Gemini 4 Argon: what Google released

Google released Gemini 4 Argon in two steps. It announced the model on September 30, 2026 and made it generally available on October 13. Argon is the first model in the Gemini 4 family and, so far, the only Gemini 4 model on the Gemini API's model list.

Cyber defenders first, then paid customers

Koray Kavukcuoglu, Google's Chief AI Architect, announced Argon in a post on The Keyword. It was "rolling out to a set of trusted cyber defenders" through Google DeepMind's Fairwind Program. Google called this "a phased approach." The company was taking part in the U.S. government's voluntary process for pre-release model access and said the wider release would start "with paid API customers and Google AI Ultra subscribers."

That release came on October 13, when the Gemini API changelog marked gemini-4-argon "generally available (GA)" and a second Keyword post listed where to use it. So release dates differ by source: Artificial Analysis lists September 30, and OpenRouter lists October 13, the day it added the model.

What Google says Argon does

The model card describes Argon as "a large scale sparse mixture-of-experts (MoE)" model. It reads text, images, audio and video (plus PDFs on the Gemini API) and writes text. Google says it "excels at deep work in complex, real-world tasks" in coding, finance, legal work, cybersecurity and science, and the Gemini app announcement says it "answers questions more directly with less hedging." The knowledge cutoff is March 2026.

Argon always reasons before it answers, according to OpenRouter's documentation, and developers pick a thinking level: low, medium (the default) or high. Google's migration notes say Argon ignores temperature, top_p and top_k, recommend an output limit of at least 32,768 tokens, and warn that hard prompts can take several minutes.

In one of Google's examples, Google used Argon to speed up vLLM for Gemma 4 E2B, the small model we run, with changes that Google says cover everything "from attention tuning to speculative decoding." End-to-end throughput on an A100 GPU rose from about 37,900 to over 108,000 tokens per second, 2.85 times faster, "with 100% token parity." That was on a data-center GPU. Gemma on our Jetson Orin Nano boards still serves about 25 to 29 tokens per second.

A 1M-token window shared by input and output

Argon's context window holds 1,048,576 tokens, and input and output share it, thinking included. A standard request returns at most 65,536 output tokens by default (32,768 for non-streaming direct HTTP calls on Google Cloud). For longer answers Google offers long decode: the Gemini API writes the answer in hops of up to 65,536 tokens, each picking up where the last stopped, up to 1,048,576 tokens minus the input.

Each hop sends the prompt and everything generated so far back as input, and Google bills every token in every hop, thinking included. Implicit caching cuts the price of the repeated input by 95%. In Google's example, a 1,000-token prompt that produces 200,000 tokens reports about 397,000 input tokens.

Where you can use it

Channel	Access
Gemini app and AI Mode in Search	Google AI Pro and Ultra subscribers
Gemini API and Google AI Studio	Paid API tier; AI Ultra subscribers can also try it in AI Studio
Google Antigravity	AI Ultra subscribers and eligible Gemini Enterprise subscriptions
Gemini Enterprise Agent Platform and Gemini for business	Enterprise users and developers; API model ID gemini-4-argon
OpenRouter	Model ID google/gemini-4-argon, served by Google AI Studio and Google Vertex

Vertex AI became part of Gemini Enterprise Agent Platform in April 2026, though OpenRouter still lists the provider as Google Vertex. Google's Argon posts call the platform "Gemini platform."

Google's Gemini API pricing page lists Argon as unavailable on the free tier and, for API and AI Studio use, on the Google AI Pro plan. Pro subscribers get Argon in the Gemini app and AI Mode, as the table shows.

Google's safety post says Argon solved all 12 of Google's end-to-end simulated cyberattack scenarios, up from five for the previous generation. Google Cloud says standard Argon "has cyber mitigations that block offensive tasks such as pentesting and exploitation." Only "a select set of trusted defenders," vetted through the Fairwind Program, get a variant without those guardrails, Gemini 4 Argon Cyber Permissive.

What it costs

API prices are introductory through November 30, 2026, and double on December 1. These are per million tokens on the paid tier, with thinking tokens billed as output:

Tier	Input through Nov 30	Output through Nov 30	Input from Dec 1	Output from Dec 1
Standard	$2.00	$10.00	$4.00	$20.00
Batch and Flex	$1.00	$5.00	$2.00	$10.00

Cached input on the standard tier costs $0.10 per million tokens through November 30 and $0.20 from December 1. On October 14, OpenRouter listed the same $2 and $10 standard rates.

Benchmarks, and who ran them

Google's model card compares Argon with three rival models. Selected rows, as Google reports them:

Benchmark	Gemini 4 Argon	GPT-6 Astra	Claude Fable 5.1	Claude Opus 5.5
Vals Index	68.9%	63.1%	65.8%	67.0%
DeepSWE v1.1	77.9%	74.1%	67.4%	74.2%
FrontierSWE v2	55.0%	65.5%	56.3%	62.3%
GraphWalks BFS F1, 256k to 1M	84.2%	71.8%	65.0%	66.8%

Argon has the top score on 13 of the 19 rows in Google's full table. It ties GPT-6 Astra on CWE-bench v1 and trails at least one rival on FrontierSWE v2, Terminal-Bench 4.0, PostTrainBench, Terminal-Bench Science 0.1 and OSWorld-2.0.

We checked two independent sources. The Vals AI leaderboard, updated October 13, ranks Argon first on the Vals Index at 68.90%, ahead of Claude Sonnet 5.5 at 67.04% and Claude Opus 5.5 at 66.97%. Sonnet 5.5 is missing from Google's table, which lists Opus 5.5 at 67.0%. Artificial Analysis scores Gemini 4 Argon (High) at 53 on its Intelligence Index, where the median for comparable models is 27. Argon generated 110 million tokens during that evaluation; the median is 85 million.

How Argon relates to earlier Gemini models

The comparison that matters to us is Gemini 3.8 Flash, which Gemma Gemini used before Argon. Google announced Gemini 3.8 Flash on September 2, 2026, the day it became generally available, and called it "our most intelligent workhorse model." Its model card says it is based on Gemini 3.7 Flash, while Argon's model card calls Argon "a new frontier model."

Gemini 3.8 Flash has the same 65,536-token output limit per request. Its introductory price is $0.75 per million input tokens and $3.75 per million output tokens, so Argon costs about 2.7 times as much per token at introductory prices.

ServiceNow, an early-access customer, said in Google Cloud's announcement that Argon delivered "roughly 20% improved task success" over Gemini 3.8 Flash on NowAI-Bench. That claim is ServiceNow's own.

Where Argon fits in Gemma Gemini

Gemma Gemini uses Argon as its large model. Messages the router judges too hard for Gemma 4 E2B on a Jetson Orin Nano, any message with an image or PDF, messages where you pick Gemini 4 Argon yourself, and messages that arrive when no Jetson can answer go to Argon through OpenRouter with reasoning effort set to high. We explain the choice in Why Gemini 4 Argon is the big model in Gemma Gemini.