Gemini 4 Argon: What's New, Who Can Use It, Pricing, and Benchmarks

Google opened Gemini 4 Argon, its new top model, to paying users on October 13, 2026, two weeks after announcing it for a small group of cyber defenders. In the Gemini app it is meant for Google AI Ultra (from $99.99 a month in the US) and Google AI Pro ($19.99), but Google’s own pages disagree on how far Pro access has reached: the announcement says Pro subscribers can use Argon now, while the Google One plans page says it is rolling out to Pro in select countries. Free accounts and the $4.99 AI Plus plan do not include it.

Developers pay for Argon separately, through the Gemini API, at an introductory $2 per million input tokens and $10 per million output tokens until November 30, then $4 and $20. On Google’s own benchmarks Argon leads on legal and finance agent work, long documents, lab biology and math, and video, and trails Claude Opus 5.5 or GPT-6 Astra on terminal coding and the OSWorld computer-use test. The independent Artificial Analysis index scores it 53, level with GPT-6 Astra and behind Opus 5.5 at 58. We have not tested the model.

Gemini 4 Argon at a glance

• Announced: September 30, 2026, for cyber defenders in Google’s Fairwind Program; opened to paying users October 13
• Model ID: gemini-4-argon
• Inputs: text, images, audio, and video; output is text
• Context: up to 1 million tokens of input and output combined in the API; 1 million tokens in the Gemini app on Pro and Ultra
• Knowledge cutoff: March 2026, per the model card
• Independent score: 53 on the Artificial Analysis Intelligence Index at high thinking (October 13, 2026)

What is new in Gemini 4 Argon?

1. Three ways in, paid separately

Argon reaches people by three routes, and a subscription covers only the first two.

• Gemini app and AI Mode in Search: included with Google AI Pro and Ultra, in chats and in Tasks.
• Coding tools on a subscription: Google’s announcement gives Argon in Google Antigravity and in AI Studio to Ultra subscribers. The Google One plans page also lists expanded Argon access in Antigravity for Pro, which the announcement does not mention.
• Gemini API: pay per token, separate from any subscription. Anyone can link a billing account to an AI Studio project and prepay at least $5 to reach the paid tier, Pro subscribers included. The pricing page notes that Argon is not available on the API’s free tier or through the Google AI Pro plan itself.

Source: Google AI plans and Gemini Apps limits, October 13, 2026.

Limits are compute-based: a prompt costs more with a stronger model, a higher thinking level, a long chat, or a large file, and the allowance refreshes every five hours until a weekly cap. Google does not publish a number of Argon messages per plan. When a paid plan runs out, the chat continues on Flash-Lite.

2. Picking Argon in the app

Google’s model help page explains why a subscriber may not be using Argon without noticing. With auto model selection on, most prompts go to Flash-Lite and harder ones may use Flash or Pro; Argon is not among the automatic choices. With auto selection off, the page says responses use Flash-Lite. The way to reach Argon is to pick it: open the model menu under the chat box, choose Argon, and set a thinking level, because an unset level defaults to Low. Medium or High suit multi-step work and use more of your allowance.

To check which model wrote an answer, open More under the response. On the web, choose Show thinking steps and scroll to the bottom; in the Android app, choose See response details.

3. Tasks, connected apps, and coding on the Mac

The same update adds a toggle in the main chat window between chat and Tasks, which 9to5Google says was previously called Spark; our Gemini Spark explainer covers the earlier version. In Tasks, subscribers can choose Flash or Argon, and Google says Gemini can keep working while you are away from your computer. Gemini now connects to more than 40 apps, including Instacart, OpenTable, Adobe, and Shopify. On macOS, the Gemini app includes Antigravity’s agentic coding interface, so you can move from chatting to building an app in one window.

4. API price: introductory until November 30

Source: Gemini API pricing, October 13, 2026.

There is no free tier for Argon. Batch requests cost half, and only implicit caching is supported. Search grounding includes 5,000 free search queries a month shared across Gemini 3 and newer models, then $14 per 1,000; one request to Gemini can run several Google searches, and each one is billed. On the paid tier, Google says your content is not used to improve its products. From December 1, Argon’s token rates match Claude Opus 5.5’s.

5. Google’s benchmarks, with the footnotes

Google compares Argon with GPT-6 Astra, Claude Fable 5.1, and Claude Opus 5.5 on 19 rows. A selection shows the pattern; the full table is in the September 30 post.

Source: Google, September 30, 2026; numbers per Google.

The methodology explains how to read this. Argon ran at the highest thinking setting in the API. Competitor numbers are mostly their makers’ own reports at maximum settings, or the best available. Several Argon scores are Google’s own runs, including DeepSWE with a mini-swe-agent harness, Terminal-Bench, and OSWorld, where Google reports the best of three runs. LVBench gave each model a different number of video frames because of API limits.

Benchmark names also hide different versions. Google’s legal scores come from Vals AI, while other launches cite Artificial Analysis’s Harvey LAB-AA v1.1, which scores the same models differently (Opus 5.5 at 3.8% here, 4.2% there). Google called its DeepSWE result state of the art on September 30; Meta reported 79.0% for Muse Wonder Full twelve days later from a different run. Compare across vendors’ tables with care.

What independent testing shows

Artificial Analysis lists Gemini 4 Argon at high thinking with a score of 53 on its Intelligence Index (v4.3.2, model page checked October 13, 2026). That is level with GPT-6 Astra and Claude Fable 5.1, one point above Muse Wonder and GPT-6.1 Sol, and 5 points below Claude Opus 5.5. Argon cost $1.99 per index task at the introductory rates, against $3.26 for GPT-6 Astra and $5.98 for Opus 5.5 in the same runs. That figure is a weighted average over the index, not a forecast for your workload, and it was computed before the December price change. Argon generated 110 million output tokens on the suite, against an 85 million median; output speed has not been measured yet.

Safety findings

The model card says Google cannot rule out that Argon reaches its critical capability levels for biological information and cybersecurity, so it shipped safeguards it assesses as bringing risk below those levels. Google reports that external biology red teaming unlocked less than 1% of tested sub-tasks, and that attack success on Gray Swan’s indirect prompt injection benchmark fell to 0.7% across 15 adaptive attempts. Fairwind partners, a vetted group of cyber defenders, get a version without cyber guardrails. Known limitations include hallucinations and occasional slowness or timeouts.

What it means for everyday users

If Argon appears in your model menu, choose it where a stronger model changes the result: a long contract or report, a spreadsheet of financial data, a long video, or a plan with many steps. Keep simple questions on auto selection to save your allowance. If you pay for Pro and Argon is not in the menu yet, access may still be rolling out in your country.

Google’s demos show the intended range: turning a folder of project notes into a client proposal deck, then drafting the email and setting up a scheduled task for future requests; and, on the Mac, designing replacement sofa legs and producing a 3D-printable file from a sketch and a photo.

Two cautions. Argon’s knowledge stops in March 2026, so ask it to search for anything recent. And for health, legal, and money questions, check the answer against your documents: the model card lists hallucinations among known limitations, and strong benchmark scores in law and finance still mean errors on many tasks.

Our guide to what free Gemini includes covers the current free tier.

Who should try Gemini 4 Argon

• Google AI Pro subscribers: try Argon on long documents, finance and legal reading, and multi-step Tasks, and watch how fast it uses your five-hour allowance.
• Developers: test on the paid API before the December 1 price change; only implicit caching applies, which matters for apps that resend long prompts.
• Coding-agent users: Argon leads Google’s DeepSWE row but trails on Terminal-Bench and FrontierSWE; compare it with your current model on your repository.
• Cyber defenders and scientists: Google is expanding the Fairwind Program and offering 10,000 scientists six months of Antigravity’s highest usage tier at no charge, subject to selection.
• Not ideal yet: teams whose main workload is terminal-based coding, where Opus 5.5 scores higher on Google’s own table.

Reported versus published

FAQ