 4 Argon and Launches Space Data Center Test as Personal-Agent Race Intensifies

Argon, unveiled on Wednesday, promises major advances in coding, cybersecurity and complex tasks, according to CNBC. Industry benchmarks cited by the network show Argon tying OpenAI on a key cybersecurity test and posting leading results in software engineering. Its introductory pricing of $2 per million input tokens and $10 per million output tokens matches OpenAI's newly discounted GPT-6.1 Sol model. A token is equivalent to about three-quarters of one word.

While Google steps up its model battle, rival Meta is racking up millions of downloads for its Muse app, which launched last month, soared to the top of Apple's App Store and topped ChatGPT, CNBC reported. Earlier this week, OpenAI rolled out Dots, its personal agent offering, capitalizing on demand from businesses and consumers for tools that can carry out tasks such as scouring emails, booking travel and organizing expenses on a user's behalf. As of Sept. 30, Muse had reached more than 5 million downloads, according to Sensor Tower.

Attention at the frontier model level has also turned to safety. Anthropic CEO Dario Amodei sparked a firestorm three weeks ago by urging the top AI labs to slow the pace of development as concerns intensify about their potential dangers, CNBC reported. Over the past three months, Alphabet's stock is down about 6%, while Meta is up 19%, sparked by September's rally, which was the sharpest for any month since 2022.

Analysts at JPMorgan Chase wrote in a note on Thursday that Google needs a significant advance in its personal agent offerings to generate consumer enthusiasm. Bank of America identified the potential for Gemini 4 to strengthen Google's cloud business and existing products, while providing a foundation for a future personal agent. In the consumer market, Google has a massive home-court advantage: billions of people use Google products including Gmail, Calendar, Chrome and Search, and their messages and schedules are already stored within Google's ecosystem, potentially giving the company a substantial edge in developing an assistant capable of working across those services.

Yet Google's personal agent, Spark, remains limited to paying subscribers, while Muse is free with usage caps. Launched in May at Google's I/O developers conference, Spark can work across Gmail and Calendar, navigate websites through Chrome and complete tasks such as filling out online forms. It is also available through Google's mobile and desktop apps. Unlike Muse, Spark cannot make outbound phone calls or complete purchases on a user's behalf. Instead, it takes users through the purchasing process before handing control back to them for final approval.

Meta's approach has also introduced complications. Reuters reported in September that the company had experimented with a human concierge system in which contractors handled some phone calls when Muse could not complete them autonomously. Internal privacy concerns prompted Meta to suspend the experiment, while its automated phone-calling feature remains in beta. The episode has raised privacy concerns associated with granting AI assistants access to so much sensitive personal information. Meta did not immediately respond to a request for comment.

Google is now evaluating whether its latest frontier model could strengthen Spark. In an interview with CNBC, Tulsee Doshi, head of product for Gemini, highlighted Argon's ability to tackle complicated assignments that require multiple steps and can run for extended periods. She said Google is still evaluating where Argon can be deployed most effectively. Personal agents do not need the most powerful frontier models to complete everyday tasks. Prior to Argon, Google spent much of the summer releasing cheaper, lighter models designed to handle routine requests at scale. Doshi said those models remain essential for everyday agents, while more powerful frontier models can provide the additional reasoning capabilities required for complex assignments.

There is also the money. Running powerful AI models is expensive, particularly for agents that operate continuously and perform multiple tasks. Google's introductory Argon pricing is competitive, but the company said its published rates will eventually double to $4 per million input tokens and $20 per million output tokens. One challenge for Google is determining how to combine its models into an agent that can handle increasingly sophisticated tasks without making the service prohibitively expensive to operate.

Separately, NPR reported that Google launched a refrigerator-sized satellite into space as part of Project Suncatcher, a research project the company hopes can pave the way for orbiting AI data centers powered by the sun. The prototype satellite has four specialty chips called Tensor Processing Units, or TPUs, that Google designed for machine learning and has already deployed in data centers on Earth. The chips will run a version of the company's Gemma AI model for 15 minutes at a time due to heat management constraints, using the open weight model to answer simple queries. In a blog post, Google said the mission will gauge how the TPUs handle “the physical stress of spaceflight and the radiation and thermal extremes of space.”

Use of AI has grown sharply over the last few years, fueling demand for data centers. At Google's annual developer summit in May, CEO Sundar Pichai said demand for AI services exceeds supply and he expects capital expenditures this year to be $180 billion to $190 billion, more than six times as large as in 2022, as the company builds computing infrastructure and develops chips. At a time when opposition to the proliferation of power-hungry data centers on Earth is growing, Google is one of a handful of companies advancing plans to put them in orbit, where there is virtually unlimited free energy and no protesters.

“The sun puts out almost all of the power in our solar system. All of the other power sources that humanity has tapped into are just a tiny fraction of a percent,” said Travis Beals, senior director and lead of Project Suncatcher. “So in some sense, this project is about tapping into the best way to use solar power to run AI compute.” The satellite will be in a sun-synchronous orbit where its solar panels will almost never be in the shade, eliminating the need for carrying heavy batteries or backup power on the spacecraft.

Last November, a venture-backed startup called Starcloud launched a spacecraft that carried an Nvidia H100 chip on board and demonstrated a version of Google's Gemini AI from space, NPR reported. Elon Musk's SpaceX is also working on space data centers and has said it expects to start deploying “orbital AI compute satellites” as early as 2028. Google, in partnership with the aerospace and satellite imagery company Planet, envisions clusters of satellites eventually orbiting together to form space-based data centers. Each satellite will carry dozens of TPU chips and communicate with the others and with Earth via lasers. Next year, the company plans to put two more satellites into orbit to test their connection.

The satellite that launched is a prototype that Beals called “a very minimal test” to make sure the chips can run in space. A SpaceX rocket carried it into orbit, and Planet will get it up and running. Beals said Google will then fire up and test the TPUs. The goal is for the mission to be operational for a year.

Brandon Lucia, a professor of electrical and computer engineering at Carnegie Mellon University, said the idea sounds “very sci-fi,” but much of the technology has been proven at a research level. Still, there are huge challenges. Heat is a big one. Computer chips get hot, and on Earth the heat can be dissipated with wind or water. That is not possible in the vacuum of space. “In a satellite in particular, using more power to do the computations means dumping more heat into the confined environment inside of the satellite,” Lucia said.

In its blog post, Google called cooling the chips “a crucial research challenge” and outlined approaches including using a combination of pipes and radiators to wick heat away from the chips. Beals said the radiators are one of the heaviest components on the current mission, a drawback because heavier objects are more expensive to launch, and his team is working on ways to cool chips more efficiently. Lucia pointed out that there are also major questions about upkeep. When a chip malfunctions in a terrestrial data center, a technician can walk in and fix it. That is harder when the data center is circling hundreds of miles above Earth. “Those all have their cost and complexity amplified by a factor of 10, maybe a factor of 100. And so there has to be a big payoff,” Lucia said.

There are environmental questions, too. While space data centers would run on solar power, the rockets that launch them run off carbon-based fuels, and eventually they will re-enter Earth's atmosphere, potentially adding pollution as they burn up. Musk has said that space is the “only way to scale” data centers to meet future demand for AI. But the economics are still a question mark. Launches are costly, repairs are difficult, and communication by laser across hundreds of miles and through Earth's atmosphere is slower. It will take constellations of many satellites to equal the computing capacity of a terrestrial data center.