On October 13 (local time), Google began generally offering its latest model, “Gemini 4 Argon.” Google AI Pro and Ultra subscribers can use it in the Gemini app and AI Mode in Search. It is also available through the API and Google Cloud. During the introductory period, API pricing is $2 per 1 million input tokens and $10 per 1 million output tokens.
A door that had been closed opened in less than two weeks
Argon first appeared on September 30. At the time, it was made available only to a limited group, including organizations responsible for cyber defense. Now, 13 days later, access has expanded at once to developers, businesses and everyday users.
There are four main ways to use it: the Gemini app, AI Mode in Search, the API and Google Cloud. Ultra subscribers can also use it in Google Antigravity, a coding tool. With so many entry points to the same model, choosing where to start becomes a decision in itself.
Select it from the “+” icon in Search
In AI Mode, switch to Argon from the menu under the “+” icon in the input field. It’s in the same place where 3.7 Flash was added in August and 3.8 Flash in September. Robby Stein, who leads Search, cited improvements in complex tasks involving multiple steps and in conversations that require back-and-forth.
The rollout is gradual, so Argon may not appear in everyone’s menu right away. The announcement also does not change the default model for AI Mode users on the free tier. Argon is an option that subscribers must select themselves.
Even for the same top-tier model, access depends on your budget
There are three paid plans in Japan: Pro at ¥2,900 per month, Ultra 5x at ¥14,500 per month, and Ultra 10x at ¥32,000 per month. All let subscribers select Argon in the Gemini app and AI Mode. Only Ultra subscribers can use it in Antigravity.
By contrast, free users and paid Plus subscribers get only Flash models under the new tier structure. Argon is positioned as a frontier-series model replacing the previous “Pro” class, and Google says it plans to release other Gemini 4 models over time. If an updated Flash model follows, lower-tier subscribers could benefit as well.
The ¥2,900 monthly difference is significant. Flash should be enough for many everyday tasks, such as drafting text or doing research. Whether you need it for demanding work, such as making code changes across multiple files, may be a useful way to decide if a subscription is worth it.
Six months of free access for researchers
Google has also opened up an option beyond personal subscriptions. It says it will provide up to 10,000 selected researchers with six months of free access to Antigravity’s highest usage tier and Argon. Applicants must be researchers affiliated with a university or research institution and apply through a designated form. Places will be allocated in application order, subject to eligibility, availability and selection.
In cybersecurity, organizations participating in DeepMind’s “Fairwind Program” will receive a version with relevant guardrails removed. The program plans to expand to hundreds of participating companies. The idea is to put powerful tools in the hands of defenders.
API prices will double after the introductory period
For developers, the key question is the price. During the introductory period, the rate is $2 per 1 million input tokens and $10 per 1 million output tokens. Cached inputs are 95% cheaper than regular inputs. When the introductory period ends, the prices will rise to $4 for inputs and $20 for outputs—exactly double.
The regular rate is on par with Anthropic’s “Claude Opus 5.5.” “Claude Fable 5.1” and “GPT-6 Astra” cost $10 for input and $50 for output, making Argon’s regular prices less than half as much. Based on the cached-input discount, the rate works out to about $0.10 during the introductory period and about $0.20 afterward. Google has not stated these figures explicitly, however, and instead lists the discount as “95% off.”
The introductory price looks attractive when prototyping, but it’s prudent to budget for the regular rate in production. The impact of doubling will be greatest for apps that repeatedly run the same processes.
How to read the numbers on performance and safety
Google has also shared examples of internal use. It says software engineers using Argon increased code submissions through agents by more than 35%, while the number of submissions sent back for changes fell by nearly a third. At Chromium, Google says it carried out a design change across hundreds of files in a codebase of around 40 million lines, completing work that would have taken months in just a few weeks.
The inference speed of the open model “Gemma” rose from about 37,900 tokens per second to more than 108,000—2.85 times faster—in a vLLM environment on an A100. All of these figures come from Google itself. Among external benchmarks, Argon tops the Vals Index, while Claude Opus 5.5 ranks higher in independent overall rankings. Because its announcement came about two weeks later, it also trails Meta’s “Muse Wonder Full,” which was released the previous day.
In safety evaluations by external red teams in the biological domain, fewer than 1% of attempts bypassed the safeguards. Agents tried to take unauthorized shortcuts in 2% of cases. On Gray Swan’s benchmark for attacks using embedded external instructions, the attack success rate was 0.7%. In the previous round of the same test, GPT-6 Astra scored 8.5% and Claude Opus 5 scored 4.8%, but the conditions may not have been exactly the same, so direct comparisons should be treated with caution.
The shadow of a potential successor, “Carbon”
Google is also reportedly testing a successor to Argon internally, codenamed “Carbon.” The model is said to be designed to outperform Argon, particularly in coding.
By the time Argon became generally available, it was already being discussed alongside what might come next. With model generations turning over every few weeks, users may need to focus less on chasing the latest release and more on choosing a plan and pricing that suit their needs.
Comments