Gemini 4 Argon explained: price, 1M context and what it really costs

Updated October 15, 2026. Researched and drafted with AI help; every figure is checked against the linked source.

Google made Gemini 4 Argon available to everyone on October 13, 2026, two weeks after announcing it for a limited group of cyber-defence partners and trusted testers. It is Google’s new top model in the Gemini API, ahead of Gemini 3.1 Pro, which is still a preview. The headline is the price: for its first seven weeks it costs the same as the mid-tier models from Anthropic and OpenAI, and then it doubles. This guide covers what it costs, what changes for developers, and how it compares with the alternatives, using the same list prices as our price table.

Gemini 4 Argon at a glance

	Through November 30, 2026	From December 1, 2026
Input (per 1M tokens)	$2.00	$4.00
Output, including thinking (per 1M)	$10.00	$20.00
Cached input (per 1M)	$0.10	$0.20
Batch input / output (per 1M)	$1.00 / $5.00	$2.00 / $10.00

Model ID: gemini-4-argon, listed as stable (not a preview).
Context: 1,048,576 tokens, shared between input, thinking and output. The default output limit is 65,536 tokens.
Input: text, images, video, audio and PDF. Output: text only.
Knowledge cutoff: March 2026, according to Google DeepMind’s model card.
Free tier: none. You need the paid tier of the Gemini API.
No long-context surcharge is listed on Google’s pricing page, unlike Gemini 3.1 Pro.

Google also lists Flex (same price as Batch) and Priority processing ($3.60 / $18.00 now, $7.20 / $36.00 from December 1).

The price doubles on December 1

Google’s blog calls $2 / $10 the introductory price and says it becomes $4 / $20 after November 30, 2026. Cached input goes from $0.10 to $0.20 per million at the same time. If you’re planning a project that runs into next year, budget with the higher number.

That shift moves Argon from one price class to another. Today it sits next to Claude Sonnet 5.5 and GPT-6.1 Sol, which also cost $2 / $10 with $0.10 cache reads. From December 1 it lines up with Claude Opus 5.5 at $4 / $20 and $0.20 cache reads.

What 1,000 requests cost

Same three workloads as our cost per 1,000 requests guide: a chat (2,000 tokens in, 500 out), a coding-agent step (2,000 fresh input tokens, 18,000 cached, 1,000 out) and a summary (10,000 in, 300 out). Each figure is the cost of 1,000 requests at standard list prices, with no batch discount.

Model	Price in / out	Cache read	1,000 chats	1,000 agent steps	1,000 summaries
Gemini 3.8 Flash (promo)	$0.75 / $3.75	$0.075	$3.38	$6.60	$8.63
Grok 4.7	$2 / $6	$0.50	$7.00	$19.00	$21.80
Gemini 4 Argon (intro)	$2 / $10	$0.10	$9.00	$15.80	$23.00
Claude Sonnet 5.5	$2 / $10	$0.10	$9.00	$15.80	$23.00
GPT-6.1 Sol	$2 / $10	$0.10	$9.00	$15.80	$23.00
Gemini 3.1 Pro (preview)	$2 / $12	$0.20	$10.00	$19.60	$23.60
Gemini 4 Argon (from Dec 1)	$4 / $20	$0.20	$18.00	$31.60	$46.00
Claude Opus 5.5	$4 / $20	$0.20	$18.00	$31.60	$46.00
Claude Fable 5.1	$10 / $50	$0.25	$45.00	$74.50	$115.00
GPT-6 Astra	$10 / $50	$1.00	$45.00	$88.00	$115.00

Until November 30, Argon is cheaper than Gemini 3.1 Pro on every workload. Its output is $10 instead of $12 per million and its cache read is half the price, so an agent step costs $15.80 instead of $19.60. There is little cost reason to stay on 3.1 Pro during the introductory period.

After December 1, it costs nearly twice as much as 3.1 Pro for chat ($18.00 against $10.00 per 1,000) and exactly the same as Claude Opus 5.5 on all three workloads. It is still well under the $10 / $50 flagships, Claude Fable 5.1 and GPT-6 Astra.

Output-heavy work still favours Grok 4.7 at $6 per million output tokens, as long as you don’t rely on caching.

Thinking tokens are billed as output

Google’s pricing page says the output price includes thinking tokens. Argon has three thinking levels (low, medium and high), and Google’s developer notes say medium is the default and recommend testing all three and picking the lowest one that is good enough.

This matters for your bill. If a chat request produces 500 visible tokens but also 3,000 thinking tokens, you pay for 3,500 output tokens. At the introductory price, 1,000 such chats cost $39.00 instead of $9.00. The thinking count varies by task, so measure it on your own prompts before comparing models on headline price alone.

Google also notes that max_output_tokens caps thinking and output together, so a low limit can cut a response short.

Caching works differently

Argon supports implicit caching only. Google’s developer notes say explicit caching (cachedContents) isn’t supported on this model; implicit caching is on by default for prompts of at least 4,096 tokens and gives a 95% discount on cached input. You don’t pay a storage fee, but you also can’t force a cache hit, so the agent column above is a best case: it assumes the 18,000 repeated tokens are served from cache every time.

If you’re coming from Gemini 3.1 Pro or 3.8 Flash and use explicit caches, you’ll need to remove that code. Google also says temperature, top_p and top_k are ignored on Argon and should be removed. Our prompt caching guide explains how much caching saves on other models.

Very long prompts: no surcharge

Several models charge more once a prompt gets very long. Google’s pricing page lists no long-context tier for Argon. Here is one request with a 300,000-token prompt and 2,000 output tokens:

Model	Rate that applies	Cost of one request
Gemini 4 Argon (intro)	$2 / $10	$0.62
Claude Sonnet 5.5	$2 / $10 (Anthropic prices the full 1M context at standard rates)	$0.62
Grok 4.7	$4 / $12 (200K tokens or more)	$1.22
Gemini 3.1 Pro	$4 / $18 (over 200K)	$1.24
Gemini 4 Argon (from Dec 1)	$4 / $20	$1.24
Claude Opus 5.5	$4 / $20	$1.24

At the introductory price, Argon is half the cost of 3.1 Pro on long documents. After December 1, the two cost about the same per request at this size. OpenAI’s pricing page also lists higher long-context rates for the GPT-6 models; see the GPT-6.1 Sol page for its threshold.

Google says the 1M-token limit is shared between input and output. A 1M-token prompt leaves no room for an answer, so in practice keep prompts well below the limit.

What Google says about quality

Google’s blog says Argon ranks first on the Vals Index, which covers finance, coding, legal and tax work. The DeepMind model card lists, among others, 68.9% on the Vals Index, 77.9% on DeepSWE v1.1, 57.4% on Terminal-bench 4.0 and 69.2% on the OSWorld-2.0 offline subset. These are Google’s own published numbers. The same card compares Argon with GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5, and rivals score higher on some individual benchmarks. Treat benchmarks as a starting point and test on your own tasks.

Long jobs can be slow: Google’s developer notes say hard prompts can take several minutes and recommend HTTP timeouts of at least 15 minutes or background execution.

Gemini 4 Argon in the Gemini app

Outside the API, Google says Argon is available in the Gemini app and AI Mode in Search for Google AI Pro and Ultra subscribers, and in Google Antigravity for Ultra subscribers. Google AI Pro is $19.99 a month and Ultra starts at $99.99. See all chat plans side by side on the plans page.

Should you switch?

From Gemini 3.1 Pro: yes, at least while the introductory price lasts. Argon is a stable release and cheaper on every workload above until November 30. Remove explicit caching and sampling parameters first.
For a long-running product: plan on $4 / $20 from December 1. At that price it competes with Claude Opus 5.5, not with Sonnet 5.5 or GPT-6.1 Sol.
For cheap bulk work: Gemini 3.8 Flash at $0.75 / $3.75 (promo through December 31) costs well under half of Argon’s introductory price on all three workloads.

Compare Argon head to head with Claude Opus 5.5 or any other model on the compare page, run your own token counts in the cost calculator, and check the Dumb meter to see how people rate it day to day.