Gemini 4 Argon Pricing: Cheap Per Token, Not Per Task

Gemini 4 Argon became generally available on October 13, 2026 at an introductory $2 per million input tokens and $10 per million output tokens. That price runs through November 30. From December 1 it doubles to $4/$20. Either way Argon is far below GPT-6 Astra's $10/$50 per token: a fifth of Astra's list price now, two fifths from December.

Per task, most of that gap disappears. Argon writes a lot: about 62k output tokens per Artificial Analysis task, roughly 2.3 times as many as Astra. Artificial Analysis measured $1.99 per task at the introductory price and $3.98 at the December price, against $3.26 for Astra. From December 1, Argon costs about 1.2 times as much as Astra per task for the same score.

This post sets out the full price list, including AIHubMix's rates, what two independent testers measured per task, the billing rules that are easy to miss, and the levers that actually lower the bill.

The price list

Google's Gemini API pricing page lists Standard, Batch, Flex and Priority prices for Argon on the paid tier. There's no free tier. Prices are USD per million tokens, and output prices include thinking tokens.

What the table doesn't show:

• No long-context surcharge. One price covers any prompt length up to Argon's 1M-token window.
• Implicit caching only. Argon doesn't support explicit caching, so there's no cache storage fee on Google's side. Repeated prompt prefixes of at least 4,096 tokens are discounted 95% automatically.
• Grounding is extra. Google Search and Google Maps grounding each come with 5,000 free requests a month, shared across Gemini 3 and newer models, then $14 per 1,000.

On AIHubMix, Argon costs $2 input, $0.10 cached input and $10 output per million tokens, the same as Google's introductory rate, according to its model page. Video and audio input are $2 per million tokens, and web search is $0.014 per request. AIHubMix hasn't said whether its price changes on December 1. [TBD: AIHubMix price from December 1]

For comparison:

Gemini 3.1 Pro Preview prices are for prompts up to 200k tokens; above that they rise to $4 input and $18 output. GPT-6 Astra's price is as reported by Engadget.

Cost per task: what independent tests measured

Figures in this section come from Artificial Analysis (its launch analysis and model pages) and Vals AI. Both ran Argon at high, its highest thinking level, not at the default medium. Neither set of numbers comes from Google.

Artificial Analysis: before and after December 1

Artificial Analysis runs every model through the same Intelligence Index tasks and reports the average cost of one task. It doesn't say how many tasks the index has.

Argon's introductory token price is a fifth of Astra's, but its cost per task is three fifths of Astra's: an 80% discount per token shrinks to 40% per task. From December 1, a 60% discount per token turns into a cost per task about 1.2 times Astra's. Artificial Analysis says the saving "is driven by lower token prices, rather than reduced token use". Argon averaged about 62k output tokens per task, against about 27k for Astra. Running the whole index, it generated 110M output tokens; the median model generated 85M, and Gemini 3.1 Pro Preview 67M.

The two Claude models score higher and cost more. From December 1, Argon's $3.98 is about two thirds of Claude Opus 5.5's $5.98 and about three quarters of Claude Sonnet 5.5's $5.46. Against Gemini 3.1 Pro Preview, Argon scores 23 points higher and costs about 1.5 times as much per task now, about 3 times from December.

Vals AI: standard prices

Vals AI ran Argon at the standard $4/$20 price, with a 262,144-token output cap, and reports the average cost of one test.

On this suite Argon costs less per test than the three Claude models just below it, even at the December price. The cost climbs on Vals' long agentic benchmarks:

CUA-bench, a computer-use benchmark, is the extreme case: the most expensive of these tests and one of Argon's weakest results, 7th of the 8 models Vals ran.

What the measured numbers show

• Against GPT-6 Astra, the introductory price carries the saving. Until November 30 Argon matches Astra's score for 60% of the cost per task. From December 1 it costs about 1.2 times as much.
• Token count is the reason. Argon writes about 2.3 times as many output tokens per task as Astra. At the introductory price that eats half of its per-token discount; at the December price it more than cancels it.
• Against Claude, Argon is cheaper with both testers. On Artificial Analysis it costs $3.98 per task at the December price, against $5.46 for Sonnet 5.5 and $5.98 for Opus 5.5, though both score higher. On the Vals Index it costs $15.68 per test, against $21.34 to $32.14 for the next three Claude models.
• Long agentic runs are where the bill grows. On Vals' agentic benchmarks a single test costs between $30 and $194.
• These are high-thinking numbers. Argon defaults to medium. Google says low and medium use fewer tokens, but no tester has published cost or accuracy at those levels.
• Compare within one tester, not across them. Artificial Analysis quotes the introductory price with the standard figure alongside; Vals uses the standard price. Their tasks and harnesses differ too.

Billing rules that are easy to miss

The introductory price ends November 30. From December 1 every Google rate doubles. A monthly budget set at $2/$10 lasts about half as long after that.

Thinking tokens are billed as output. Google's price list says so. The usage field reports them separately as thought tokens, but they cost the output rate.

A cap that's too low still costs money. max_output_tokens limits thinking and output together. If a response hits it, the request ends incomplete, with truncated or even empty output, and the thinking tokens are still billed. Google's advice is to lower the thinking level to save tokens rather than set a small cap, and to raise caps below 32K to at least 32,768.

Long decode re-reads its own output. Responses longer than 64K tokens use long decode, which generates in 64K "hops". You pay for every token in every hop, including hops that contain only thinking, and each hop sends the prompt plus everything generated so far back to the model as input. Google's example: a 1,000-token prompt that produces 200,000 tokens takes four hops and reports about 397,000 input tokens, although the context is only about 201,000. Implicit caching discounts that repeated input by 95%. A cancelled or failed run still bills the work done before it stopped.

Cache hits need a stable prefix. Implicit caching applies automatically to prompts of at least 4,096 tokens. The AIHubMix Gemini guide notes that a hit needs the same content, model and parameters, with repeated context placed first. Reordering the prompt between turns turns cached input back into full-price input.

Search grounding is billed per query. After the free 5,000 a month, Google charges $14 per 1,000 queries, and one request can trigger several. AIHubMix lists web search at $0.014 per request.

Declined answers still cost money. Argon declines more often than other frontier models when it isn't sure. A declined answer still bills its input and thinking tokens, and retrying the same question on Argon pays again.

How to lower the bill

1. Budget at the December price. The introductory rate is a bonus for the next few weeks, not a plan.
2. Start at medium, then try low. Google recommends testing all three thinking levels and picking the lowest that meets your quality bar. Every cost figure above was measured at high.
3. Send work that can wait to Batch or Flex. Both cost half the standard price.
4. Keep the prompt prefix stable. System prompt, tool definitions and documents go first and stay identical; the user's new message goes last.
5. Use long decode only when one answer needs more than 64K tokens. Google says chat, Q&A, summarization, classification and extraction finish well under 32K.
6. Route by difficulty. Send extraction, classification and formatting to Gemini 3.8 Flash and keep Argon for the hard reasoning steps.
7. Send declines somewhere cheaper. When Argon says it isn't sure, route the question to a person or another model rather than retrying Argon.

Estimate your own cost

A small calculator for token counts from your own logs. Output covers thinking tokens too, so pass the full completion count.

PRICES = {  # USD per 1M tokens; output includes thinking tokens
    "intro":    {"input": 2.00, "cached": 0.10, "output": 10.00},  # Google, through Nov 30, 2026
    "standard": {"input": 4.00, "cached": 0.20, "output": 20.00},  # Google, from Dec 1, 2026
    "batch":    {"input": 2.00, "cached": 0.10, "output": 10.00},  # Google Batch or Flex, from Dec 1
    "aihubmix": {"input": 2.00, "cached": 0.10, "output": 10.00},  # AIHubMix model page, now
}


def task_cost(uncached_in, cached_in, output, tier="standard"):
    p = PRICES[tier]
    return (uncached_in * p["input"] + cached_in * p["cached"] + output * p["output"]) / 1_000_000


# Fill in the counts from the usage field of your own responses:
# task_cost(uncached_in, cached_in, output, "standard")

The most direct way to get real counts is a few runs of your own workflow on Argon through AIHubMix, at the thinking level you plan to use. If you size things on Gemini 3.8 Flash or 3.1 Pro Preview instead, expect Argon's counts to be higher: on Artificial Analysis' index it generated 110M output tokens where Gemini 3.1 Pro Preview generated 67M. For long decode runs, the reported input total includes every hop's re-read, so use it for billing, not to measure context size.

On AIHubMix

Argon is live on AIHubMix under the model ID gemini-4-argon, served through Vertex AI and Google AI Studio. You can call it on the OpenAI-compatible endpoint (https://aihubmix.com/v1) or with Google's own SDK. It bills $2 per million input tokens, $10 per million output tokens and $0.10 per million cached input tokens.

For the cheap steps in the same workflow, Gemini 3.8 Flash is still the lower-cost option.

FAQ

How much does Gemini 4 Argon cost?
Through November 30, 2026, $2 per million input tokens, $0.10 per million cached input tokens and $10 per million output tokens. From December 1, $4 input, $0.20 cached input and $20 output. Output prices include thinking tokens.

When does the Gemini 4 Argon introductory price end?
On November 30, 2026. The standard price applies from December 1.

How much does Gemini 4 Argon cost on AIHubMix?
$2 per million input tokens, $10 per million output tokens and $0.10 per million cached input tokens, the same as Google's introductory rate. Web search is $0.014 per request. AIHubMix hasn't said whether its price changes on December 1.

Is Gemini 4 Argon cheaper than GPT-6 Astra?
Per token, yes, by a wide margin. Per task, Artificial Analysis measured about 60% of Astra's cost at the introductory price and about 1.2 times Astra's cost at the December price, because Argon writes about 2.3 times as many output tokens.

Does Gemini 4 Argon charge more for long prompts?
No. There's no long-context surcharge. Explicit caching isn't supported, but implicit caching discounts repeated prefixes of at least 4,096 tokens by 95%.

How is long decode billed?
You pay for every token in every 64K hop, and each hop re-reads the prompt and everything generated so far as input. Implicit caching discounts that repeated input by 95%.

What does one task cost on Argon in independent tests?
On Artificial Analysis' Intelligence Index, $1.99 per task at the introductory price and $3.98 at the December price, against $5.46 for Claude Sonnet 5.5 and $5.98 for Claude Opus 5.5. On the Vals Index, run at the standard price, $15.68 per test, rising to $193.78 per test on the CUA-bench computer-use benchmark.

Keep reading: the Gemini 4 Argon series

• Before budgeting, check whether Argon is the right model for the job at all: Gemini 4 Argon: Frontier Scores, a 1M-Token Output Limit, and How to Call It Now
• Thinking level and long decode decide most of the bill. How to set both: Gemini 4 Argon Thinking Levels and Long Decode: Defaults, Costs and When to Use Each
• Output caps, cache-friendly prompts and a route for declined answers belong in the switch itself: Moving to Gemini 4 Argon from Gemini 3.1 Pro or 3.8 Flash: What Breaks and How to Fix It

Sources

• Gemini Developer API pricing (Google AI for Developers)
• Long decode (Google AI for Developers)
• Gemini 4 Argon: Google is back as one of the top three labs in intelligence achieved (Artificial Analysis)
• Gemini 4 Argon benchmarks, cost and capabilities (Vals AI)
• Gemini 4 Argon on AIHubMix (AIHubMix)