Gemini 4 Argon Opens Up at $2 per Million Tokens: What the New API Pricing Changes for Builders

Posted on by
Contents show
Gemini 4 Argon API pricing falling, shown as descending bars and a downward trend line

Gemini 4 Argon spent its first two weeks behind a velvet rope. Gemini 4 Argon was announced on 30 September and handed first to a set of trusted cyber defenders, with Google saying access would widen gradually. The company's Gemini API documentation now lists Gemini 4 Argon as a stable model, last updated on 14 October, which closes that earliest phase and opens the model to the paid API customers Google said would come next. The model also reached consumers through AI Mode in Google Search, which changed the surface publishers see.

Gemini 4 Argon is the headline elsewhere. For anyone paying the bill, the more consequential number is smaller and easier to miss: an introductory price of 2 dollars per million input tokens and 10 dollars per million output tokens, with cached input tokens discounted by 95 percent.

Read alongside Anthropic's October price moves on its small models, that single line describes the shape of API pricing across the market, and the shape decides which products get built. The capability side of the model is a separate story, and we covered it in what Gemini 4 Argon did in an overnight coding run. The interesting question is no longer which frontier model is best. It is which tier each job actually needs, and how much of your bill caching can absorb.

What Google is charging, line by line

Four numbers decide most of the API pricing you will pay for Gemini 4 Argon.

Input tokens cost 2 dollars per million. That covers everything you send, including the PDFs, images, audio and video Gemini 4 Argon accepts. It is the dominant cost in document-heavy and retrieval-heavy work, because the same context tends to be resent on every turn.

Output tokens cost 10 dollars per million. Output is the expensive direction by five to one, which changes how you should think about prompting. A model that thinks longer, or an agent that narrates every step, spends real money.

Cached input tokens are 95 percent off. That brings cached reads to roughly 10 cents per million tokens on Gemini 4 Argon, and it is the largest single lever in the pricing table. Repeated context, such as a system prompt, a codebase excerpt or a long document set, should be structured so that it can be reused rather than retransmitted at full price.

Three consumption modes are supported. Batch API, flex inference and priority inference. Batch is for work that can wait and typically comes with a discount, priority is for work that cannot, and the choice is now a line item rather than an afterthought.

One detail deserves emphasis because it undercuts the caching story slightly. The model supports implicit caching only. Explicit caching, where you deliberately create and manage a cache, is not supported for Gemini 4 Argon. That means cache hits depend on how you structure prompts and how the platform detects reuse, rather than on you pinning context deliberately.

Google describes the 2 and 10 dollar figures as introductory pricing, and the model page does not commit to how long that lasts. Budget accordingly: a cost model built on an introductory price is a model with an expiry date.

The batch, flex and priority decision

Gemini 4 Argon supports three consumption modes, and treating them as interchangeable leaves money on the table.

Batch is for anything that can wait. Overnight content generation, bulk classification, backfills and evaluation runs belong here, and providers price batch work below interactive rates.

Flex inference sits between batch and priority, for work that is not urgent but cannot wait for a daily job.

Priority inference is for the requests a person is actually waiting on. Paying a premium there is often correct, because a slow answer that arrives after the user has given up costs more than the tokens saved.

The discipline is deciding the mode when you design the feature rather than after the first bill arrives.

The specifications that decide your bill

The Gemini 4 Argon documentation answers the questions that matter for cost, and several of the answers shape architecture rather than prompting.

PropertyDetail
Context window1,048,576 tokens, shared between input and output
Default maximum output65,536 tokens
Long decode maximum outputUp to 1,048,576 tokens
Input typesText, image, video, audio and PDF
Output typeText only
Thinking levelsLow, medium and high
CachingImplicit only, explicit caching is not supported
Consumption optionsBatch, flex and priority inference
Not supportedImage generation, audio generation, Live API

Two of those rows drive real money.

The 1 million token context window in Gemini 4 Argon is not an invitation to use it. Sending a million input tokens costs 2 dollars every time you do it, and a conversation that resends a large context on each turn multiplies that figure quickly. The value of the window is that you can stop chunking documents and losing the thread, not that you should paste everything into every request.

The expanded output limit is the more interesting capability, because long decode lets the model produce hundreds of thousands of tokens in a single trajectory. That is what makes long-running agent work feasible, and it is also what makes a generous output budget expensive at 10 dollars per million. The default synchronous cap of 65,536 tokens is where most applications should live.

What a workload actually costs

Three worked examples of Gemini 4 Argon pricing, using list prices and ignoring discounts, because the arithmetic matters more than the exact result.

A customer support assistant on Gemini 4 Argon. A fixed system prompt of 4,000 tokens, 600 tokens of conversation history per turn, and 300 tokens of answer. Over 10,000 turns that is roughly 100 million input tokens and 3 million output tokens, or about 200 dollars plus 30 dollars. If the system prompt is cached and served at the discounted rate, the input side of that bill falls sharply, which is why caching design is now a core part of product engineering rather than an optimisation.

A document analysis pipeline on Gemini 4 Argon. Two hundred contracts, 30,000 tokens each, analysed once with a 1,500 token summary returned. That is 6 million input tokens and 300,000 output tokens, which at list price comes to about 12 dollars plus 3 dollars. This is the workload class where a frontier model at these prices became viable this month; at the output prices of a year ago it was not.

A weekly reporting job on Gemini 4 Argon. Forty client accounts, each with a 15,000 token data extract and a 2,000 token written summary. That is 600,000 input tokens and 80,000 output tokens, or about 2 dollars per weekly run. This is the class of work where the question is not which model to use but whether a person spends two hours writing the same summaries. At these prices the build argument is straightforward, and the limiting factor becomes data preparation rather than model cost. The output side dominates, which is the argument for choosing a cheaper tier for the routine steps and reserving the flagship for the parts that need it.

None of those figures include the discounts Google may apply to batch work, and none of them are audited bills. They are the shape of the arithmetic, and the shape is what changes decisions.

The other half of the market is moving too

Flagship pricing only tells half the story, so here is the other half. On 7 October, Anthropic released Claude Haiku 5.5 with a pricing table that lands in a different universe: 10 cents per million input tokens and 50 cents per million output tokens for prompts up to 100,000 tokens, rising to 50 cents and 2 dollars 50 above that threshold. Cache reads sit at one cent per million.

Anthropic says the model averages around 75 percent cheaper to run than its previous small model, and that prompts under 100,000 tokens make up around 90 percent of requests to that model line. The same announcement halved the price of cache reads on Claude Sonnet 5.5, which Anthropic calculates reduces the cost of most agentic work on that model by about 20 percent.

Put the two side by side and the tiering lesson is hard to miss. The support example above spends about 230 dollars at list prices, most of it on input tokens that Gemini 4 Argon discounts heavily when they are cached. Routing the routine steps such as summarisation, compaction and classification to a small model can shed a large share of that without touching the parts of the work that need the expensive tier.

Gemini 4 Argon and Claude Haiku 5.5 are not substitutes. Gemini 4 Argon is a reasoning flagship with a million token window and a long decode mode built for sustained agent work. Haiku 5.5 is a fast small model that Anthropic itself says is best used for narrowly scoped tasks such as compaction, summarisation and subagent work, and it points developers to Sonnet and Opus for complex agentic coding. The mistake is not choosing the wrong one. It is choosing one for both jobs.

Which platforms and models enterprises actually put into production is a separate question, and our survey write-up on agent orchestration answers it with data rather than pricing tables.

Why caching is the lever that matters most

Notice what both companies did in the same month. Google priced cached input at 95 percent off, and Anthropic cut cache reads on its mid-size model in half while pricing Haiku's cache reads at a tenth of its input rate. Two competitors, one conclusion: the cost problem in agentic systems is resending the same context.

That produces a short list of engineering habits worth adopting now.

  • Make the stable part of a prompt byte-identical. A system prompt with a timestamp or a random ordering inside it will not cache. Moving volatile content to the end of the request is the cheapest optimisation available.
  • Keep retrieved documents in a fixed order when the same set is reused across turns, rather than reshuffling by relevance score on every call.
  • Stop treating the context window as free. A million token window in Gemini 4 Argon at 2 dollars per million is a 2 dollar request every time you fill it.
  • Check how caching actually behaves on implicit-only models, because you cannot pin context deliberately and hit rates will vary with prompt shape.
  • Measure cost per completed task, not cost per token. A cheaper model that needs three attempts can cost more than the expensive one that needs one.

The reason both vendors moved here is structural. In an agent loop, most of what a model reads on turn twenty is what it read on turn nineteen. Without caching, you pay full input price to resend the same content dozens of times per task. With it, the cost of memory drops by an order of magnitude. Gemini 4 Argon puts that discount at 95 percent, which is about as clear a statement as a pricing table can make about what the vendor wants you to stop doing. The mechanics are the same ones behind inference caching in production, which we covered in detail.

How to decide which tier each task needs

Routing is the skill this market rewards, and the split is easier than it looks. It is also the pattern that production grade agentic systems are built on.

Send to Gemini 4 Argon: work where a wrong answer costs more than the tokens. Multi-step reasoning, code changes with consequences, document synthesis where nuance matters, and agent tasks that have to hold a plan for a long time. The million token window and the long decode mode exist for these cases, and they are what the premium buys.

Send to a small model: classification, extraction, formatting, summarisation, compaction and the subagent work inside a larger agent. These tasks have checkable answers, they run at volume, and a small model at a tenth of the price returns the same result. Anthropic's own guidance for its cheap model points exactly here.

Send to neither: work you can answer with a rule, a query or a template. The fastest way to shrink an AI bill is to stop sending requests that never needed a model.

Measure the boundary. The right tier for a task is not a matter of taste, it is whether the cheap option fails often enough to cost more than the expensive one. Instrument both and compare cost per completed task.

Five things to do this week

  1. Rebuild your cost model with these numbers. If it was built when frontier output tokens cost five to ten times more, the conclusion about what is affordable has changed.
  2. Route by tier. Flagship for reasoning and synthesis, small model for classification, extraction, summarisation and compaction.
  3. Design for cache hits before you write another prompt. It is the difference between a viable margin and an unviable one at scale.
  4. Move offline work to batch. Anything that does not need a human waiting for it belongs in the cheaper queue.
  5. Keep the abstraction layer. Prices moved twice in eight days in October. Code that assumes one provider's pricing structure is code you will rewrite.

Where the numbers stop

Four limits on everything above.

  • The 2 and 10 dollar figures are introductory. Google has not said how long they last, and API pricing in this market moves monthly, so every worked example here is a snapshot.
  • The benchmark scores behind the Gemini 4 Argon capability claims come from Google, including the AutomationBench and long video results. They are vendor run, and no independent cost or capability comparison exists yet.
  • Gemini 4 Argon's most distinctive capability, autonomous vulnerability discovery, is not generally available. Google released the model without cyber guardrails only to trusted defenders and internal teams, so the security claims are not something a normal API customer can test.
  • The two pricing tables describe models with different strengths and different limits. Comparing their lists is not comparing what they can do.

FAQ

Can any developer use Gemini 4 Argon now?

Gemini 4 Argon is listed as stable in Google's API documentation and Google's stated sequence begins with paid API customers, so the practical answer for most builders is yes through the paid API. Free tiers and consumer availability have not been confirmed by Google at the time of writing, and the documentation is the page to check before you plan around it.

Is the 2 dollar input price permanent?

No, it is described as introductory. Treat it as a launch window rather than a rate, and build the cost model so that a doubling of input price does not break the business case.

Should I move everything to Gemini 4 Argon now that prices have fallen?

No, and the pricing tables argue against it. Both Google and Anthropic price the cheap tier at roughly a twentieth of the flagship on input and similar ratios on output, and Anthropic's own guidance points summarisation and compaction work at the small model. The winning pattern in this market is routing, not replacement.

The short version

Gemini 4 Argon is now listed as stable on the Gemini API at an introductory 2 dollars per million input tokens and 10 dollars per million output tokens, with cached input discounted 95 percent. Anthropic spent the same month cutting prices on its small models and on cache reads. Taken together, the two moves say that the frontier is cheaper than it was, that the small tier is dramatically cheaper still, and that the biggest lever inside any agentic system is not the model you pick but how much repeated context you avoid resending.

Sources
Subscription Form