Google's Gemini 4 Argon debuts with strong benchmark leads, but access remains restricted

Google's latest large language model, Gemini 4 Argon, demonstrated superior performance across the majority of tested benchmarks when compared to offerings from OpenAI and Anthropic, though the company is controlling its release carefully. The announcement arrived on Wednesday, marking the debut of what Google has positioned as its primary flagship offering. In head-to-head comparisons, Gemini 4 Argon prevailed outright or tied for first place in 13 of 18 benchmark tests against Anthropic's Opus 5.5 and Fable 5.1, as well as OpenAI's GPT-6 Astra, with some advantages reaching double-digit margins.

Google stated it is "actively engaged in the U.S. government's voluntary process for pre-release model access while we gradually expand access." The company announced this rollout just one day after CEO Sundar Pichai joined other technology leaders—including representatives from Anthropic, Meta, Nvidia, OpenAI, and SpaceX—in signing a commitment to self-regulate artificial intelligence development following a meeting with President Donald Trump. The commitment, which lacks any specified enforcement mechanisms, came ahead of Argon's release.

Argon represents the successor to Gemini 3.5 Pro, which Google had originally planned to introduce at its May I/O developer conference. The company had initially targeted a June launch for the new Pro model, but instead released a series of Flash variants during that period. Early access participants in the Fairwind Program will help shape the model's safety guidelines before broader availability. When released more widely, paid API customers and AI Ultra subscribers will receive access first, followed by developers, enterprises, and consumers.

Coding performance shows uneven results; knowledge work stands out

While Google emphasized Argon's coding capabilities, its performance in this domain proved inconsistent. The model achieved 77.9% on DeepSWE v1.1, establishing a new benchmark high, and scored 91.9% on Vibe Code Bench. However, Argon ranked last among the four tested models on both FrontierSWE v2 and Terminal-Bench 4.0, trailing GPT-6 Astra by 10.5 points and Opus 5.5 by nine points respectively.

The model's strongest performance emerged in knowledge work and automation tasks. On Zapier's AutomationBench, Argon scored 51.3%, nearly nine points ahead of Opus 5.5. For the GraphWalks test using inputs between 256K and 1M tokens, it achieved 84.2%, exceeding GPT-6 Astra by more than 12 points. "Its 19.6% on Harvey's Legal Agent Benchmark is nearly triple Fable 5.1's score, though that means it still fully completes only about one in five tasks." Additional advantages appeared on the Vals Index, Agent's Last Exam, Chartography, and shorter GraphWalks testing, though these leads typically remained under two points.

On CWE-bench v1, the sole cybersecurity benchmark included with competitor scores, Argon tied both GPT-6 Astra and xAI's Grok 4.7 at 68%, with Opus 5.5 scoring one point lower. The comparison carries an important caveat: OpenAI and Anthropic models operate within their respective agent frameworks—Codex and Claude Code—meaning the results reflect each model's performance alongside its integrated tooling rather than the models in isolation.

Output token capacity reaches one million

Argon introduces a notable capability where most competitors have not yet advanced: the ability to generate up to one million output tokens. This represents a substantial increase from the 64,000-token output limit of previous Gemini models, while one million input tokens has become standard across frontier models. Google explained that "when the model has the headroom to think deeply and generate hundreds of thousands of tokens in a single trajectory, it adds a new level of depth in reasoning to solve tough problems in one go."

Google trained Argon to independently discover, verify, and remediate software vulnerabilities. For Fairwind participants and Google's internal teams, the company is deploying the model without cybersecurity guardrails. Wiz, which Google acquired for $32 billion in March, has already integrated Argon into its Scan for Good initiative. Google reported that the model identified a critical vulnerability in healthcare software deployed across hospitals globally that earlier frontier models had not detected.

Argon achieved 85.8% on Google's internal vulnerability discovery benchmark and 70.9% on Wiz's penetration testing benchmark. Google provided comparisons only against its own Gemini 3.8 Flash Cyber, which scored 71.0% and 58.2% on these respective tests.

Pricing structure

During its introductory period, Argon will be priced at $2 per million input tokens and $10 per million output tokens. Following this initial phase, pricing will increase to $4 per million input tokens and $20 per million output tokens. The later output token rate aligns with Anthropic's pricing for Opus 5.5 at $20 per million output tokens.