Google Just Opened Gemini 4 Argon to Everyone. Here's What It Actually Does
Google opened Gemini 4 Argon to developers and subscribers on October 13. Here's what the new frontier model does, what it costs, and who should actually use it.
What the Gemini 4 Argon release changes
Google opened Gemini 4 Argon to the wider world on October 13, two weeks after a debut that only a small group of cybersecurity defenders got to see. The model is now available in the paid tier of the Gemini API, in the Gemini app for Pro and Ultra subscribers, in AI Mode in Search, and in Antigravity, Google's coding tool, for Ultra subscribers. API pricing sits at $2 per million input tokens and $10 per million output tokens.
That price matters because Argon is a frontier model, not a budget tier. Anthropic's Claude Opus 5.5 costs $5 per million input tokens and $25 per million output. Google is undercutting the flagship class while claiming frontier-level ability. Whether that claim holds up depends on which benchmark you look at, but the pricing alone puts pressure on everyone else.
The headline technical detail is a 1 million-token output limit. Most models cap output far below input, which means long jobs get chopped into pieces. A model that can produce a million tokens in one run can, in theory, refactor a large codebase or draft a long report without losing the thread. Google says its own engineers used Argon on a Chromium refactoring project spanning hundreds of files, work expected to take months, finished in weeks. Company anecdotes deserve skepticism, but the output limit itself is real and unusual.
The benchmark picture is mixed, as usual
Argon topped the Vals Index on October 12 with a score of 68.90%, ahead of Claude Sonnet 5.5 at 67.04% and Claude Opus 5.5 at 66.97%. The Vals Index tests finance, coding, legal, and tax tasks, so it is closer to real work than a trivia quiz. On the Artificial Analysis Intelligence Index, Argon tied OpenAI's GPT-6 Astra and trailed two Anthropic models. No model led every category.
Google also reported gains in science and robotics. The model made the open-source xTB quantum chemistry package up to 2.3 times faster in Google's testing, and scored 87% on the Embodied Reasoning benchmark, which tests physical and spatial reasoning. In one internal experiment, Argon watched a video of a person playing cornhole, built a simulation, and helped a humanoid robot reproduce the activity. Treat these as directional, not decisive: vendors pick the tests that flatter them.
Google says Argon ships with its most comprehensive safety and security stack yet, tested with external partners. That language showed up because the model's first audience was cyber defenders in Google's Fairwind program, and because the AI agent wars have made safety claims part of the marketing. Independent verification will take time.
Who should actually care
If you build software, Argon is worth a test drive for one reason: agentic coding. Google says internal use in Antigravity increased agentic code submissions by more than 35% while cutting rollbacks by nearly a third. The company did not disclose the measurement period or baseline, so take the numbers with a grain of salt. But the direction is consistent with what every lab is chasing: models that can plan, execute, and check multi-step work without constant supervision.
If you run a business, the more useful comparison is cost per task. Argon at $2/$10 sits between cheap small models like Claude Haiku 5.5 and expensive flagships. For drafting, research synthesis, and code review, that middle tier is where most real workloads live. The cheapest model that does the job reliably wins, and Argon is priced to be that model for long, complex jobs.
What to try first
Start with the jobs where context loss hurts most: refactoring a messy module, reviewing a long pull request, or synthesizing a stack of research papers into a briefing. Retrieval-augmented setups still help, but a model with a huge output window reduces the need to stitch answers together by hand.
Don't switch your whole stack on day one. Run Argon next to whatever you use now on the same tasks for a week, and compare the outputs on accuracy, not vibes. Frontier models converge fast, and today's leader is next month's second place. The Gemini 4 Argon release is a strong entry, but the race it joined is far from settled.
