Google Gemini 4 Argon is the company’s new flagship model for demanding Artificial Intelligence (AI) tasks. It is aimed at paying Gemini app users, businesses, and people accessing the model through an application programming interface. Its global launch also covers Switzerland in principle, although access and cost vary considerably between Google’s products.
What Argon changes for you
Google positions Argon as a model for complex, multi-step work in fields such as finance, law, science, cybersecurity, and coding. According to Google’s official model announcement, it should handle these jobs with greater depth and precision than earlier Gemini models. Calling it a flagship does not mean that it will answer every request better; it means that Google considers Argon the most capable model in its current lineup.
For general queries, the model is supposed to provide more direct answers while reducing clichés, repetition, and excessive hedging. That could help when you want to summarize a lengthy business report, plan a multi-stage project, or structure a difficult financial question. Whether the result is genuinely shorter and more useful will still depend on the task and the quality of your instructions.
Argon is rolling out worldwide in the Gemini app and AI Mode in Google Search, but only for Google AI Pro and Ultra subscribers. Ultra subscribers also receive access through Antigravity, an environment where the model can work on larger tasks with some autonomy. Free accounts and Plus subscriptions remain on less capable Flash models; t3n specifically identifies Flash-Lite for free accounts.
For users in Switzerland, the announced global availability means Argon is not restricted to a small group of launch markets. The supplied sources do not mention separate Swiss pricing or provide reliable assessments of its performance in Swiss Standard German or regional dialects. They also give no specific details about data storage or processing in Switzerland.
What access and usage cost
Argon is not a free model in the Gemini app. Reports identify Pro and Ultra as eligible subscriptions, but the supplied material does not contain a complete price list for those plans. Simon Willison also reports access for customers on a $100-per-month subscription, without providing a full mapping of every app plan in the available text. The clear point is that app access costs money; the sources do not establish which plan offers the best value in every country.
Billing works differently through the application programming interface (API), a technical connection that lets other software use the model. Google’s introductory rate is $2 per million input tokens and $10 per million output tokens. Tokens are small units of text into which prompts and generated responses are divided for billing.
Cached input receives a 95 percent discount during the launch period, bringing the rate to $0.10 per million tokens. After the introductory offer, the standard prices double to $4 for input and $20 for output, while cached reads rise to $0.20. The Verge says the offer ends after November, while Simon Willison gives December 1 as the date of the increase.
A low per-token price does not by itself reveal the cost of completing a job. A model that produces many intermediate steps and lengthy answers may still become expensive. Willison therefore points to a total cost of $2,407 for running one benchmark with Argon and reports that a complex SVG image he generated cost 17.166 cents; SVG is a format for scalable vector graphics. These examples are useful reference points, not universal estimates for your work.
What the benchmarks actually show
A benchmark is a standardized test used to compare models on predefined tasks. Google says Argon ranks first on the Vals Index, which measures performance in economically relevant areas including finance, law, tax work, and coding. The company also says Argon outperforms competitors such as GPT-6 Astra, Claude Fable 5.1, and Claude Opus 5.5 across several tests.
The independent picture is less decisive. According to The Decoder’s report, Argon leads the Vals Index but remains behind Claude Opus 5.5 and Meta’s Muse Wonder Full in independent overall rankings. Google’s claim that Argon beats other frontier models across multiple benchmarks is not necessarily incompatible with every individual ranking, but it presents a more favorable overall picture than those independent evaluations do.
Google also provides a specific result for spatial understanding. Argon scores 87 percent on the Embodied Reasoning Question Answering benchmark (ERQA), which tests physical and spatial problem-solving in three-dimensional environments, according to Google DeepMind’s robotics report. This result is not independently verified in the supplied sources and says relatively little about ordinary chat tasks.
Practical cases better illustrate the type of work Google has in mind. At YouTube, Argon reportedly identified redundant emoji rendering, and removing it reduced server load by 45 percent. In the physics simulator used by Alphabet’s Wing drone project, the model found critical bugs that had allegedly remained in the code since 2018. These are concrete results from Google’s internal use, but they do not guarantee similar performance with another organization’s data and processes.
Pros and Cons of Gemini 4 Argon
Pros:
- Broad availability – Google says Argon is rolling out worldwide through the app, Search, and the API.
- Low introductory rate – At launch, the API costs considerably less than several flagship models compared in the supplied reports.
- Complex workflows – The model is designed for multi-step work in finance, law, science, security, and coding.
- Specific real-world results – Google’s examples range from reducing YouTube server load to finding bugs in a drone simulator.
Cons:
- No free app access – Free and Plus accounts do not currently receive Argon.
- Temporary API discount – Input and output rates are scheduled to double after the introductory period.
- Mixed rankings – Google’s strong benchmark claims are moderated by independent overall evaluations.
- Unresolved safety concerns – A capable model for cybersecurity and autonomous workflows can also be misused.
Google initially took a cautious approach to broad availability because Argon’s capabilities could be used for harmful purposes. The company says the model is trained to reject malicious instructions and that an automated monitor can stop processing when it detects an unintended workflow. These safeguards are described by Google and have not been independently verified in the supplied material.
The defensive value is nevertheless illustrated by a concrete example. Selected security specialists received early access through the Fairwind Program, and the company Wiz used Argon to find a critical flaw in hospital software that earlier models had missed, according to Google DeepMind’s safety report. Google says the discovery helped protect hundreds of thousands of patient records. The same example also shows why sensitive business or health data should not be placed in an AI system without proper controls.
What you can use right now
If you are a beginner, start with one clearly defined, moderately complex task in the Gemini app. For example, you could ask Argon to break a long report into decision options, risks, and unresolved points instead of requesting only a brief summary. Compare the output with a Flash model or another tool you already use, and verify figures, names, and conclusions against the original material.
If you are an advanced user, you can run recurring prompts through the API and track the cost of completing an entire assignment. Reusing cached, unchanged context may reduce the bill considerably if your tool supports that feature. The key measure is not only the price per million tokens, but also how much text and how many processing steps Argon needs to produce a dependable result.
Antigravity extends this approach with agentic workflows, meaning the model can carry out parts of a multi-step task with some autonomy. According to t3n, its Task mode can run in the background and connect to more than 40 third-party services, including Shopify, Adobe, and OpenTable. In professional use, that could allow a process to continue while you are away from your computer, but it also increases the need for careful permissions, supervision, and account protection.
Ars Technica reports that an automated monitoring system can halt unintended workflows and that thousands of Google employees have already used Argon. Even so, autonomous actions should not be confused with error-free automation. The more services a model is permitted to control, the greater the potential impact of a mistaken interpretation or inappropriate action.
For robotics, Google demonstrates applications ranging from controlling a four-legged Spot robot to building simulations from camera footage or hand-drawn sketches. These are not ordinary features in the Gemini app; they illustrate Argon’s spatial reasoning and its possible role in specialized workflows. For most users, the immediate value remains in text work, research support, analysis, and structured tasks.
Gemini 4 Argon combines a relatively inexpensive introductory API rate with broad availability and credible examples of demanding work. At the same time, some of its strongest performance results come from the provider, app access requires payment, and API rates will double after the launch period. The unresolved risk is less about missing features than about how reliably such a capable model can be controlled when handling sensitive data and taking autonomous actions.
Sources
- Google veröffentlicht Gemini 4 Argon: Neues Spitzenmodell ab sofort weltweit verfügbar – The Decoder, 2026-10-13
- Überraschung: Google veröffentlicht Gemini 4 Argon – t3n, 2026-10-13
- Gemini 4 Argon begins rolling out in Gemini app and Antigravity, but only if you pay – Ars Technica, 2026-10-13
- Google finally releases Gemini 4 Argon – The Verge, 2026-10-13
- Gemini 4 Argon: Built to solve your hardest problems – Google Blog, 2026-10-13
- Developing Gemini 4 Argon responsibly and safely – Google DeepMind, 2026-10-13
- Gemini 4 Argon for Robotics – Google DeepMind, 2026-10-13
- llm-gemini 0.35 – Simon Willison’s Weblog, 2026-10-13
Image: Matheus Bertelli via Pexels