Before Gemini 4 Argon API Prices Double: Full Pricing Table and Selection Advice

Google released its flagship model Gemini 4 Argon on October 13, and gave it a price tag with an expiration date: $2 per million tokens for input and $10 per million tokens for output at the standard tier, valid through November 30. Starting December 1, prices double across all tiers. If you're planning to build a product on it, the next few weeks are the cheapest it will ever be.

What Is Gemini 4 Argon

gemini-4-argon is Google's new flagship, positioned for "deep work"—complex multi-step tasks in coding, finance, law, cybersecurity, and scientific research. It has been generally available (GA) through the paid Gemini API since October 13. Two facts worth knowing upfront: it's not available on the API free tier, and this launch comes after a turbulent period at Google—the originally planned Gemini 3.5 Pro was scrapped, and DeepMind's leadership was reshuffled after Demis Hassabis stepped back. This model has a lot to prove.

Full Pricing Table

Google published tier-by-tier pricing with an introductory period running through November 30:

Tier	Input/1M	Output/1M	Cached Input/1M	From Dec 1 (Input/Output/Cached)
Standard	$2.00	$10.00	$0.10	$4.00 / $20.00 / $0.20
Batch & Flex	$1.00	$5.00	$0.05	$2.00 / $10.00 / $0.10
Priority	$3.60	$18.00	$0.18	$7.20 / $36.00 / $0.36

A few details worth noting: output pricing includes thinking tokens—the model's internal reasoning is billed. Cached input (text the model has already processed) gets a 95% discount, which is unusually generous and has a big impact on multi-agent workflows that repeatedly read context. The Batch and Flex tiers trade speed for half price, while the Priority tier costs more for faster, more stable responses.

For reference, Gemini 3.8 Flash is listed on the same page at $0.75 input / $3.75 output—at the introductory price, Argon is roughly 2.7x Flash, and 5.3x after December.

Thinking tokens: Intermediate tokens generated when the model performs internal reasoning before producing its final answer. Traditional chat models bill only for visible replies, but frontier models with chain-of-thought reasoning (such as OpenAI's o-series, Anthropic's extended thinking, and DeepSeek^Ad-R1) "think" for a while before answering. This thinking content consumes compute even if it isn't shown to users, so it's counted toward output costs. For developers, this means the same question can produce wildly different bills depending on how long the model reasons—in complex multi-step tasks, thinking tokens can even exceed the tokens in the visible answer. When designing prompts, controlling irrelevant reasoning and encouraging more concise thinking paths is a key way to reduce costs.

Cached input (prompt caching): When the same long text (such as a system prompt, codebase, or knowledge base document) appears repeatedly across multiple API calls, model providers allow the intermediate state of that content to be cached, with subsequent calls billed only at the "cache hit" price. Argon's 95% discount means input originally priced at $2/1M costs just $0.10/1M on a cache hit. This is critical for multi-agent, RAG, and long-context conversation scenarios: if a 100K-token system prompt is cached, the cost per call can drop from $0.20 to roughly $0.01. Caches usually have a time-to-live limit (ranging from minutes to hours) and require a strictly consistent prompt prefix to hit, so engineering-wise you should keep static content at the front and variable content at the back.

Where You Can Actually Call It

• Paid Gemini API and Google AI Studio (gemini-4-argon)
• Antigravity, Google's agentic coding environment (individual users need an Ultra subscription)
• Gemini Enterprise enterprise subscription
• AI Mode in the Gemini app and Search (Pro and Ultra subscribers)
• Android Studio and Firebase AI Logic (for Android developers)

The API has a uniform limit of one million input and output tokens, with longer tasks handled through a continuation protocol. In other words, extremely long-context tasks can't be stuffed in all at once—they need to be stitched together according to the protocol.

Antigravity: This is Google's "agent-first" coding environment launched in 2025, positioned somewhere between a VS Code extension and a cloud IDE. Its core isn't code completion but letting multiple AI agents autonomously plan, execute, and verify development tasks—for example, fixing a bug, implementing a feature module, and automatically running tests. Agents can read the entire codebase, call the terminal, browse the web, and commit code. Google uses it to demonstrate Argon's "deep work" capabilities while also using it as an enterprise distribution channel: individual developers need an Ultra subscription, while enterprises access it through Gemini Enterprise. For developers, Antigravity is the most direct way to experience Argon's agentic capabilities, but it also means the official performance numbers all come from this controlled environment.

Which Google Claims Deserve Scrutiny

Google says teams using Argon in Antigravity saw agent code commit volume increase by more than 35%, while rollback rates dropped by nearly a third. Early enterprise testers include Deloitte, LexisNexis, Ambience Healthcare, and Isomorphic Labs. These numbers come from Google's own tools and its own methodology, and no independent third-party benchmarks have been published yet. Until third-party verification arrives, treat them as marketing material.

Google is also making two distribution moves at the same time: opening the Fairwind program to hundreds of cybersecurity partners, giving vetted defenders access to a version of Argon with network guardrails removed; and providing 10,000 scientists with six months of free access to Antigravity's top tier. Giving scientists free credits is a clever distribution play—getting the model into labs before competitors do.

Independent benchmarking: AI model capability evaluation typically falls into two categories—vendor self-testing and third-party evaluation. Vendor self-testing often uses in-house tools and in-house task sets, making it easy to cherry-pick favorable scenarios; third-party evaluations (such as LMSYS Chatbot Arena, SWE-bench, ARC-AGI, and GAIA) use publicly standardized tasks and are run by external organizations, making results reproducible and comparable. For coding and agentic tasks, SWE-bench Verified and Terminal-Bench are common yardsticks. The article's emphasis on "no independent benchmarks published" means Argon has not yet appeared on these public leaderboards or in evaluations by external organizations. In the absence of third-party data, numbers like a 35% increase in commits and lower rollback rates can't be compared across models, nor can bias from tooling or task selection be ruled out. For enterprise procurement, waiting for Arena or SWE-bench results is standard practice for reducing selection risk.

Should You Build on It Now

Let's do an honest calculation: if your workload fits, the introductory pricing window (through November 30) is a real price advantage—the standard tier's $2/$10 is lower than most frontier models. But architecturally, you should design for December's pricing ($4/$20), because that's the cost you'll bear long term. The 95% cached input discount is an easily overlooked key item: for agents that repeatedly read long contexts, actual costs may be far below list price.

The risk lies in Google's recent track record—the scrapped 3.5 Pro and leadership turmoil leave little confidence in roadmap stability. The sensible approach: prototype while it's cheap, then decide whether to commit seriously once independent benchmarks arrive.

Frequently Asked Questions

Is Gemini 4 Argon on the free tier? No, it's only available through the paid API.

When does the introductory price end? November 30, 2026; standard rates apply from December 1.

What's the cheapest way to try it? The Batch tier during the introductory period at $1/$5 per million tokens—provided you can accept slower responses.

How does it compare with Claude and GPT pricing? At the introductory price it's in the competitive range with frontier models, and after December it lands in the same bracket. The cached input discount is more generous than most competitors, which is crucial for agentic scenarios.

---

Pricing data reflects rates published by Google; performance claims come from Google and have not been independently verified.