10 hrs ago
Google’s Gemini 4 Argon targets coding, finance and legal work
Gemini 4 Argon is a new artificial-intelligence model from Google.
It is designed to help with coding, research, writing, finance and legal work.
Google says many of its employees are already using it.
In one example, it helped make a quantum-computing method use 40% fewer resources.
It also helped find memory improvements that freed more than 300 TiB in a Google data center.
The model can help move very large software projects from C and C++ to Rust.
It performed strongly on some tests, but other models scored higher on certain software-engineering benchmarks.
Google says Argon can understand charts, long videos and information spread across several documents.
The company is also adding controls that can stop it if it begins doing something outside the user’s instructions.
Google says Gemini 4 Argon is being used by thousands of employees for coding, research and writing.
In examples, Argon reduced quantum-algorithm resource needs by 40% and freed more than 300 TiB of data-center memory.
The model is helping migrate large C/C++ codebases to Rust, including projects exceeding 800,000 lines.
Argon scored 68.9% in overall economic-impact evaluations and 77.9% on the DeepSWE v1.1 software-engineering test.
Google says Argon can analyze charts, long videos and multiple documents while monitoring its actions for safety.
- Who
- Google and its Gemini 4 Argon artificial-intelligence model.
- What
- Google described Argon’s capabilities, use cases, benchmark results and safety controls.
- Where
- In Google’s projects and data-center examples, with the model used by Googlers.
- When
- Why
- To support coding, research, writing, finance, legal analysis and other complex tasks.
Google’s overall performance case
Competitors’ specialized benchmark advantage
Benchmark performance
Google’s overall performance case
Google says Gemini 4 Argon leads overall economic-impact evaluations with a 68.9% score and achieved 77.9% on DeepSWE v1.1.
Competitors’ specialized benchmark advantage
Argon trails GPT-6 Astra on FrontierSWE v2 and Claude Opus 5.5 on Terminal-Bench 4.0, according to the reported scores.
Autonomous task safety
Google’s overall performance case
Google says it monitors Argon’s reasoning and actions and can stop tasks that move beyond the user’s intended boundaries.
Competitors’ specialized benchmark advantage
The article presents this as a Google claim and does not provide independent test results showing how effective the controls are.
Key facts
- Economic-impact evaluation
- 68.9% across tax, legal, coding and finance
- DeepSWE v1.1 score
- 77.9%
- Quantum-algorithm improvement
- 40% reduction in resource requirements compared with a published baseline
- Freed memory
- More than 300 TiB through data-center telemetry improvements
- Code migration
- Includes C/C++ projects exceeding 800,000 lines of code
- Other capabilities
- Analyzing charts and graphs, understanding long videos, and using multiple documents
- Safety controls
- Google says Argon’s reasoning and actions are monitored, with tasks stoppable when boundaries are exceeded
Sources
Related news
Google Launches Gemini 4 Argon for Cybersecurity and Coding
Google Unveils Gemini 4 Argon With One-Million-Token Output
Google Unveils Gemini 4 Argon for Long-Horizon AI Workflows
Google's Gemini 4 Argon Challenges GPT-6 Astra, Claude Opus
Google Launches Gemini 3.8 Flash Models With Cybersecurity Tools
Google Announces Gemini 4 Flagship Argon After Months Of Delays
Anthropic IPO Filing Warns Advanced AI Could Threaten Humanity
White House Seeks US Review Before UK AI Model Access
OpenAI and Anthropic Release Cheaper Models Amid AI Slowdown Calls
