Gemini 4 Argon Pricing: What's New and Who It's For

Gemini 4 Argon is Google's flagship reasoning model, generally available in the Gemini API since 13 Oct 2026. Gemini 4 Argon pricing is $2/$10 per million input/output tokens through 30 Nov 2026, then $4/$20 from 1 Dec 2026. It has a unified 1M-token window, long decode up to 1M tokens, and implicit caching only.

That introductory price is the story for buyers. For the next five weeks, Google's top model costs the same per token as mid-tier models such as GPT-6.1 Sol and Claude Sonnet 5.5. On 1 Dec 2026 the price doubles, and Argon lands on exactly the same rate card as Claude Opus 5.5.

We checked every price in this post against the vendors' own pricing pages on 26 Oct 2026. This guide covers what is new in Argon, what it costs on every tier before and after the change, what real workloads cost against its rivals, and what to change in your code before you switch.

Gemini 4 Argon At A Glance

Google calls Argon "our most capable model built for complex coding, agentic, math, and science tasks" in its Gemini API release notes. So far it is the only Gemini 4 model on Google's models page.

• Released: Generally available on 13 Oct 2026, with the model ID gemini-4-argon.
• Price now: $2 input and $10 output per million tokens, through 30 Nov 2026.
• Price from 1 Dec 2026: $4 input and $20 output per million tokens.
• Cached input: $0.10 per million tokens, rising to $0.20. Implicit caching only.
• Batch and Flex: $1/$5 now, $2/$10 from 1 Dec 2026.
• Context: One window of 1,048,576 tokens, shared by input, thinking and output.
• Output: 65,536 tokens by default, or up to 1,048,576 with long decode.
• Inputs: Text, image, video, audio and PDF. It outputs text only.
• Where to use it: The Gemini API and Google AI Studio.
• Free tier: Not available. You need a paid Gemini API account.

Our Google Gemini provider page lists Argon beside every other Gemini model we track.

What's New In Gemini 4 Argon

Google's What's new in Gemini 4 Argon guide says the model "can reason longer on hard problems for better results on complex coding, agentic, math, and science tasks". Most of the API changes follow from that idea.

One 1M Token Window For Input, Thinking And Output

On Argon, input, thinking and output share one window of 1,048,576 tokens. A 900K token prompt therefore leaves room for about 148K tokens of thinking and answer combined.

Google's pricing page lists one rate for Argon, with no separate rate for long prompts. That matters for document-heavy work, because several rivals charge more once a prompt passes a size. We come back to this in the cost examples below.

Long Decode: Responses Up To 1M Tokens

A single Argon response can now run past 64K tokens, up to 1M tokens minus your input. Google calls this long decode. The model writes long answers in hops of up to 64K tokens, and each hop continues where the last one stopped.

To use it, you turn on background execution and set max_output_tokens above 65,536. Google's servers then run the continuation hops for you. If you turn on background execution without setting max_output_tokens, the server stops after 64K tokens.

For complex reasoning and coding, Google suggests a max_output_tokens of 131,072 or 262,144. That is a lot of output at $10 per million, so set the limit with your budget in mind.

Thinking Levels Replace Sampling Controls

Argon has three thinking levels: low, medium and high. Medium is the default, and Google recommends it for most tasks, including complex code and agents.

• Low: Fewest tokens and lowest latency. Suited to simpler and cost-sensitive work, though quality can drop on hard tasks.
• Medium: Balanced. Google's recommended starting point.
• High: Most tokens and the slowest first token. For hard maths, science and the most difficult code or agent tasks.

Google's advice is to run your evaluations at all three levels and pick the lowest that meets your quality bar. That is also the simplest way to control cost, because thinking tokens are billed as output.

The sampling controls are gone. Argon ignores temperature, top_p and top_k, so remove them from your requests.

Agents, Video And Tools

Two more additions arrived with Argon. Agentic video understanding is now generally available, and the Antigravity agent in Gemini Managed Agents supports Argon.

According to Google's Gemini 4 Argon model page, the model supports function calling, code execution, structured outputs, file search, URL context and grounding with Google Search and Google Maps. Computer use is supported as a preview. Argon does not support the Live API, image generation or audio generation, so voice and image work needs a different Gemini model.

Gemini 4 Argon Pricing Explained

Introductory Price Vs List Price

All rates below are US dollars per million tokens, from Google's Gemini Developer API pricing page. Output prices include thinking tokens.

Every rate doubles on 1 Dec 2026. Google has given the end date, so plan your budget on the list price from December rather than hoping the offer is extended.

Grounding with Google Search comes with 5,000 free search requests a month, shared across all Gemini 3 and newer models. After that it costs $14 per 1,000 requests. Grounding with Google Maps has the same allowance and price.

Batch, Flex And Priority Tiers

Batch and Flex both halve the standard rate, including cached input. If a job does not need an instant answer, such as a nightly run over a document archive, one of these tiers is the simplest saving available. Argon supports both, as well as Priority.

Priority costs 1.8 times the standard rate. During the introductory period, Priority Argon at $3.60/$18 is still a little cheaper than standard Claude Opus 5.5 at $4/$20.

Implicit Caching Only

Argon does not support explicit context caching. You cannot create a cache and reference it from later requests. Instead, implicit caching is on by default for prompts of at least 4,096 tokens, and a cache hit discounts the cached input by 95%.

That is a generous discount, but you do not control when a hit happens. For cost planning, assume fewer hits than you would get with an explicit cache, then measure your real hit rate in the first week. Our cost optimisation FAQ covers caching and batch in more depth.

What Real Workloads Cost On Argon And Its Rivals

Per-token prices only tell you so much. Here are three worked examples at standard rates, using the same token counts for every model. We leave out any cache-write charges, which some vendors bill on top.

The output counts include thinking tokens. Models think different amounts on the same task, so treat these as a guide and test your own traffic before relying on them.

Coding agent task: 1.5M input tokens across the whole task, 90% of them cache hits, plus 100K output tokens. Every request stays under 200K tokens, so no long-prompt rates apply.

Long document review: One 600K token prompt with no cache hits, such as a large contract bundle, and a 20K token answer.

Support assistant: 100,000 conversations a month. Each one has a 6K token system prompt and knowledge base that hits the cache, 2K tokens of fresh input and a 500 token reply.

Four things stand out.

• The introductory price is a real bargain: Until 30 Nov, Argon costs the same as GPT-6.1 Sol and Claude Sonnet 5.5 on the agent task and the support assistant, and about half as much as Opus 5.5 on all three workloads.
• From December, Argon and Opus 5.5 cost the same: The two flagships match on input, output and cached input, so the choice comes down to quality and token use on your own tasks.
• Long prompts favour Argon over GPT-6 Astra: Astra bills the whole request at $20/$75 once a prompt passes 272K tokens. Argon keeps its single rate across the full window. GPT-6.1 Sol rises to $4/$15 past the same size, which Argon beats until 30 Nov but not after.
• High volume rewards a smaller model: The support assistant costs $382.50 a month on Gemini 3.8 Flash, against $1,920 on Argon from December. Short, repetitive replies rarely need a flagship.

Grok 4.7 looks competitive on the agent task, but its cache reads cost $0.50 per million, five times Argon's introductory rate. Its 500K token window also rules it out for the largest documents.

Try your own token counts in our AI pricing calculator.

Gemini 4 Argon Vs Claude Opus 5.5, GPT-6 Astra And GPT-6.1 Sol

Against Claude Opus 5.5, Argon is half the price until 30 Nov, then exactly the same. Opus 5.5 launched on 22 Sep 2026 with a 1M token context window and a 128K max output on standard requests.

Argon's long decode lets a single response run much further, up to 1M tokens minus the input. Both charge $0.20 for cached input from December, but Argon's caching is implicit only. Our Claude Opus 5.5 guide covers Anthropic's side in detail.

Against GPT-6 Astra, Argon is far cheaper. Astra costs $10/$50, five times Argon's introductory rate and two and a half times its list price. Astra's cached input costs $1.00 per million, ten times Argon's introductory cache rate. Above 272K tokens Astra's price rises to $20/$75, which is why the long document example costs nearly ten times Argon's introductory figure on Astra.

Against GPT-6.1 Sol, the match is close until December. Both cost $2/$10 and both charge $0.10 for cached input. Sol's rate rises to $4/$15 above 272K tokens, while Argon's does not. From 1 Dec 2026, Sol becomes the cheaper option: half Argon's list price up to 272K tokens, and $4/$15 against $4/$20 above it.

OpenAI also launched GPT-6.1 Terra on 26 Oct 2026 at $1/$5, with cached input at $0.05. That is a quarter of Argon's list price, and worth testing if you need a lower-cost model for everyday work. For a wider view, see our AI API comparison.

The Cheaper Google Option: Gemini 3.8 Flash

If Argon's list price is more than you need, Gemini 3.8 Flash is Google's "most intelligent Flash model". It costs $0.75/$3.75 through 31 Dec 2026, then $1.50/$7.50 from 1 Jan 2027. Unlike Argon, it is available on the free tier.

Migration Checklist Before You Switch

Moving to Argon is more than a model name change. Google's migration checklist includes these steps:

• Update the model ID: Change your model string to gemini-4-argon.
• Raise max_output_tokens: Google recommends at least 32,768, because Argon can spend more of the limit on thinking than earlier models.
• Set a thinking level: Choose low, medium or high, and test your workload at all three.
• Remove explicit caching: Delete calls that create or reference cached content. Implicit caching takes over automatically.
• Remove sampling parameters: Temperature, top_p and top_k are ignored.
• Use background execution for long outputs: Anything over 64K tokens needs background set to true and max_output_tokens above 65,536.
• Increase client timeouts: Hard prompts can take several minutes, so Google suggests HTTP timeouts of at least 15 minutes.
• Keep conversation state on the server: Use previous_interaction_id for multi-turn and function calling conversations.

The max_output_tokens change is the one most likely to catch you out. An app that caps output at 8K tokens may see answers cut short once Argon spends part of that limit thinking. For step-by-step help with model upgrades, see our migration guides.

Who Gemini 4 Argon Is For

Try Argon now if you run:

• Long document work: Contract review, research synthesis or codebase analysis with prompts above 272K tokens, where GPT-6 Astra's long-prompt rate bites.
• Very long outputs: Reports, full code files or proofs longer than the 128K max output that Opus 5.5 and GPT-6 Astra offer on standard requests.
• Hard coding, maths and science: The tasks Google built Argon for, ideally tested at medium and high thinking.
• Video and multimodal agents: Agentic video understanding is now generally available on Argon.
• Anything you can test before December: The introductory price makes November a cheap month to run evaluations at scale.

Look elsewhere if you need:

• A free tier: Argon is paid only. Gemini 3.8 Flash has free access.
• Explicit caching: Argon does not support creating a cache and reusing it. If your costs depend on predictable cache hits, test your real hit rate before you commit.
• Voice or image generation: Argon has no Live API, image generation or audio generation.
• The lowest cost for short, simple tasks: From December, GPT-6.1 Sol and Claude Sonnet 5.5 cost half Argon's list price, and GPT-6.1 Terra a quarter.

Frequently Asked Questions

How Much Does Gemini 4 Argon Cost?

Gemini 4 Argon pricing is $2/$10 per million input/output tokens through 30 Nov 2026. From 1 Dec 2026 it costs $4/$20. Cached input costs $0.10 per million, rising to $0.20. Batch and Flex are half the standard rate.

When Does The Gemini 4 Argon Introductory Price End?

Google's pricing page says the introductory rates run through 30 Nov 2026. The $4/$20 list price starts on 1 Dec 2026, and every tier doubles at the same time.

Is Gemini 4 Argon Available On The Free Tier?

No. Google's pricing page says Gemini 4 Argon is not available on the Free tier. You need a paid Gemini API account to use it.

What Is Gemini 4 Argon's Context Window?

Argon has one window of 1,048,576 tokens, shared by input, thinking and output. A response defaults to 65,536 output tokens, or up to 1,048,576 with long decode, minus your input.

Is Gemini 4 Argon Cheaper Than Claude Opus 5.5?

Until 30 Nov 2026, yes: $2/$10 against $4/$20, with cached input at $0.10 against $0.20. From 1 Dec 2026, the two cost the same on input, output and cached input. You can follow every change on our price changes page.

The Bottom Line

Gemini 4 Argon is Google's most capable model and, until 30 Nov 2026, it is priced like a mid-tier one. At $2/$10, it costs half as much as Claude Opus 5.5 and a fifth as much as GPT-6 Astra. Its flat rate across a 1M token window makes it especially strong for long documents, where GPT-6 Astra's price rises to $20/$75 past 272K tokens.

From 1 Dec 2026, Gemini 4 Argon pricing rises to $4/$20, and it becomes a straight contest with Opus 5.5. Use November to run your evaluations at the introductory price, measure cost per task rather than cost per token, and decide before the rise. Then plug your numbers into the compareai.today pricing calculator and watch the pricing history explorer for Google's next move.