Gemini 4 Argon’s API price is two date columns. The model code is gemini-4-argon. On standard, per million tokens: through 30 November 2026, input is $2.00 and output is $10.00; from 1 December, input is $4.00 and output is $20.00. The output price includes thinking tokens. Cached input is 5% of the matching input price, $0.10 in November and $0.20 in December, and the only cache on offer is implicit.
Today is 14 October 2026. The figures come from the Gemini API pricing page opened that day, and the specs come from the model page. On that row the free tier and the Google AI Pro plan are marked not available. They are not a $0 price. Check the pricing page again before you go live. This is a worksheet, not a quote. If you buy through OpenRouter or another relay, add that route’s markup on the model line. See the 10 October OpenRouter comparison.
What input and output cost
Every unit price is US dollars per million tokens. The pricing page does not add a long-context surcharge for Argon, and it does not price audio by the minute. Text, images, video, audio, and PDFs all take the same input price. Output is text. Thinking tokens have no price of their own. They are already inside the output price.
| Per million tokens | Through 2026-11-30 | From 2026-12-01 |
|---|---|---|
| Standard · input | $2.00 | $4.00 |
| Standard · output (thinking included) | $10.00 | $20.00 |
| Standard · cached input | $0.10 | $0.20 |
| Batch / Flex · input | $1.00 | $2.00 |
| Batch / Flex · output (thinking included) | $5.00 | $10.00 |
| Batch / Flex · cached input | $0.05 | $0.10 |
| Priority · input | $3.60 | $7.20 |
| Priority · output (thinking included) | $18.00 | $36.00 |
| Priority · cached input | $0.18 | $0.36 |
Standard
Interactive calls use this column. Through 30 November, $2 of input sits against $10 of output, so the output price is 5× the input price. Take a request with 20,000 input tokens, 8,000 thinking tokens, and 2,000 tokens of visible answer. Billed output is 10,000 tokens: $0.04 of input and $0.10 of output, $0.14 together. The visible answer is only 2,000 tokens. Thinking already accounts for four-fifths of that $0.10.
Batch and Flex
The two lanes share one price, half of standard. In November that is $1.00 input, $5.00 output, and $0.05 cache. From 1 December it is $2.00 input, $10.00 output, and $0.10 cache. Evals that can wait in a queue, offline labeling, and overnight batches belong on this lane. Latency for the lane itself follows the docs. Half price here is a price fact. It is not a claim that the latency matches.
Priority
Priority is 1.8× standard, exactly. In November that is $3.60 input, $18.00 output, and $0.18 cache. From 1 December it is $7.20, $36.00, and $0.36. Send the $0.14 request above on Priority and it is $0.072 of input plus $0.18 of output, $0.252 together. Leave the default lane on Priority only when the wait itself costs more than that gap.
Standard, Batch, Flex, and Priority are four bills. A request lands in one of them. Average $2 and $1 into $1.50 and the monthly total will not match the invoice.
Free tier
Argon has no free tokens you can drop into a formula. The pricing page says the free tier is not available, and the Google AI Pro plan is not available either. Content on the paid tier is not used to improve the product. Whether you can select the model in an app or in AI Mode does not change this API row. Calling gemini-4-argon means the account has billing turned on.
| Free tier | Paid tier | |
|---|---|---|
| Argon input / output | Not available | Table above |
| Google AI Pro plan | The pricing page says Argon is not included | |
| Used to improve the product | Argon not available | No |
| Google Search grounding | Not available | First 5000 calls a month free, then $14 / thousand |
| Google Maps grounding | Not available | Same rule: first 5000 calls a month free, then $14 / thousand |
Those 5000 calls are shared by Gemini 3 and newer models. They are not a bucket reserved for Argon. One user request can issue several search queries, and the count is queries, not requests. Maps is a separate row with the same arithmetic. Free tokens on Flash and other models can still exist. Book those free calls under Argon’s name and the first row of this table becomes $0.
Thinking tokens and cache
The model page gives one window of 1048576 tokens, input and output together. The default synchronous output cap is 65536. Long decoding can raise output to at most 1048576, and input then uses only what remains in the window. The default thinking level is medium, with low and high beside it. Whether thinking can be switched off follows the model page. The price list has no row of its own for it.
Fill the default synchronous cap and November charges 65536 × $10 / 1,000,000 ≈ $0.66 for that one call. December is about $1.31. The visible answer can be very short. Once thinking has filled the cap, the bill is still the full output. Filling long decoding, about 1048576 output tokens, is about $10.49 in November and about $20.97 in December, and that is before input. Keep maximum output off the everyday cap.
Implicit cache at 5%
Cached input equals the standard input price then in force × 0.05. That is $0.10 in November and $0.20 in December. Batch / Flex is $0.05 and $0.10. Priority is $0.18 and $0.36. The pricing page lists implicit cache only. Explicit cache is not supported. The Argon card also has none of the hourly storage fee that shows up on the Flash price list. Hits bill at the cache price. Misses stay on the input price. Cache replaces the input row and leaves the output row alone. Thinking is still billed in full at the output price.
This sheet’s base is 24,000 thinking tokens plus 4,000 of visible output per session. Halve the visible output and billed output moves from 28,000 to 26,000. The output line falls by about 7%. To bring that line down, lower the thinking level before you delete the last paragraph the user can see.
Monthly call cost
The base is an assumption, so a unit price can become a monthly fee. It is not an official Google usage figure. One month is 400 sessions. Each session sends 80,000 input tokens, 24,000 thinking tokens, and 4,000 tokens of visible output. Billed output is 28,000. Across the month that is 32 million (32M) input tokens and 11.2 million (11.2M) billed output tokens. Those 80,000 input tokens already include the growth from carrying earlier turns back in. Multiply only the user’s first message and the sheet comes out short.
The model line is millions of input × the input price, plus millions of output × the output price. Search and the machine are added afterward. They stay out of this line.
One volume, four bills
| Standard · November price | Standard · December price | Batch / Flex · November price | Priority · November price | |
|---|---|---|---|---|
| Input 32M | $64.00 | $128.00 | $32.00 | $115.20 |
| Output 11.2M | $112.00 | $224.00 | $56.00 | $201.60 |
| Total | $176.00 | $352.00 | $88.00 | $316.80 |
One session on the November standard price is 80,000 × $2 / million = $0.16, plus 28,000 × $10 / million = $0.28, $0.44 together. 400 × $0.44 = $176. The $112 output line is 1.75× the $64 input line, because thinking stretched the output. The visible answer is not 11,000 tokens.
If 70% of input hits implicit cache and 30% stays on the input price, with output unchanged: uncached 9.6M × $2 = $19.20, cached 22.4M × $0.10 = $2.24, output still $112. November standard totals $133.44. The same hit rate in December is 9.6M × $4 = $38.40, 22.4M × $0.20 = $4.48, output $224, total $266.88. A prefix that has not hit yet sits outside this discount. When the system prompt and the repo map repeat, read the hit rate before you drop the lane to Batch.
December Batch / Flex equals November standard: $64 of input and $112 of output, still $176. December Priority is 32M × $7.20 = $230.40 and 11.2M × $36 = $403.20, total $633.60.
Search on its own line
This sheet’s base leaves search out, so the model line is the table above. Two Google searches per session is 800 for the month, still inside the 5000 free calls, so this line is $0 — as long as Gemini 3 and newer models have not already used those 5000 searches this month. Twenty searches a session is 8,000 for the month, 3,000 past the cap, and 3 × $14 = $42. That $42 sits outside the model line. Maps uses the same free allowance and the same overage price, on a row of its own.
The machine on its own line
- A person at the computer sends the request. The machine line is $0. Close the lid and the process stops, so overnight jobs have nowhere to land. The month stays on the model line, and November standard is $176.
- You need HTTPS around the clock, and every tool is an API. The process that sends requests is light. If a machine is already on, skip renting a Mac just to send HTTP.
- The tool has to be macOS. Examples are
xcodebuild, a simulator, or signing. Add a Cloud Mac. On the day this was written, the homepage starting price was M4 · 16GB / 256GB from $101.7 per month. Checkout follows the plan page. A starting price is not a contract price.
November standard plus that 16GB machine: $176 + $101.7 = $277.70. December standard plus the same machine: $352 + $101.7 = $453.70. The Mac can stay awake while Argon output is still billed at $10 or $20. Monthly rent does not swallow the API. Bill one machine per command. If the command runs in some other sandbox, bill that sandbox. If the command moves onto the Mac, write the sandbox line as 0. For how to split the model, the tools, and the machine, see how much an AI agent costs per month.
After 1 December
The same 32M / 11.2M takes standard from $176 to $352, exactly double. Batch, Flex, Priority, and the cache prices double with it. The doubling covers the whole price card. If the work can leave the interactive path, December Batch / Flex ($176) lines up with November standard. It does not line up with November’s half-price total.
- A stable prefix: let implicit cache hit before you talk about dropping a lane. A 70% hit takes December standard from $352 to $266.88. The saving is on the input line.
- The output line is still the larger one: December standard output is $224 and input is $128. Lowering the thinking level does more than deleting the visible answer.
- The free tier does not show up on 1 December. What doubles is the paid-tier unit price. Argon stays off the free tier.
Swap in your own three numbers
Change only three numbers: how many sessions, input tokens per session, and billed output tokens per session. Billed output is thinking plus the visible answer. The model line is (millions of input × input price) + (millions of output × output price). If 400 sessions become 40, divide the model line by 10. Cloud Mac rent stays put. The search allowance is a bucket of 5000 calls for the month. It does not grow with the session count.
- Drop thinking and keep only 4,000 tokens of visible output: 11.2M becomes 1.6M. The November standard output line goes from $112 to $16, and the model line becomes $64 + $16 = $80. That is an assumption. It is not a measured result of setting the thinking level to 0.
- Raise input to 200,000 tokens per session and keep the session count: input is 80M. November standard input is $160, output stays $112, and the total is $272. That is the point where cache is worth reading first.
- Buy through a relay: multiply the model line by that route’s factor. Whether tools and search are extra follows that route’s invoice.
FAQ
Does Gemini 4 Argon have a free tier?
No. The pricing page marks both the free tier and the Google AI Pro plan as not available. Calling gemini-4-argon requires the paid tier. Whether you can hold a conversation with it inside an app does not change that row.
How are thinking tokens priced?
They are billed at the output price, and they count against the output cap. This sheet's base is 24,000 thinking tokens plus 4,000 of visible output per session, billed as 28,000 output tokens. Shortening only the visible answer barely moves the bill.
How much can cache save?
Argon has implicit cache only, with no explicit cache and no hourly storage fee on the price list. Cached input is 5% of the standard input price. If 70% of input hits on this sheet, November standard drops from $176 to $133.44.
How does the price change after 1 December?
On the same token volume, the input, output, and cache unit prices for standard, Batch, Flex, and Priority all become twice the November price. This sheet's standard line goes from $176 to $352.
Are Batch and Flex cheaper?
On the price list they match, and both are half of standard. November is $1 of input and $5 of output. From 1 December it is $2 and $10. This sheet's base drops from $176 to $88. Priority is 1.8× standard, and November comes to $316.80.
Match the date and the lane, then multiply by your tokens
Through 30 November, multiply the standard lane by $2 and $10. From 1 December, multiply by $4 and $20. Thinking counts as output. The free tier has no zero you can enter. Once search passes the shared 5000 calls, add $14 per thousand. Add the Mac only for the step that cannot run once it leaves macOS.
Put overnight calls on a machine that stays awake
The laptop is where you match one Argon request’s tokens against this sheet. The overnight loop, and the step that has to run on macOS, go on a Cloud Mac that does not sleep. Keep the token price and the monthly rent on two lines.
Further reading
- How much does an AI agent cost per month? API tokens, tool calls, and the cloud server
- OpenRouter rankings, October 2026: which of GPT, DeepSeek, and Gemini is growing fastest, and how far apart are the API prices?
- OpenRouter pricing guide 2026: GPT, Claude, Gemini, and DeepSeek API costs and a monthly budget
- Claude Opus 5.5 vs Claude Code: how to choose a 2026 coding model — price, tokens, coding ability, and agent workflow