I Tested Google’s New Gemini 4 Argon Satellite, and Here’s How It Compares to GPT-6.
The Gemini 3.1 Pro processor was released back in February of this year—quite a while ago, in terms of AI. Now Google has moved up to the next level with the Gemini 4 Argon, which, according to the company, “excels at deep learning on complex real-world problems.”
The new AI model is available right now in the Gemini app if you have the AI Pro ($19.99/month) or AI Ultra ($99.99/month) plan. Free and AI Plus ($4.99/month) users cannot access it yet, though a scaled-down version of Gemini 4 Argon is expected to be available for these plans in the future.
What is the Gemini 4 Argon spacecraft capable of?
Google cites a long list of areas in which the new AI model demonstrates superior results. They cite complex engineering tasks (such as reviewing millions of lines of code for efficiency improvements), predicting molecular structures in chemistry, and creating 3D visualizations. It also improves knowledge work in legal, financial, scientific, and programming fields.
According to Google, Gemini 4 Argon is now better at simulating complex physical processes, analyzing genetic databases, protecting software from vulnerabilities, and solving problems in physical space (which is useful for robotics development). The new model, Google says, “answers questions more directly, with fewer evasions, doesn’t repeat itself more than necessary, and avoids clichés in its answers.”
Amid concerns from many in the industry about an AI-fueled apocalypse , Google has strengthened its protections against abuse and hacking. The company has significantly improved its malicious query blockers, limited the capabilities of its AI agents, and increased resistance to malicious suggestion injections.
All these statistics are what you’d expect to see in an announcement of a new AI model, whether from Google, Anthropic, or OpenAI: “Our new model is better and safer than ever, and especially good on these specific tests we’ve chosen.” But how much more useful is this for the average user using the Gemini app?
I don’t have any pressing scientific discoveries to explore or large-scale software optimizations to make, but I did test Gemini 4 Argon against GPT-6 Sol, the latest model available to consumers through the ChatGPT app—though you’ll need at least the Plus plan ($19.99/month) to do so.
Argon testing on board Gemini 4 compared to GPT-6 Sol
Make no mistake: both of these new AI models are quite capable. First, I gave Gemini 4 Argon and GPT-6 Sol the following command: “Using information from the internet, create a web app I can run locally that will display the world’s lighthouses on a map. Clicking on a lighthouse should display information about it, including its range and a simple 2D model of its shape and height (if known). Include a lighthouse keeper in each pop-up window of the model for scale.”
At the highest and slowest levels of ” thinking, ” these AIs took some time to respond to me: almost 10 minutes for Gemini and almost 26 minutes for ChatGPT. Both faced the same problem—limited access to online mapping data—but both were able to create their own rudimentary world map when asked a second time.
ChatGPT’s “Beacon” app was decent, except for the map’s a bit clunky (his mapping skills clearly need some work). It provided a wealth of information, organized everything well, and gave me neat, clear diagrams of each beacon. The day/night mode worked perfectly, and it was an impressive piece of programming, even though it took 26 minutes to create.
However, the Gemini project, dubbed “Pharos,” was superior in almost every way. The map was more detailed, beacons actually showed the range of their illumination (with an impressive animation), and there was more information for each location—perhaps thanks to Google’s expertise in web search. Some of the design decisions were a bit unfortunate, but Gemini 4 Argon created a clearly superior app in less than half the time.
Next, I fed each AI model PDFs of 10 recent scientific studies. The task was: “Give a simple summary of each of these studies, one sentence for each of the main findings, one strength of the study, one weakness of the study, and one main implication.” Although I didn’t read the studies in full, I did read news articles about all of them and set the “thinking” level to low.
This time, ChatGPT completed the task faster—25 seconds versus Gemini’s 55 seconds—and, in my opinion, provided better summaries, at least for my purposes. Gemini performed well, its answers were precise and to the point, but ChatGPT’s summaries were easier to understand and shorter. ChatGPT also offered some general considerations for the most promising research, something Gemini did not.
Finally, I tested the models’ imagination and creativity: “Given the success of Apple TV series ‘Severance,’ ‘Slow Horses,’ ‘Silo,’ and ‘Widow’s Bay,’ please propose five ideas for new shows on the platform that combine elements of some or all of these series. Try to be as original as possible and limit each proposal to approximately 100 words.” I again raised the “thinking” level to “high” for this assignment.
(Is this really how Apple TV executives get their ideas these days? I hope not.) As for the answers the AI gave me, Gemini was faster than ChatGPT (1 minute and 28 seconds versus 3 minutes and 48 seconds, respectively), and both produced answers that were a bit more literal in their mix than I was going for: botched Secret Service agents and supernatural horrors were all over the place, which I suppose was my intention.
Most of the ideas boiled down to ” Widows Bay, set somewhere else” or ” Slow Horses, set somewhere else,” and for the most part didn’t make much sense, but Gemini’s best work was about a Silo- style underground bunker whose operators go through a 24-hour routine similar to the one in The Breach : the outside of the bunker defends against zombies at night; inside, the bunker plans social events for the residents during the day. It’s not perfect, but I might at least watch the trailer.
Of ChatGPT’s proposals, I liked the idea of a show about a remote weather station, similar to Severance , whose staff discovers they’re not recording the weather, but creating it. Again, it’s not entirely logical, but overall, GPT-6’s proposals seemed more thoughtful, more suited to a real show, and better “thought through.” Perhaps Gemini should have given itself more time to “think things over.”
This is just a narrow set of quick tests, but in my opinion, Gemini is at least back on a level playing field with ChatGPT. It performed significantly better on the beacon task, though its summarization and TV show presentation features weren’t as impressive. Access to the best models requires a subscription to both platforms, but if you have the option, it’s worth comparing them against each other on the tasks that most often require AI.