Vals AI Benchmarks: Gemini 4 Argon Takes Lead

Overview

Google’s latest frontier model, Gemini 4 Argon, has claimed the top spot on the Vals Index, an independent evaluation suite that tests AI systems on economically valuable real-world tasks. Released on September 30, 2026, the model achieved a performance score of 68.90%, narrowly edging out Anthropic’s recently launched Claude Sonnet 5.5, which secured second place with 67.04%.

The results highlight a increasingly competitive landscape in large language model development. While Gemini 4 Argon leads the aggregate index, the gap between the top contenders remains slim, suggesting that different models retain distinct advantages depending on the specific domain being tested. The evaluation covered a broad spectrum of challenges, including finance, software engineering, cybersecurity, and legal research.

Gemini 4 Argon places #1 of 41 on the Vals Index at $15.68 per test, ahead of Claude Sonnet 5.5 and Claude Opus 5.5.

Benchmark Leadership and Domain Performance

The Vals Index aggregates results from dozens of specialized benchmarks to provide a holistic view of model capability. Gemini 4 Argon’s overall lead is built on exceptional performance in specific high-stakes categories. It took the #1 position on Finance Agent v2, achieving a score of 65.40%, outperforming even the faster Gemini 3.8 Flash model. Additionally, it tied with GPT-6 Astra for a perfect 100.00% score on the IOI benchmark, demonstrating elite problem-solving abilities in computational tasks.

However, the model’s dominance was not universal across all domains. In software engineering, Sonnet 5.5 maintained a slight edge, scoring 92.39% on the Vibe Code Bench compared to Gemini 4 Argon’s 91.91%. Similarly, Sonnet 5.5 led in Code Migration (69.83% vs. 68.17%) and Tax Agent Bench (77.64% vs. 76.23%). This suggests that while Google’s new model excels at financial reasoning and hard logic puzzles, Anthropic’s offering remains highly competitive for code generation and migration tasks.

Other notable placements for Gemini 4 Argon include a top-five finish on several critical infrastructure benchmarks. It ranked #4 on CyberBench, falling just 0.12 points behind GPT-6 Sol, and placed #5 on Terminal-Bench 4.0 and the Vals RSI Index. The model also showed strong generalist capabilities, landing in the top ten for Legal Research Bench and EMB.

In contrast, the model showed relative weakness in medical transcription and complex UI automation. It dropped to #15 on MedScribe and struggled significantly on CUA-bench, ranking #7 out of 8 participants with a score of only 4.83%.

Cost Efficiency and Evaluation Methodology

One of the defining features of the Gemini 4 Argon release is its pricing structure, which positions it as a cost-effective option for high-throughput applications. At $4.00 per million input tokens and $20.00 per million output tokens, it is notably cheaper than the Claude models that trail it on the index.

On the main Vals Index, the cost per test for Gemini 4 Argon is $15.68, compared to $21.34 for Claude Sonnet 5.5 and $32.14 for Claude Opus 5.5. This price advantage is significant for enterprises running large-scale evaluations or deploying agents in production environments where token usage scales rapidly.

Despite the lower base rates, the total cost of evaluation can vary dramatically depending on the complexity of the task. For long-running agentic workflows, such as those measured by CUA-bench, the cost per test rose to $193.78. Other expensive evaluations included Code Migration at $57.82 per test, Terminal-Bench Science at $44.90, and SRE Bench at $30.45. These figures underscore the importance of looking beyond headline prices when estimating operational costs for complex AI workflows.

The evaluations were conducted under consistent conditions to ensure fairness. All models were tested with a 1M-token context window and a 262k max output limit. Gemini 4 Argon’s reasoning effort was set to “high,” and temperature was fixed at 1.0. Vals emphasizes that these are independent, unbiased benchmarks designed to measure performance on real work rather than leaked or memorized knowledge.

Recent Developments in AI Research

Beyond the Gemini 4 Argon launch, the Vals platform has continued to release updates that reflect the rapid pace of AI advancement. On September 28, 2026, the team published a comprehensive evaluation of Anthropic’s Claude Sonnet 5.5. That model ranked #2 on the Vals Index with a score of 69.22%, sitting just behind the newly released Opus 5.5 variant. Sonnet 5.5 demonstrated strong value, delivering top-tier performance at approximately two-thirds of the cost of Opus 5.5.

The platform also recently partnered with Exa, Keenable, Parallel, and Tavily to launch the Vals Web Search Index. This new benchmark addresses a critical gap in AI evaluation: the lack of independent standards for web search capabilities. By providing a unified test set for legal and finance tasks, Vals aims to help customers distinguish between search tools based on actual utility rather than marketing claims.

Additionally, recent insights have shed light on the evolving role of AI detectors. A September 30 analysis explored whether the Bitter Lesson—the principle that scalable compute eventually dominates hand-crafted solutions—might also apply to AI content detection. The findings suggest that as models improve, specialized detectors face increasing difficulty in maintaining accuracy against sophisticated rewrites.

In a separate scientific breakthrough highlighted in early October, Claude Opus 5.5 agents were tasked with discovering new materials. The agents successfully identified KV[Cr(CN)₆] as a room-temperature magnetic semiconductor candidate, showcasing the growing ability of LLMs to assist in genuine scientific discovery and material science research.