Gemini 4 Argon Benchmarks: Strong Agent, Not a Clear Frontier Leader

A long-awaited release with a split verdict

Google DeepMind has shipped Gemini 4 Argon, its first proprietary model outside the Flash line in more than seven months. 3 Google is presenting it as a frontier contender. Its own comparisons show Argon beating or tying GPT-6 Astra and Claude Opus 5.5 on 14 of the 19 benchmarks the company chose to publish. 5 Independent evaluators tell a narrower story. Argon looks like a genuinely strong agentic and knowledge-work model, but on broad general intelligence it sits level with the pack rather than ahead of it. 24

The gap between those two pictures is the real story of this launch.

What Google is claiming

Google's headline numbers lean toward professional workflows. Argon reports 68.9% on the Vals Index, ahead of Opus 5.5 at 67.0% and GPT-6 Astra at 63.1%. 14 On Google's AutomationBench figure, which measures multi-step business processes, it scores 51.3%, roughly nine to ten points clear of Opus 5.5 and Astra. 45 It also posts 65.4% on Vals Finance Agent v2. 5

The most striking result is Harvey's Legal Agent benchmark. Google reports 19.6% for Argon against single-digit scores for every rival listed. 1 The company also cites leads in long-context retrieval, with GraphWalks at 99.7% up to 128K tokens and 84.2% in the 256K–1M range. 14 Further claimed leads include RiemannBench for math at 76.0%, the computer-use test Agent's Last Exam at 39.5%, and Vibe Code Bench at 91.9%. 4 Google has additionally promoted advantages in biology research. 4

Where the cracks show

Coding is the clearest weak spot, and even Google's own tables admit it. Argon sets a new high on DeepSWE v1.1 at 77.9%, ahead of Opus 5.5 and Astra at about 74%. 15 It then falls behind on FrontierSWE v2, trailing Astra by 10.5 points with 55.0% to 65.5%. On Terminal-Bench 4.0 it scores 57.4% against Opus 5.5's 66.4%. 1

The pattern is consistent. Argon is weaker where agents operate in a terminal or reason across large codebases. 1 Third-party results agree: Vals ranks it second behind Sonnet 5.5 on app building and code migration. 2

The independent index

Artificial Analysis gives Argon 53 on its Intelligence Index at high reasoning. 23 Outlets agree that this ties a top OpenAI configuration, though they describe that configuration differently. One says it ties GPT-6 Astra at max reasoning and edges past GPT-6.1 Sol at 52. 3 Another frames the tie as being with "GPT 6.1 Astra." 2 Anthropic stays ahead on this measure, with Claude Opus 5.5 at 58 and Sonnet 5.5 at 56. 2 One write-up calls the Opus gap "a point," but by those figures Opus leads by five points, which is a meaningful margin on a composite index. 2

Arena's Agent Arena leaderboard is less flattering. Argon ranks eighth there. 4

Artificial Analysis does credit Argon with two clear wins. It ranks first on that firm's AutomationBench-AA at 78%, and it has the lowest hallucination rate, 15%, of any model scoring 45 or higher on the index. 23 Note that the 78% figure comes from a different benchmark version than Google's 51.3% and should not be compared directly. Artificial Analysis also judges Argon's agentic performance to be notably stronger than earlier Gemini models. 3

How to read it

The vendor benchmarks are not wrong, but they are curated. Google chose the 19 tests and is highlighting the ones where it wins. 5 Neutral composite measures flatten those peaks into a tie. That gap is normal for frontier launches, yet it matters here because some coverage leads with "14 of 19." Other coverage, accurately, describes Argon as "a competitive entrant" rather than a leader. 45

Our reading: Argon is a specialist dressed as a generalist. It is best suited to document-heavy, multi-step professional work: legal drafting, financial research, business automation and long-context retrieval. Its low hallucination rate reinforces that fit, since accuracy matters most in those settings. 235

Teams building terminal-driven coding agents have less reason to switch. The evidence still favors Anthropic and OpenAI there. 12 On general reasoning, Google has closed the gap after a long quiet stretch but has not taken the lead. 34

For buyers, the practical takeaway is to test Argon on your own workload rather than trusting any single headline number.