Google Releases Gemini 4 Argon, Tops Vals Index and Boosts Agentic Coding by 35%

Google has launched Gemini 4 Argon, a new frontier model designed for complex engineering, scientific, and business tasks. The model secured the top spot on the Vals Index, which measures performance across finance, coding, law, and tax sectors, outperforming other frontier models. It is available globally via AI Studio, the Gemini Platform API, Google Antigravity, and the Gemini app for subscribers.

Engineering and Coding Performance

Internal tests at Google demonstrated significant efficiency gains when using Argon in Antigravity. Agentic code submissions increased by over 35%, while code rollbacks decreased by approximately one-third. In a major refactoring effort for the Chromium codebase, Argon separated legacy browser functions from core web page processing across 6 million lines of code in weeks rather than months. The model also accelerated the Google Agent Development Kit (ADK) by reducing in-memory session prep time by 40% and speeding up cold imports by 21%.

Scientific and Domain-Specific Breakthroughs

Argon shows strong capabilities in specialized fields. In quantum computing, it helped translate the AlphaQubit error correction decoder into a single hardware kernel, nearly doubling inference throughput. For nuclear fusion research, the model assisted in discovering counterexamples to "Grad's conjecture," a mathematical problem considered unsolvable since 1967. In rare disease research, Argon identified chromosomal rearrangements in newborn genomes that routine software had missed for years, scoring 57% on the RareBench diagnostic benchmark.

Business Workflow and Safety

Partners reported high success rates in long-running tasks. Deloitte achieved a 100% success rate in long-running execution evaluations, while Kepler matched analyst-built reference models with over 99% numerical accuracy in complex financial workflows. Box noted a 14% accuracy improvement over Gemini 3.1 Pro in document analysis. On the safety front, Argon reduced shortcut attempts in coding benchmarks to 2% and achieved a 0.7% attack success rate against indirect prompt injection in the Gray Swan benchmark.

Availability and Pricing

The model supports a 1 million unified input-output token limit. Introductory pricing is set at $2 per million input tokens and $10 per million output tokens, with a 95% discount for cached inputs. After November 30, 2026, prices will rise to $4 for input and $20 for output per million tokens. Scientists can apply for six months of free access to the highest usage tier of Google Antigravity.