Gemini 4 Argon Pricing: API Rates, Plans and the December 1 Increase

Gemini 4 Argon pricing on the paid API is $2.00 per million input tokens and $10.00 per million output tokens through November 30, 2026. On December 1, 2026 both rates double to $4.00 and $20.00, and every other tier doubles with them. In the Gemini app, Argon comes with Google AI Pro at $19.99 a month and Google AI Ultra from $99.99. All prices here are as of October 16, 2026.

Rates are only half of the budget question. Argon writes long answers: Artificial Analysis measured about 62,000 output tokens per task against 27,000 for GPT-6 Astra, as reported by The Next Web. A low price per token does not guarantee a low price per task.

This post covers the full rate table, app plan prices, a list-price comparison with four rivals, and a cost formula with four worked scenarios you can rerun. For what the model is and does, start with our Gemini 4 Argon overview.

How much does the Gemini 4 Argon API cost per million tokens?

Gemini 4 Argon costs $2.00 per million input tokens and $10.00 per million output tokens on the Standard tier through November 30, 2026, then $4.00 and $20.00 from December 1. Output includes thinking tokens. Cached input is $0.10, rising to $0.20. Batch and Flex cost half of Standard, and Priority costs 1.8 times Standard.

Google’s Gemini API pricing page lists these paid-tier rates in US dollars per million tokens. The table below copies them as of October 16, 2026.

Four details change how you read that table:

• Thinking tokens bill as output, so a reply that thinks for 20,000 tokens and answers in 2,000 is billed for 22,000 output tokens.
• There is no free API tier. Google’s page lists Argon as unavailable on the Free tier, and on the Google AI Pro plan for API use.
• Google lists no prompt-size surcharge and no cache storage fee for Argon. Caching is implicit only, and explicit caching is not supported.
• Google Search grounding includes 5,000 free requests per month, shared across Gemini 3 and newer models, then costs $14 per 1,000 requests.

The tiers differ in behavior as well as price. According to Google’s Flex inference documentation, Flex targets 1 to 15 minutes of latency and can be preempted when capacity is tight, so you need your own retry logic. Priority is built for second-level responses and, per Google’s Priority documentation, overflow requests drop to Standard processing at the Standard rate. Google’s Batch documentation sets a 24 hour target turnaround.

When does the Gemini 4 Argon price increase happen?

The Gemini 4 Argon price increase takes effect on December 1, 2026. Google’s pricing page states the introductory rates apply “through November 30, 2026.” Every paid tier doubles, cached input included, so a job that costs $100 on November 30 costs $200 on December 1 at the same token counts.

For a rough anchor, one million input tokens plus one million output tokens costs $12.00 on Standard through November 30 and $24.00 from December 1. Early coverage said Google had given no end date, and some posts still repeat that. The pricing page now states one, so budget from the page rather than from older articles.

How do you estimate your own Gemini 4 Argon cost?

Estimate Argon cost with one formula: uncached input tokens times the input rate, plus cached input tokens times the cached rate, plus output and thinking tokens times the output rate, divided by one million. Run it twice, once at introductory rates and once at standard rates, and budget from the second result.

cost = (uncached_input x input_rate
      + cached_input x cached_rate
      + (output + thinking) x output_rate) / 1,000,000

cache savings = cached_input x (input_rate - cached_rate)

Five steps turn the formula into a number you can defend.

Step 1: Count input per call, including resent history.

Chat and agent loops resend earlier context on every turn, so total input is usually far larger than the prompt you wrote.

Step 2: Decide your cache hit share.

Google’s caching documentation lists Argon with a 4,096-token minimum and says savings pass through only when a request actually hits the cache. Treat a high hit rate as a best case until you measure it.

Step 3: Count output and thinking together.

Both bill at the output rate, which is five times the input rate at both the introductory and standard prices.

Step 4: Pick a tier.

Use the Standard, Batch, Flex or Priority row from the table above.

Step 5: Multiply by volume and run it at both rates.

The December figure is the one to plan around.

What would four typical Gemini 4 Argon workloads cost?

These four scenarios are illustrations built from stated assumptions, not measurements from a live account. Each is priced at the introductory rate and the standard rate. The December bill is exactly double in every case, but caching, tier and output length move the starting point far more than the date does.

The assumptions, in order:

• A, document Q&A: one 200,000-token document and 20 questions on Standard. Each call adds 300 question tokens and returns 2,500 output and thinking tokens. The first call misses the cache and later calls hit it on the document.
• B, agentic coding task: 62,000 output tokens, the per-task average reported for Argon, plus an assumed 1,200,000 input tokens across turns with 90% cache hits, on Standard.
• C, batch job: 5,000 documents at 8,000 input and 1,500 output tokens each, no cache hits, priced on Batch.
• D, long-decode run: Google’s own example of a 1,000-token prompt producing 200,000 output and thinking tokens over four hops, with 397,216 input tokens in total.

Scenario A is the clearest case for caching. The same 20 questions cost $8.51 without cache hits and $1.29 with them at introductory rates, because the 200,000-token document is read 20 times.

Scenario B shows the part you cannot discount. The 62,000 output tokens alone cost $0.62 at introductory rates and $1.24 at standard rates, before any input. These figures are not Artificial Analysis’s own per-task costs of $1.99 and $3.98, which come from its test suite.

Scenario C shows tier choice. Batch at the December rate costs $155.00, the same as Standard through November 30.

How does Argon’s list price compare with GPT-6 Astra, GPT-6.1 Sol, Claude Sonnet 5.5 and Claude Opus 5.5?

At the introductory rate, Argon’s $2 input and $10 output matches GPT-6.1 Sol and Claude Sonnet 5.5, and costs one fifth of GPT-6 Astra’s $10 and $50. From December 1, Argon’s $4 and $20 matches Claude Opus 5.5 and sits 60% below Astra’s short-context list price. Rival prices are as of October 16, 2026.

The OpenAI figures come from OpenAI’s API pricing page and the Anthropic figures from Anthropic’s pricing documentation, both read on October 16, 2026. For plan-level pricing on those vendors, see our guides to Claude pricing and ChatGPT pricing.

Two caveats apply. OpenAI lists higher long-context rates for both models but its page does not say where long context starts, so confirm that before pricing a very large prompt. Anthropic also says its newer tokenizer produces roughly 30% more tokens for the same text on Claude 4.7 and later models, so a token is not the same unit across vendors. Count tokens on your own data with each vendor’s counter.

Why is cheap per token not cheap per task?

Because Argon is verbose. Artificial Analysis measured about 62,000 output tokens per task for Argon against 27,000 for GPT-6 Astra, as reported by The Next Web. Output is the expensive side, so extra tokens can erase a lower rate. Artificial Analysis put Argon at $1.99 per task at introductory prices and $3.98 at standard, against $3.26 for Astra.

An output-only test shows how much room Argon has. Argon is cheaper than a rival on output alone as long as it writes fewer tokens than the multiple below. The multiple is the rival’s output rate divided by Argon’s.

Read the table against Argon’s measured ratio of about 2.3 times Astra’s output (62,000 divided by 27,000). That sits under Astra’s 5.0x break-even at introductory rates, so output alone costs $0.62 for Argon against $1.35 for Astra. At standard rates the break-even is 2.5x, so Argon’s $1.24 is only about 8% below Astra’s $1.35. The margin nearly disappears on December 1. Artificial Analysis ran Argon at high thinking and Astra at max, so your own settings will shift these numbers.

Against Sol and Sonnet 5.5, the introductory prices are identical, so every extra token costs more. From December, Argon would need to write less than half as many tokens to win on output. Against Opus 5.5 the rates match, so token count alone decides. We have no verified per-task token counts for those three models, so measure them on your own work.

The test ignores input and caching. Argon’s cached input at $0.10 to $0.20 sits well below Astra’s $1.00, which favors Argon on input-heavy jobs. Whether the extra tokens buy better answers is a separate question, covered in our Gemini 4 Argon benchmarks post and the Argon vs Astra vs Opus comparison.

What do Google AI Plus, Pro and Ultra cost, and which include Argon?

In the Gemini app, Google AI Pro costs $19.99 a month and Google AI Ultra starts at $99.99, and both include Gemini 4 Argon. Google AI Plus costs $4.99 a month. Its column carries an Argon mention, but Google’s launch banner points to Pro and Ultra, so confirm in your account. App plans are separate from API billing.

These prices are from Google’s Gemini subscriptions page as of October 16, 2026. App usage limits depend on prompt complexity, features used and chat length, so a plan price is not a guarantee of unlimited use.

A subscription does not give you API tokens. Developers pay per token on the paid API tier from the table above. Our guide to accessing Gemini 4 Argon covers each route, and our Gemini pricing guide covers the wider plan lineup.

Why can a long-decode run cost more than its output suggests?

Long decode generates in hops of about 64K tokens, and each hop resends the prompt plus everything written so far. Google’s example, a 1,000-token prompt producing 200,000 output and thinking tokens, counts 397,216 input tokens across four hops. You pay for every token in every hop, including hops that contain only thinking.

Google’s long decode documentation shows the hop inputs adding up as 1,000 + 66,536 + 132,072 + 197,608. It says implicit caching discounts repeated input in each hop by 95%, but it does not say exactly how many of those tokens count as cached. That is why scenario D is a range: $2.04 to $2.79 at introductory rates, of which $2.00 is output.

The output share dominates, so the practical lever is not caching but whether you need a long run at all. Google says long decode is not meant for chat, Q&A, summarization, classification, extraction, fixed-length requests, agents making many short tool calls or latency-sensitive work. Lowering max_output_tokens cuts a response off rather than making the model think less, so control cost with the thinking level instead.

What should you do before December 1 to control the bill?

Before December 1, budget every Argon workload at the $4.00 and $20.00 rates, structure prompts so the cache can hit, and move any job that can wait to Batch or Flex. Then measure tokens per task in a pilot, because output length, not the rate card, drives most of the bill.

The checklist:

• Budget at the standard rate. Treat the introductory price as a discount that ends, not as your baseline.
• Structure prompts for caching. Put large, stable content first and keep requests with a similar prefix close together in time, as Google’s caching guidance advises.
• Move latency-tolerant work to Batch or Flex. Both list at half of Standard, and Flex needs client-side retries.
• Lower the thinking level before cutting the output cap. A tight cap truncates answers, while Google advises lowering the thinking level to control cost.
• Test concise-output instructions. Prompt wording affects response length, and our guide on how to prompt Gemini 4 Argon covers it.
• Log input, cached input, output and thinking tokens per task from the first day.

What should you do next?

Take 50 to 100 real tasks from your workload and run them on Argon at the thinking level you plan to use. Record input, cached input, output and thinking tokens for each. Put the averages into the formula above at both rates. That gives you a number you can defend, which no rate card can.

Then run the same exercise on the model you use today. If Argon’s output stays under the break-even multiple for that model, the December price still works for you. If it does not, the answer may be Batch, a lower thinking level or a different model.

Check Google’s pricing page again on December 1 and after any announcement, because every figure in this post is dated October 16, 2026.