Gemini 4 Argon Reaches the API at $2 and $10 Per Million Tokens, With a No. 1 Ranking and No Score

Google made Gemini 4 Argon available through the Gemini API on October 13, pricing it at $2 per million input tokens and $10 per million output tokens until November 30, when the rates double. The company's launch post says the model ranks first on the Vals Index, a benchmark of economic-sector tasks, and does not give the score.

The post is signed by Koray Kavukcuoglu, senior vice president of Google DeepMind and Chief AI Architect at Google. Argon first appeared on September 30 in a limited rollout to cybersecurity partners, according to a Reuters report carried by Khaleej Times. As of Monday afternoon, the October 13 post is the broadest statement of what Google is selling.

What the Argon API page lists

The model page gives the identifier gemini-4-argon, marks it Stable, and sets a context window of 1,048,576 tokens shared between input and output. Default output is capped at 65,536 tokens, and Google says long decode can raise that to the full window.

Input can be text, image, video, audio or PDF. Output is text only. Thinking is supported at low, medium and high settings. The page carries no knowledge cutoff and sends readers to separate pricing and rate-limit pages.

Item	Introductory, to Nov. 30	From Dec. 1
Input, per million tokens	$2	$4
Output, per million tokens	$10	$20
Cached input	95% off	95% off

The launch post says the lower rate applies to the public-access model. The 95% cached-input discount appears once in the post with no end date, so the table repeats it for both periods.

1. Introductory, to Nov. 3010 USD
2. Standard, from Dec. 120 USD

Source: Google, Gemini 4 Argon launch post, accessed 2026-10-19

Which numbers Google published

The post lists ERQA, a test of embodied reasoning, at 87%, and says that is the highest among frontier models. It reports 57% on RareBench and 52.2% on LabBench, both from private evaluations run by Gamow Labs, and a footnote says Gamow's views are not Google's.

On safety, Google says attack success on the Gray Swan indirect prompt injection benchmark fell to 0.7% across 15 adaptive attempts, and that shortcut-taking on coding benchmarks dropped to 2%. All of these are vendor-supplied or partner-supplied figures. The post links a model card as a PDF.

The post gives no figure for the Vals Index ranking it leads with, and no score on any named coding benchmark.

What Reuters reported on September 30

Reuters said Google's spokesperson described Argon as comparable to OpenAI's Astra and Anthropic's Opus on key coding and cyber benchmarks, all self-reported. Argon trailed on two of the four coding benchmarks Google included in that release, Reuters said.

Reuters also reported that Google no longer plans to release Gemini 3.5 Pro, which Sundar Pichai had said in the spring would launch in June. It reported that Demis Hassabis, founder and chief executive of DeepMind, is stepping aside in a restructuring of the lab, and that several Gemini model leaders have left. The Reuters text gives no reason and no successor. The October 13 post does not mention either change.

Cyber access and scientists

Google said it is adding hundreds of partners to its Fairwind Program, and that defenders can apply for access without cyber guardrails on a rolling basis. Wiz is named as using Argon in its Scan for Good initiative.

Separately, Google is offering 6 months of free access to Antigravity's highest usage tier to 10,000 scientists, subject to eligibility and selection. Consumers on Google AI Ultra and Pro get Argon in the Gemini app and in AI Mode in Search.

The introductory pricing ends on November 30. Google has not said whether it will publish the Vals Index score before then.

Related coverage: EmbeddingGemma 2's MTEB figures and Google Research's Harness2 repository.