Google Unveils Gemini 4 Argon with 1M Output Tokens and Restricted Early Access
Google's new Gemini 4 Argon model targets complex, long-horizon tasks with a 1-million-token output limit, but initial access is limited to select cybersecurity partners.
Google has introduced its latest frontier AI model, Gemini 4 Argon, designed to manage complex, long-horizon tasks across software engineering, cybersecurity, and enterprise analysis. However, the model is not yet widely available. To comply with the US government's voluntary safety review process, Google is initially limiting access to a select group of security organizations through its Fairwind Program. This release follows a seven-month gap in major model launches for Google, during which the company delayed and eventually bypassed the promised June release of Gemini 3.5 Pro due to development hurdles in coding and reasoning.
Technical Capabilities and Benchmark Performance
Argon's most notable upgrade is its output capacity, which has expanded to 1 million tokens from the 64,000-token limit of previous Gemini models. This expansion is designed to support extended reasoning and multi-step tasks within a single execution trajectory. Google has already applied Argon internally to optimize memory usage in its data centers, identifying savings that could free up over 300 TB of memory, with ultimate goals of 500 TB to 1 PB.
On industry benchmarks, Argon shows competitive but mixed results. It scored 68.9% on the Vals Index, ahead of Anthropic's Claude Opus 5.5 at 67.0%, and achieved 77.9% on DeepSWE v1.1 compared to 74.2% for Opus 5.5. In cybersecurity, Argon tied for the top spot on CWE-bench v1 with a 68% score, matching Grok 4.7, GPT-6 Astra, and Claude Opus 5.5. However, it lags behind Claude Opus 5.5 on PostTrainBench for machine learning engineering, scoring 45.3% against the competitor's 49.3%.
Pareekh Jain, CEO of EIIRTrend and Pareekh Consulting, noted that while Argon successfully completes multi-step tasks using automated tools and adheres closely to facts, its standard coding capabilities remain average. Jain pointed out that the model still trails competitors in creative writing, nuanced explanations, and operating command-line terminals.
Pricing and the Economics of Switching
Google is offering Argon with an introductory price of $2 per million input tokens and $10 per million output tokens, with cached inputs discounted by 95%. This introductory rate will eventually increase to a standard price of $4 per million input tokens and $20 per million output tokens, though Google has not announced when this change will occur.
This future pricing aligns closely with Claude Opus 5.5 ($4 per million input, $20 per million output) and remains significantly lower than OpenAI's GPT-6 Astra ($10 per million input, $50 per million output). According to Jain, enterprises should calculate their business cases using the standard $4/$20 pricing rather than relying on the introductory rates. He advises CIOs to evaluate models based on successful business outcomes and task accuracy rather than simple token costs, suggesting that organizations add Argon for specific workloads—like legal or financial document analysis—rather than migrating entire workflows.
What it means for developers
For developers, the launch of Gemini 4 Argon represents a significant step forward in handling long-horizon reasoning, but immediate access remains restricted to select security partners. This means most developers cannot integrate Argon into their production pipelines just yet. Once it becomes widely available, the 1-million-token output limit will allow for more complex agentic workflows and multi-step automation without the model losing context.
However, because Argon does not clearly lead across every category—particularly in everyday coding and command-line tasks—developers will likely need to adopt a multi-model strategy. Rather than migrating existing applications entirely to Google's ecosystem, developers should benchmark Argon against their specific workloads. While waiting for broader access to frontier systems, developers can try top AI models cheaply through one API at https://apixoai.online. This approach allows teams to compare the strengths of different models, utilizing Argon for data-heavy, long-document tasks while retaining other models for creative writing or standard coding assignments.
Source: Google makes Gemini 4 AI model available to a trusted few — InfoWorld AI. Written by the Apixo team from that report.
One key for Claude, GPT, GLM, DeepSeek and more. Pay per token with crypto.
Get your API key
