Gemini 4 Argon Benchmark Analysis: How It Stacks Up Against GPT-6 and Claude

Google's latest flagship model, Gemini 4 Argon, made its official debut this week. Among the 19 benchmarks published by Google, it took first or tied for first place in 14 of them, going head-to-head with OpenAI's GPT-6 Astra and Anthropic's Claude Fable 5.1 and Opus 5.5. This isn't a clean sweep—independent testers' data confirms some conclusions while painting a completely different picture in other dimensions.

Knowledge Work and Coding: Argon's Strong Suit

The widest gaps in Google's published results are concentrated in professional services scenarios. On the Harvey legal agent benchmark, Argon scored 19.6%, nearly four times GPT-6 Astra's 5.4%. On the Vals Index, which aggregates finance, coding, legal, and tax tasks, Argon reached 68.9%, ahead of Astra's 63.1% and Opus 5.5's 67.0%.

In coding, Argon leads on DeepSWE v1.1 (a real-world software engineering task set) with 77.9%, ahead of Opus 5.5's 74.2% and Astra's 74.1%. It also took first place on Vibe Code Bench (91.9%), the long-video understanding test LVBench (91.7%), and GraphWalks (84.2%), a long-context test supporting up to one million tokens.

Independent testing firm Vals AI ranked Argon first among 41 models on its Vals Index with a score of 68.90%. Artificial Analysis gave it an Intelligence Index score of 53, tied with GPT-6 Astra and 23 points above Google's previous-generation Gemini 3.1 Pro Preview. Cross-validation from these two independent sources essentially confirms that Argon is genuinely in the top tier for overall intelligence.

The Harvey legal agent benchmark was developed by legal AI company Harvey to specifically evaluate large models on legal practice tasks, covering contract review, due diligence, legal research, and document drafting—real lawyer workflows. Most questions are drawn from actual cases and demand deep reasoning, long-document processing, and legal domain knowledge. Because legal tasks have low error tolerance and high specialization requirements, scores tend to be low—19.6% is already quite a lead and serves as a key reference for measuring a model's "professional services readiness."

The Vals Index is a comprehensive index from independent evaluation firm Vals AI that weights and aggregates results across finance, coding, legal, tax, and other task categories into a single score, making it easy to compare different models' overall capabilities in real knowledge-work scenarios. The companion Artificial Analysis Intelligence Index is published by another independent firm and focuses on standardized evaluations of reasoning, knowledge, and coding abilities using a unified scoring scale. Neither index relies on vendor self-reported data, so the industry often uses them as cross-validation sources for judging a model's true capabilities.

Where Competitors Still Hold Their Ground

If you only look at the 19 tests Google itself listed, it's easy to miss the other side. On the harder FrontierSWE v2 software engineering evaluation, GPT-6 Astra leads Argon by a wide margin, 65.5% to 55.0%. Claude Opus 5.5 wins on Terminal-Bench 4.0 (terminal-driven agent tasks) at 66.4% to 57.4%, while Astra takes Terminal-Bench Science at 68.1% to 57.6%.

On computer operation tasks, Astra scored 72.6% on the OSWorld-2.0 offline subset, slightly above Argon's 69.2%. The machine learning engineering test PostTrainBench was led by Opus 5.5 at 49.3% to 45.3%. For teams building shell automation pipelines, these are precisely the most critical workloads—terminal operations, ML engineering, and hard SWE tasks—and Argon holds no advantage here.

Terminal-Bench is a class of agent evaluations focused on terminal operations, requiring models to complete file operations, process management, and script debugging in real shell environments. OSWorld focuses on graphical user interface operations, such as clicking menus, filling out forms, and cross-application collaboration. These two types of tests measure a model's "hands-on ability"—unlike pure text Q&A, they require multi-step planning, observing feedback, and error correction. PostTrainBench targets machine learning engineering workflows such as data cleaning, fine-tuning configuration, and hyperparameter tuning, assessing whether a model can handle an ML engineer's daily tasks. These are exactly the dimensions that teams building automation pipelines care about most.

The Footnote Behind the Benchmark Table

One thing bears repeating: all 19 results were run and reported by Google itself, and no third party has fully reproduced the table at publication time. It's industry convention for flagship models to self-report benchmarks at launch, but when the evaluator and the evaluated are the same party, credibility takes a hit.

Argon's design goal was long-horizon "deep work" from the start, and its one-million-token output ceiling (more than a 15x increase over Gemini 3's 64,000) is itself a notable spec leap. This capability has already shown up in long-context tests like GraphWalks, but stability in real workflows still needs more independent verification.

What This Means

Roughly summarizing the current landscape: Argon won more benchmark categories, independent organizations confirm it's genuinely top-tier, but competitors still hold advantages in several key categories. This isn't the endgame of the LLM race—just a change in who's leading.

The practical advice for enterprises: don't just look at total benchmark counts. If your use case involves legal agents, long-context reasoning, or comprehensive knowledge work, Argon's lead is backed by independent data. But if your core needs are terminal automation, hard software engineering, or ML engineering tasks, Claude Opus 5.5 and GPT-6 Astra have clear advantages in those specific dimensions—and those advantages come from Google's own report, making the competitors' leads in those categories even harder to dispute.

One-line Summary

Gemini 4 Argon leads most benchmarks but still trails GPT-6 Astra and Claude Opus 5.5 in terminal, SWE, and ML tasks.

Paragraph Summary

Google released flagship model Gemini 4 Argon, which ranks first or tied for first in 14 of 19 self-reported benchmarks, with independent firms Vals AI and Artificial Analysis both confirming its top-tier status. Its strengths are concentrated in legal agents (Harvey benchmark 19.6% vs Astra's 5.4%), comprehensive knowledge work (Vals Index 68.9%), and long-context tasks (GraphWalks 84.2%), plus a one-million-token output ceiling. But independent testing also reveals weaknesses: GPT-6 Astra leads on FrontierSWE v2 and OSWorld-2.0, while Claude Opus 5.5 dominates Terminal-Bench 4.0 and PostTrainBench. Since all benchmarks are self-reported by Google with no full third-party reproduction, credibility should be discounted. Enterprise model selection should depend on specific use cases.