When major technology companies introduce a new generation of AI systems, the gap between what appears on paper and what ordinary people can actually try often stretches longer than the announcements suggest.
Benchmark tables arrive first, accompanied by careful language about safety reviews and phased access.
Ordinary accounts remain on older systems while the newest model stays with a narrow set of partners.
That pattern held for Gemini 4 Argon when Google first described it.
Coverage at the time noted the model’s strong showing across a range of published evaluations. Rarly reports described Argon as a frontier system aimed at complex, multi-step work in software engineering, enterprise knowledge tasks such as legal and financial analysis, and cybersecurity defense.
Google's own comparison table placed it ahead or tied for first on the majority of the rows it chose to publish, with particularly wide margins on certain knowledge-work measures.
It was a disappointment because almost no one outside a limited group could run it.
Now, Google finally announced that Gemini 4 Argon is finally out.
The model supported a one-million-token output limit, far higher than previous Gemini releases, which in principle allowed longer continuous reasoning on a single problem.
Pricing was set at an introductory rate of $2 per million input tokens and $10 per million output tokens, with a steep discount on cached inputs.
Google DeepMind posted that Gemini 4 Argon was now globally available and built to handle hard problems from start to finish across coding, enterprise workflows, and cyber defense.
In practive, Argon can break tasks into steps and then continues past an outline to produce usable results, such as a shareable web application, ready-to-print three-dimensional files, or finished slides and reports.
Another noted major improvements in coding and the addition of Google Antigravity's agentic coding interface inside the Gemini app on macOS, so users can move between conversation and code work without leaving the same environment.
A further update highlighted connections to more than forty external services, including Instacart, OpenTable, Adobe, and Shopify, allowing the system to handle actions ranging from grocery orders and restaurant reservations to design and business-building tasks.
A later post showed the model turning a video or webcam feed into a playable synthesizer by generating an interactive audio application in Canvas. The account stated that these capabilities were reaching Google AI Pro and Ultra subscribers globally that day.
The accompanying blog post on the Gemini app confirmed the same access path.
Pro and Ultra subscribers can now select Argon in the app and in AI Mode in Search.
The system is described as answering more directly, with less hedging and fewer repeated phrases, while providing clearer progress updates during longer jobs. A toggle lets users shift from ordinary chat into task mode, in which the model can continue work across Google services even when the person is away from the computer.
Developers and enterprises gain access through the paid Gemini API in AI Studio and the broader Gemini platform.
Ultra subscribers can also use the model inside Google Antigravity. Scientists and researchers may apply for a free six-month period of higher usage on Antigravity, subject to selection.
Cybersecurity teams remain able to request Fairwind access if they want the version without certain defensive guardrails. Free accounts and the lower Plus tier stay on lighter models for the time being.
The introductory pricing remains in place through the end of November 2026 before rising to $4 per million input tokens and $20 per million output tokens.
The one-million-token output capacity is available via the API.
Early external measurements, such as those from Artificial Analysis, had placed the model in the same broad range as several other leading systems on aggregate intelligence scores while showing advantages on specific long-context and knowledge-work tasks.
The release moves Argon out of the narrow initial cohort that drew the earlier commentary and into the hands of paying users who already hold Pro or Ultra plans, while the broader free tier continues to wait.
In the comparisons Google published at launch, Argon led or tied for first on most of the rows against GPT-6 Astra, Claude Fable 5.1, and Claude Opus 5.5, with the clearest margins appearing in knowledge-work benchmarks such as the Vals Index, AutomationBench, finance-agent tasks, and a legal-agent evaluation.
It also posted the highest score among those models on DeepSWE for long-horizon software engineering. The same table showed it trailing on a handful of coding and terminal-oriented tests, including FrontierSWE and Terminal-Bench.
Independent aggregate scoring from Artificial Analysis placed it level with Astra and Fable and a few points behind Opus 5.5.
It's worth noting that Google has indicated that Argon is the first in the Gemini 4 series, so further variants are expected to follow.
