Gemini 4 Argon Goes Public With Big Claims and Mixed Benchmarks

Google has released Gemini 4 Argon for broad use, opening access to a model built for coding, knowledge work, enterprise workflows, and cyber defense. The launch arrives on October 14, 2026, after Google first unveiled Argon on September 30 through its Fairwind Program, which initially limited access to approved cybersecurity defenders.

The release was a major focus at AIE NYC, where 20,000 people joined today. Argon is now available through the paid Gemini API and Gemini platform, Google AI Pro and Ultra in the Gemini app, AI Mode in Search, and Ultra in Antigravity. Google is also offering 10,000 scientists free access to Antigravity’s highest usage tier for six months through an application program.

Argon’s workplace pitch is built around long, difficult tasks

Google says Argon answers questions more directly with less hedging, avoids unnecessary repetition, and skips cliches throughout its responses. On longer tasks, it shares progress more often so users can follow what it is doing. The Gemini app now offers frontier reasoning and improved performance across complex tasks, while Gemini can connect to more than 40 apps, including Instacart, OpenTable, Adobe, and Shopify.

Google is also bringing Antigravity’s agentic coding interface directly into the Gemini macOS app, making it easier to move between chatting and coding. The company says Argon is available in Cursor, Devin Desktop/CLI, and GitHub Copilot, putting the model inside several developer tools rather than limiting it to Google’s own products.

Google’s internal results describe a model aimed at software work that normally takes months. Argon helped separate core webpage handling from a legacy Chromium folder containing approximately six million lines of browser feature code inside Chromium’s roughly 40-million-line codebase. Google says the work finished in weeks and eliminated a category of browser crashes.

Over four weeks, Google’s engineering teams increased agentic code submissions by 35% and reduced post-submission rollbacks by 30%. Google also says Argon reduced in-memory session preparation time by 40% and accelerated cold imports by 21% in its Python framework. On an Nvidia A100 GPU, inference throughput rose from approximately 37,900 to more than 108,000 tokens per second while preserving identical output.

Other examples reach beyond ordinary application code. Argon examined 134 aging software bugs and identified 97 as stale or duplicated in under three hours. Google says it helped identify and remediate a serious vulnerability in healthcare software through Wiz’s Scan for Good initiative, found redundant emoji rendering in YouTube’s comment system that reduced server load by 45%, and uncovered critical bugs in the proprietary drone physics simulator used in Project Wing.

The benchmark picture is strong, but not perfect

Google’s benchmark results place Argon first on the GDP-weighted Vals Index with a score of 68.90%. Claude Sonnet 5.5 scored 67.04%, while Claude Opus 5.5 scored 66.97%. Argon also led the Finance Agent v2 benchmark at 65.40% and scored 19.6% on Harvey’s Legal Agent Benchmark, compared with 5.4% for GPT-6 Astra and 3.8% for Claude Opus 5.5.

On Zapier’s AutomationBench, Argon reached 51.3%, ahead of Opus 5.5 at 42.5% and Astra at 41.4%. It scored 77.9% on DeepSWE v1.1, compared with 74.2% for Opus and 74.1% for Astra. Terminal-Bench 4.0 tells a different story: Opus 5.5 scored 66.4%, while Argon scored 57.4%.

Artificial Analysis assigns Argon an Intelligence Index score of 53, tying it with GPT-6 Astra and Claude Fable 5.1. That score sits below Claude Opus 5.5, Claude Sonnet 5.5, and Muse Wonder Full. Cursor reports 52.8% on CursorBench at default effort, while Cognition reports 60.5% on FrontierCode 1.1 and offers a 75% Devin discount through October 27.

Sam Witteveen, CEO and co-founder of Red Dragon AI, sees Argon’s strongest results on tasks that combine visual understanding, software generation, and creative media production. He called it “a really, really, really good model,” but added, “is it the best model out there? I actually don’t think it is.” Witteveen also said Antigravity is “nowhere near as good as Claude Code or Codex.”

Consumer access and pricing widen the launch

Google says Argon is better at solving spatial problems in 3D environments, a capability tied to its multimodal design. More than 100,000 Google employees use Argon each month through an internal version of Antigravity, giving the company a large workplace setting for testing agentic software tasks.

Google is pricing Argon at an introductory rate of $2 per million input tokens and $10 per million output tokens, with a 95 percent discount for cached tokens until the end of November. After November 30, the rates will rise to $4 per 1M input tokens and $20 per 1M output tokens.

The wider competition is not limited to coding. Meta’s launch of Muse emphasizes personal assistants as well as coding, and Alexandr Wang identifies consumer use in Muse as the primary development use case. Muse’s MemoryBench evaluates updated facts within a long conversation, but it does not establish durable cross-session memory. Wang also acknowledged deployment problems and said Meta had been fixing rollout issues.

For users, Argon’s launch comes down to access, price, and the kind of work they need done. Google is offering broad availability, app connections, coding tools, and a six-month program for 10,000 scientists. The benchmark results show clear wins in several agent tasks, alongside a weaker Terminal-Bench score and an Intelligence Index tie. Argon is powerful, but the numbers leave room for rivals to claim the lead.

Stay connected via Google News

Follow us for the latest AI News