Google DeepMind's Gemini 4 Argon Now Generally Available
Google DeepMind has announced the general availability of Gemini 4 Argon, its latest frontier model, to a broad audience including developers, enterprise clients, and subscribers of paid Gemini services. This AI model, distinguished by its sparse mixture-of-experts architecture, is specifically engineered to handle extensive, multi-stage tasks across diverse sectors.
Initially introduced to a select group of cybersecurity professionals on September 30, Argon is now part of the paid Gemini API offerings. Its pricing is set at $2 per million tokens for input and $10 per million tokens for output. A pivotal enhancement for developers is Argon's ability to produce continuous text outputs reaching an unprecedented one million tokens within a single response, a substantial increase from the 64,000 token limit of previous Gemini iterations.
Understanding Gemini 4 Argon's Core Capabilities
Argon is presented as an entirely new model, not a refined version of an existing Gemini AI. It operates as a large-scale sparse mixture-of-experts transformer, providing native support for multimodal inputs. This means the model can process and understand information presented as text, images, audio, and video, subsequently generating text-based responses. Google utilized its advanced TPUs for training, leveraging JAX and ML Pathways. The model's knowledge base extends up to March 2026, though specific parameter counts have not been disclosed by the company.
The Significance of a 1 Million Token Output Limit
The elevated output capacity to one million tokens is a game-changer for applications requiring extensive context or detailed generation. This capacity is part of a combined input-output window, meaning that longer prompts will naturally reduce the available space for the model's output. Google has also introduced a new continuation protocol to facilitate these lengthy decodes, as detailed in its developer and enterprise documentation. At its introductory rate, generating a full one million token response would incur an output cost of $10.
Performance Benchmarks: Strengths and Areas for Improvement
Google has positioned Gemini 4 Argon against prominent competitors like GPT-6 Astra, Claude Fable 5.1, and Claude Opus 5.5, showcasing its performance across various benchmarks:
- Leading Scores: Argon achieved top results in several critical areas, including 91.7% on LVBench for long video understanding. It also outperformed rivals on DeepSWE v1.1 (77.9%), Vals Index (68.90%), AutomationBench (51.3%), and demonstrated a nearly threefold lead on Harvey's Legal Agent Benchmark (19.6%). Impressive results were also seen in GraphWalks (84.2% for 256K to 1M tokens) and CWE-bench v1 (68.0%, matching GPT-6 Astra). An 87% score on the ERQA embodied reasoning benchmark was also reported.
- Areas for Development: While strong in many areas, Argon displayed lower performance on specific metrics. It trailed behind competitors on FrontierSWE v2 (55.0%), Terminal-Bench 4.0 (57.4%), Terminal-Bench Science 0.1 (57.6%), PostTrainBench (45.3%), and OSWorld-2.0 (offline subset) (69.2%).
It is important to note that evaluation methodologies can vary, with some scores being self-computed or sourced from public leaderboards. For instance, Argon's DeepSWE score was derived using a mini-swe-agent harness.
Acknowledged Limitations and Safety Protocols
Google's model card for Gemini 4 Argon candidly lists certain limitations, including the potential for hallucinations and occasional instances of slowness or timeouts. Adhering to its Frontier Safety Framework, Google acknowledges that it cannot entirely eliminate risks associated with biological and cybersecurity critical capability levels. However, the company states it has implemented various safeguards designed to mitigate these identified risks effectively.
This article is a rewritten summary based on publicly available reporting. For the original story, visit the source.
Source: MarkTechPost
