A day after Meta said it had finally reached the frontier, Google answered with a model it has been sitting on for weeks. On Oct. 13 the company made Gemini 4 Argon available worldwide: through the Gemini API in AI Studio, for Google AI Ultra subscribers in the Antigravity agent platform, and for Ultra and Pro subscribers in the Gemini app and AI Mode in Search. Koray Kavukcuoglu, Google DeepMind's SVP and Google's chief AI architect, framed the drop as access to a system built for the hardest problems in coding, finance, legal work, cybersecurity and science.
The timing is not accidental. Argon was first revealed in late September and held for a phased Fairwind cybersecurity release before a broad paid rollout. In the months since Gemini 3 in late 2025, Anthropic and OpenAI shipped models that left Google looking late. The Verge noted that Argon lands one day after Meta's Muse Wonder Full, a reminder that the frontier is no longer a two-lab contest. Google's first major frontier release in nearly a year is meant to close that gap in public.
Independent numbers give the claim a sharper edge than a press blog alone. The Vals Index, which weights agent performance across finance, coding, legal and tax tasks by each sector's share of U.S. GDP, ranked Gemini 4 Argon first at 68.90 percent as of Oct. 13. Claude Sonnet 5.5 sat at 67.04 percent and Claude Opus 5.5 at 66.97 percent. Meta's Wonder Full placed fifth at 64.57 percent. Google's model card reports the same Vals Index figure and adds AutomationBench at 51.3 percent, Vals Finance Agent v2 at 65.4 percent, and Harvey's Legal Agent Benchmark at 19.6 percent, all above the GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5 scores listed in that table. No single model leads every component of the Vals Index. Argon leads Finance Agent v2. Claude Opus 5.5 still leads Terminal-Bench 4.0. The headline is leadership on a GDP-weighted composite, not a clean sweep.
What Google wants to sell is depth of work, not a prettier chat window. The blog post stresses a one-million-token unified input-output limit on the Gemini API, a continuation protocol for long agent runs, and internal results: more than 35 percent more agentic code submissions with nearly a third fewer rollbacks inside Antigravity; a Chromium refactor across hundreds of files that the company says compressed months into weeks; and open-source speedups on quantum chemistry and Gemma inference that Google says landed in days rather than months. Wiz, named as a Fairwind partner, is described as using Argon to find and fix a public file download bypass that exposed a master key and administrator access in healthcare infrastructure. Those are company-reported case studies. Outside labs will stress-test the harnesses. The product story is clearer: Argon is aimed at multi-step professional workflows where hedging and cliché answers have been a Gemini complaint for a year.
Pricing is the other lever. Argon launches at an introductory $2 per million input tokens and $10 per million output tokens, with cached input discounted 95 percent, through Nov. 30, 2026. After that the rate becomes $4 and $20, matching Meta's Wonder API and Anthropic's Opus 5.5 list prices. Ars Technica notes that even the post-promo rate undercuts GPT-6 Astra and Claude Fable 5 on a per-token basis, though token use per task can erase that gap. Access is gated: Pro and Ultra get the Gemini app; Antigravity and AI Studio agent workflows stay Ultra-only for now; free and Plus users are left waiting. Google has not said when, or if, Argon reaches lower tiers.
Safety paperwork tracks the Frontier Safety Framework. The model card says Google cannot rule out biological and cybersecurity critical capability levels without mitigations, and cannot rule out a stealth and situational awareness threshold. Residual risk assessments, the company says, bring those risks below the relevant thresholds with jailbreak defenses, an automated misalignment monitor, and prompt-injection hardening that cuts Gray Swan Indirect Prompt Injection attack success to 0.7 percent across 15 adaptive attempts. Shortcut attempts on coding benchmarks where models have no incentive to hold back fell to 2 percent, Google's best reported result for Gemini. A fuller safety and security report is linked from the card. None of that settles how enterprises should govern an agent with calendar, code and document access. It does show Google publishing the same style of risk case its peers now ship with a frontier drop.
For the industry, Oct. 13 resets the scoreboard more than it ends the race. Meta spent Monday claiming a seat at the frontier table. Google spent Tuesday putting a GDP-weighted leaderboard win and a temporary half-price API in front of paying customers, while keeping the model off free tiers and fencing the heaviest agent tooling behind Ultra. Argon is Google's answer to a year of trailing Anthropic and OpenAI: available today, scored first on Vals, priced to invite migration through November, and still short of a promise that every subscriber will get it. The frontier just got crowded again.
