Gemini 4 Argon: Pricing, Access, Benchmarks, Who Gets It

Google’s Gemini 4 Argon, the company’s newest flagship model, has moved out of its cyber-defenders-only preview and into paying customers’ hands. Following a two-week restricted period, Google opened Argon to Google AI Pro and AI Ultra subscribers and to developers on its paid API starting October 13, 2026, according to multiple reports, with an introductory price that is set to double after November 30. Coverage this week has focused on that pricing deadline and on a gap between Google’s benchmark claims and early third-party results.

From Fairwind to general access

Google announced Gemini 4 Argon on September 30, 2026, calling it its most powerful model to date. Unusually for a frontier launch, the first release went only to vetted security organizations in Google’s Fairwind Program. Google’s announcement said Argon can find, validate, and patch critical software vulnerabilities, and that a cyber-defense variant without standard guardrails would be available only to those vetted defenders. Google said it plans to expand Fairwind to hundreds more partners.

The staged approach mirrors moves by rivals. OpenAI restricted the cyber capabilities of its flagship in a tiered release, as covered in OpenAI’s Astra restricted cyber release, and Google previously shipped a locked-down security sibling alongside Gemini 3.8 Flash.

After safety testing, the broader rollout began October 13. According to reporting from Android Authority, Forbes, and others:

• AI Pro and AI Ultra subscribers get Argon in the Gemini app and in AI Mode in Google Search.
• AI Ultra subscribers also get Argon inside Antigravity, Google’s agentic coding environment. One report put that tier at $99.99 per month.
• Free-tier users do not have access.
• Developers and enterprises can call the model through the paid API as gemini-4-argon, with one report on October 15 describing general availability across Google’s APIs and cloud platforms.

Pricing: an introductory window

Google set introductory API pricing at $2 per million input tokens and $10 per million output tokens. Several outlets report those rates run through November 30, 2026, after which they rise to $4 per million input and $20 per million output from December 1. At least one benchmarking site already lists the higher rate, so developers should confirm current billing in their own consoles.

Forbes framed the launch as putting “a deadline on its price advantage”: at the introductory rate, Argon undercuts rival flagships on a per-token basis, but the doubled price narrows that gap. Pricing pressure has been intense this year, with OpenAI repeatedly cutting rates, as detailed in our coverage of the GPT-5.6 price cut and AI price war.

Specs and capabilities

According to third-party listings and early testing reports:

• Context window: about 1.05 million tokens, shared across input, reasoning, and output. One report puts the default output cap at 65,536 tokens. For what that budget means in practice, see our explainer on the context window.
• Reasoning levels: low, medium (the default), and high, with at least one API listing also showing a minimal setting.
• Targeted improvements: Google says Argon reasons better, writes with fewer clichés and less repetition, and performs more capably on finance, legal, cybersecurity, and science tasks. Google has positioned the model for long-running agent work.

Benchmarks: Google’s claims vs. early third-party results

Google’s announcement said Argon scored significantly higher than OpenAI’s GPT-6 Astra and Anthropic’s Claude Fable and Opus models across a range of benchmarks. Independent results so far are mixed:

• Artificial Analysis reportedly places Argon at 53 on its Intelligence Index, tied with GPT-6 Astra, and behind Anthropic’s Claude Sonnet 5.5 and Opus 5.5, according to one outlet’s summary.
• Vals AI ranks Argon first on its own composite index at 68.90%.
• One report cites 77.9% on the DeepSWE v1.1 coding benchmark, while noting competing models lead on other coding tests.

Tech Times reported on October 15 that Google’s headline scores reflect Argon’s high reasoning mode rather than the medium default subscribers receive in the Gemini app. A critical analysis published during the Fairwind period also noted that no independent lab had at that point reproduced Google’s figures. Benchmark comparisons between frontier models frequently depend on reasoning settings, tool access, and prompt formats, and results this early should be treated as provisional.

For context on the models Google is benchmarking against, see our coverage of Claude Opus 5.5 and GPT-6 Astra.

Where it fits in Google’s lineup

Argon is the first model to carry the Gemini 4 name. It follows a rapid series of Gemini 3.x releases through the summer, including multiple Flash updates aimed at high-volume, cost-sensitive workloads. Google has not said when lighter Gemini 4 variants will follow.

Putting Argon into AI Mode in Search gives the model immediate reach to a large paying audience and ties Google’s flagship directly to its core product. Gating Antigravity access to the Ultra tier, meanwhile, signals Google sees agentic coding as a premium feature — the same market where Anthropic and OpenAI have concentrated their enterprise efforts.

What it means

Google is back in the flagship race on its own terms. A staged, security-first launch followed by a broad paid rollout lets Google claim both responsibility and reach. The Fairwind period also positioned Argon as a cyber-defense tool before it was a chatbot, a framing that matters to the enterprise and government buyers all three frontier labs are courting.

The price is the pitch — for six weeks. At $2/$10, Argon is aggressively priced for a frontier model. The scheduled doubling to $4/$20 turns that into a trial offer. Developers who build on Argon now will face a real cost decision in December; Google is betting that switching costs and quality will keep them.

Benchmark claims need independent confirmation. Early third-party indexes put Argon in a tie or a narrow lead depending on methodology, not the clear sweep Google’s announcement described. Reports that headline scores used the high-reasoning setting reinforce the need to compare models at matched settings and matched cost.

Who wins and loses. Paying Google subscribers and API customers get a capable new model at a discount. Anthropic and OpenAI face renewed price pressure at the top end, especially in coding agents. Free-tier users are left out, and developers on tight budgets will need to plan for the December increase.

What to watch next:

• Independent evaluations at the default reasoning level, particularly on agentic coding and long-running tasks.
• Whether Google extends the introductory pricing past November 30 or adjusts it in response to competitor moves.
• Lighter Gemini 4 models for high-volume workloads, following the Flash pattern of the 3.x generation.
• Expansion of the Fairwind Program and any security findings Google publishes from its cyber-defense partners.