Gemini 4 Argon is open to paying users: what a task really costs, where it wins and trails, and what changes on December 1

Google opened Gemini 4 Argon to paid API developers, businesses and Gemini app subscribers on October 13. This page puts the live API rates next to an independent full-run cost, redoes that cost at the December 1 price, sorts Google's 19 published benchmarks into wins and losses by job, and maps who can use Argon where today.

This explains reporting by

Google, "Gemini 4 Argon" launch post (Koray Kavukcuoglu, October 13, 2026). Read the original first:

https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-argon/

In one minute #

• On October 13, Google made Gemini 4 Argon generally available. Paid Gemini API developers get it in AI Studio as gemini-4-argon, and Google AI Pro and Ultra subscribers get it in the Gemini app and AI Mode in Search. On September 30 it went only to cyber defenders.
• API price: $2 per million input tokens and $10 per million output tokens through November 30, 2026, then $4 and $20 from December 1. Batch is half that. Cached input is $0.10, then $0.20. There is no free tier.
• Per token, Argon matches GPT-6 Sol and costs a fifth of GPT-6 Astra and Claude Fable 5.1, according to Simon Willison's comparison.
• Per task is what you pay. Artificial Analysis spent $2,407 running its full test suite on Argon, against $5,324 for GPT-6 Astra and $8,708 for Claude Opus 5.5. At the December 1 price, the same run would cost about $4,814, still below both.
• Thinking is billed as output. In that run, reasoning cost $815 and the visible answers $313, so about 72% of the output bill was thinking.
• On Google's own table, Argon leads on knowledge work, long documents and video, and trails on terminal-driven coding and science tasks. Every one of those scores was run by Google.

What changed on October 13 #

Google announced Gemini 4 Argon on September 30, but only trusted cyber defenders in its Fairwind Program could use it. On October 13, in a post signed by Koray Kavukcuoglu of Google DeepMind, Google opened it to paying users.

• Developers: the Gemini API paid tier in AI Studio, plus Android Studio, with the model ID gemini-4-argon. Google's developer post says it works with the Interactions API and the google-genai SDKs from day one.
• Consumers: the Gemini app and AI Mode in Search, for Google AI Pro and Ultra subscribers.
• Google Antigravity, Google's agent coding tool: Google AI Ultra subscribers.
• Businesses: Gemini for business, the API and Antigravity.
• Scientists: applications are open for 10,000 free six-month grants of Antigravity's highest usage tier.
• Cyber defenders: a version without cyber guardrails stays in the Fairwind Program, on a rolling basis.

For most readers, the practical news is simple. Two weeks ago you could plan for Argon. Today you can test it.

The prices, all of them #

A token is a chunk of text, about three quarters of a word. Prices are per million tokens, from Google's Gemini API pricing page.

• Input: $2.00 through November 30, 2026, then $4.00 from December 1.
• Output, including thinking tokens: $10.00, then $20.00.
• Cached input: $0.10, then $0.20. Only implicit caching is supported, so you cannot pin a cache yourself.
• Batch: half price, $1.00 in and $5.00 out until November 30, then $2.00 and $10.00.
• No separate rate above 200,000 tokens of context. One flat rate.
• No free tier.

Simon Willison sets the launch price beside the rivals: the same as GPT-6 Sol, half of Claude Opus 5.5, and a fifth of Claude Fable 5.1 and GPT-6 Astra. After December 1, Argon's list price equals Opus 5.5's.

Why price per token is the wrong number #

A model that thinks longer, retries more or writes longer answers uses more tokens for the same job. So a cheap token can still mean an expensive task. The number that matters is what one finished task costs.

Artificial Analysis runs the same full test suite on every major model and publishes what it spent. Simon Willison charted those totals for 14 models. Gemini 4 Argon at high thinking cost $2,407. GPT-6 Astra at max cost $5,324. Claude Opus 5.5 at max with fallback cost $8,708, and Claude Fable 5.1 cost $13,129. The cheapest model on the chart, Qwen3.8 27B, cost $1,336, and GLM-5.3 cost $2,503.

So at launch prices, the same work cost about 2.2 times as much on GPT-6 Astra and about 3.6 times as much on Opus 5.5.

Now redo it for December 1. Every Argon rate doubles, so if the model uses the same tokens, the same run costs about $4,814. That is still below Astra's $5,324 and a little over half of Opus 5.5's $8,708. The gap narrows a lot, but it does not close. This is our arithmetic on Artificial Analysis's numbers, not a figure either company published.

Where the money goes: thinking is billed as output #

Google's pricing page says output pricing includes thinking tokens. The Artificial Analysis breakdown for Argon shows how much that matters:

- Reasoning (thinking): $815
- Output (the visible answer): $313
- Cache write: $897
- Cache read: $264
- Input that was not cached: $118

Thinking is about 72% of the output side of the bill. At $10 per million, that run spent about 112.8 million output tokens, most of them on reasoning the user never reads.

That gives you a lever. Argon runs at low, medium and high thinking levels. Willison's high-level test drawing cost him about 17 cents, and he rated the low and medium results excellent too. If a lower level passes your own checks, you may cut the biggest line on the bill.

Wins and losses, sorted by job #

Google published 19 benchmark scores comparing Argon with GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5. Google ran all of them. On that table Argon has the top score on 13, ties on 1 and trails on 5.

Where it leads:

• Knowledge work: Vals Index 68.9% (Opus 5.5 67.0%), AutomationBench 51.3% (Opus 5.5 42.5%), Vals Finance Agent v2 65.4% (Fable 5.1 58.9%), Harvey's Legal Agent Benchmark 19.6% (Fable 5.1 6.7%).
• Long documents: GraphWalks from 256k to 1M tokens, 84.2% against 71.8% for Astra.

- Video and charts: LVBench 91.7% (Astra 87.5%), Chartography 71.6% (Astra 71.0%).
- Some coding: DeepSWE v1.1 77.9% (Opus 5.5 74.2%), Vibe Code Bench 91.9%.

Where it trails:

• FrontierSWE v2: 55.0%, against 65.5% for Astra and 62.3% for Opus 5.5.

• Terminal-bench 4.0: 57.4%, against 66.4% for Opus 5.5.

• Terminal-Bench Science 0.1: 57.6%, against 68.1% for Astra.

• PostTrainBench: 45.3%, against 49.3% for Opus 5.5.

• OSWorld-2.0 computer use: 69.2%, against 72.6% for Astra. CWE-bench v1, which tests finding software weaknesses, is a tie with Astra at 68.0%.

The pattern is useful. Argon is strongest when the job is reading a lot and reasoning over it: contracts, finance, long logs, video. It is weaker when the job is driving a terminal step by step. If your agent lives in a shell, test before you switch.

Safety numbers worth knowing for agents #

Google says Argon leads the Gray Swan indirect prompt injection benchmark, with a 0.7% attack success rate across 15 adaptive attempts. Prompt injection is when text inside a web page, email or file tricks an agent into following it. For anyone letting an agent read customer email or browse, this is the number to watch.

Google also says harmful-request violations fell nearly 2-fold against earlier Gemini models, that external testers unlocked less than 1% of bio sub-tasks, and that Argon tried shortcuts on coding benchmarks 2% of the time. These are Google's own figures.

Who is affected #

What to do #

• Pick three real tasks you already run on another model, and run them on gemini-4-argon before November 30.
• Record the cost of each finished task and how often it passes your own checks, not the token price.
• Try low and medium thinking first; thinking tokens are billed as output and were most of the output bill in the Artificial Analysis run.
• Use batch mode for anything that can wait; it halves the price.
• Put a reminder on December 1 to rerun the cost math at $4 and $20.
• If an agent of yours reads untrusted email or web pages, add a prompt-injection test to your comparison.

What is still unknown #

• Every benchmark score on Google's table was run by Google. None of the 19 was run by an outside lab.
• The $2,407 full-run cost is Artificial Analysis's figure, as charted by Simon Willison, at launch prices. The $4,814 December figure is our arithmetic and assumes the same token use.
• Google's pricing page says only implicit caching is supported and lists no cache-write price, yet the Artificial Analysis breakdown shows $897 for cache writes. How cache writes are billed is not clear from Google's pages.
• Google does not state one context window. Its launch post describes a 1M unified input-output token limit, and its model page says outputs can run to 1M tokens.
• Google's pricing page says Argon is not included in the Google AI Pro plan, while the launch post says Pro subscribers get it in the Gemini app and AI Mode. The likely reading is that Pro covers the app, not the API, but Google has not said so in one place.
• The safety figures, including the 0.7% prompt-injection rate, are Google's own.

Sources #

AI News Report· every headline, every morning.