The full Gemini 4 Argon price table
These are the figures on Google's Gemini API pricing page, per 1M tokens in USD[1]. The page labels the output column as including thinking tokens.
Three details matter more than the headline doubling.
First, there is no free tier for this model: the page states that Gemini 4 Argon is not available on the Free tier[1], and Google's launch post says building requires the paid tier of the Gemini API[3]. You cannot prototype for free and then watch the price rise.
Second, caching is implicit only. The pricing page says explicit caching is not supported, and no storage price is listed[1]. You get the cache-read discount when Google's system detects a repeated prefix, not when you decide to pay for one. Your hit rate is something to measure, not assume.
Third, the Contributor tier is an invite-only preview at Batch-level rates, and Google's page says it does use your content to improve products. The Standard, Batch, Flex and Priority tiers do not[1]. Check that before anyone on your team picks the cheapest row for production data.
What the model page says about size
The Gemini API model page lists a context window of 1,048,576 tokens, input and output combined. Maximum output is 65,536 tokens by default on synchronous requests, or up to 1,048,576 with long decode[4]. Thinking is supported at low, medium and high levels. Inputs are text, image, video, audio and PDF; output is text only. Batch, Flex and Priority inference are all supported[4]. Note that the pricing page lists no separate long-context surcharge for Argon, unlike OpenAI's GPT-6.1 Sol, which steps up above 272K input tokens[6].
How Argon compares with Opus 5.5, Sonnet 5.5, GPT-6.1 Sol and DeepSeek
All prices below are per 1M tokens on each vendor's official pricing page, checked on 18 October 2026. They match the dataset behind our AI model cost calculator for the Claude and OpenAI models. The last column prices a month of 50M input and 10M output tokens with no caching: input millions times input price, plus output millions times output price.
Read the table in two passes. Today Argon is priced identically to Sonnet 5.5 and GPT-6.1 Sol. On 1 December it moves to Opus 5.5's price card, to the cent, on all three columns.
Two caveats on the comparison. Anthropic bills cache writes separately (for example $5 per 1M on Opus 5.5 for a 5-minute write[5]), which this table ignores, while Argon lists no write charge because writes are implicit. And DeepSeek's cache-miss input price is the one shown; its cache-hit input is $0.003 off-peak. DeepSeek peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday, excluding Chinese public holidays[7].
Price per token says nothing about quality or task completion. Nothing here claims Argon is better or worse than any of these models; that depends on your workload, and we do not quote benchmark numbers.
Three workloads priced before and after 1 December
Every figure below uses the Standard tier unless stated. Divide token counts by one million, multiply by the per-1M price, and add the lines.
Workload 1: a customer chatbot
50M input tokens and 10M output tokens a month, no cache hits assumed.
- To 30 Nov: 50 x $2.00 = $100.00 input, plus 10 x $10.00 = $100.00 output. Total $200.00.
- From 1 Dec: 50 x $4.00 = $200.00 input, plus 10 x $20.00 = $200.00 output. Total $400.00.
That is a $200 monthly increase, or $2,400 a year, with no change in traffic. Now add thinking. Suppose the same chatbot's responses carry 15M thinking tokens a month on top of the 10M visible output tokens. Billable output becomes 25M.
- To 30 Nov: $100.00 input + 25 x $10.00 = $250.00 output. Total $350.00.
- From 1 Dec: $200.00 input + 25 x $20.00 = $500.00 output. Total $700.00.
Same visible answers, $350 more per month, because 60% of the output tokens were thinking. The pricing page says output includes thinking tokens[1], so this is how the meter works, not a loophole. The 15M figure is an illustration, not a measurement of Argon.
Workload 2: an agent with a 60% cache hit rate
An always-on agent re-sends its context at every step. Say 400M input tokens and 20M output tokens a month, with 60% of input served from cache: 240M cached, 160M uncached.
- To 30 Nov: 160 x $2.00 = $320.00 uncached input; 240 x $0.10 = $24.00 cached input; 20 x $10.00 = $200.00 output. Total $544.00.
- From 1 Dec: 160 x $4.00 = $640.00; 240 x $0.20 = $48.00; 20 x $20.00 = $400.00. Total $1,088.00.
- If implicit caching never hits (0%): 400 x $2.00 = $800.00 + $200.00 = $1,000.00 now, and 400 x $4.00 = $1,600.00 + $400.00 = $2,000.00 after.
The same workload elsewhere, cache writes ignored: Sonnet 5.5 and GPT-6.1 Sol both come to $544.00 (the same $2 / $10 / $0.10 card). Opus 5.5 comes to $1,088.00. DeepSeek deepseek-flash off-peak: 160 x $0.15 = $24.00, 240 x $0.003 = $0.72, 20 x $0.60 = $12.00, total $36.72; at peak rates, $73.44. Cache behavior is the biggest variable here: the gap between a 60% and a 0% hit rate is larger than the December increase on the cached case. Our agent run cost simulator lets you vary agents, steps and cached share and see the monthly bill move.
Workload 3: an offline summarization batch
300M input tokens and 30M output tokens a month, no deadline pressure.
- Standard, to 30 Nov: 300 x $2.00 = $600.00 + 30 x $10.00 = $300.00. Total $900.00.
- Batch, to 30 Nov: 300 x $1.00 = $300.00 + 30 x $5.00 = $150.00. Total $450.00.
- Standard, from 1 Dec: 300 x $4.00 = $1,200.00 + 30 x $20.00 = $600.00. Total $1,800.00.
- Batch, from 1 Dec: 300 x $2.00 = $600.00 + 30 x $10.00 = $300.00. Total $900.00.
Batch halves the bill at either date, but it does not dodge the doubling. The cheapest cell is Batch before 1 December at $450; the dearest is Standard after, at $1,800. A fourfold spread for the same tokens.
What to do before 1 December
The pricing page gives dates but offers no way to lock a rate. Nothing in the docs we checked lets you prepay or reserve today's price, so the levers are your own workload choices.
1. Move offline work to Batch or Flex now
Anything that can wait for results (nightly summaries, classification of an archive, eval runs, embeddings-style backfills) belongs on Batch or Flex. Both list half of Standard's price on every line[1]. One-off backlogs are the real opportunity: if you have a pile of documents to process once, run it before 30 November at $1 / $5 instead of $2 / $10 later. Google's page does not say whether a batch job is billed at the rate on the day you submit it or the day it completes, so submit early enough that the question never matters.
2. Measure your thinking-token share
Pull a week of usage from your own API responses and split output into visible tokens and thinking tokens. If thinking is more than half of your output, it is your largest line item. The model page lists low, medium and high thinking levels[4]. Run your eval set at each level and record cost per completed task, not per token. Many classification and extraction tasks do not need the high setting.
3. Measure your real cache hit rate
Because caching is implicit, nothing guarantees a hit. Keep stable content (system prompt, tool definitions, reference documents) at the start of each request and variable content at the end, then compare billed cached tokens with total input. Our AI prompt cost calculator prices a prompt against several models if you want a quick per-request figure.
4. Set a budget and alerts in Google Cloud Billing
In the Cloud console, open Budgets & alerts under Billing and create a budget. Pick a scope (projects, services, labels), set an amount, and add alert threshold rules. The defaults are 50%, 90% and 100%, and each can track actual or forecasted spend[8]. Email goes to billing admins; Pub/Sub notifications can feed your own automation. One warning from Google's docs: an alerts-only budget does not cap usage or spending[8]. A separate spend cap budget type exists, but only for supported services, and we could not confirm that the Gemini API is among them. Set the alert first, then edit the thresholds so the 1 December step does not trip your 90% line on day one. For the agent-side version of this problem, see our post on hard spend caps and runaway coding-agent bills.
5. Decide whether Argon earns twice the price
After 1 December, Argon costs double Sonnet 5.5 and GPT-6.1 Sol on every column. That means Argon only wins on cost if it uses half the tokens or fewer to finish the same task. Fewer retries, fewer tool-call loops and shorter thinking all count. Test it: take 50 real tasks, run each model, and record total billed tokens and pass rate. If Argon needs 40% fewer tokens, it is still 20% more expensive overall (0.6 x 2 = 1.2). If it needs 60% fewer, it is 20% cheaper (0.4 x 2 = 0.8).
If your tasks are simple, a cheaper model is the right answer regardless. The lower-cost Gemini lines are covered in our Gemini 3.8 Flash migration guide.
Other Gemini API changes in the same window
The Gemini API changelog lists several dates that can affect a bill or a deployment[2]:
- 13 October 2026:
gemini-4-argon generally available.
- 8 October 2026:
gemini-3.7-flash deprecated and replaced by gemini-3.8-flash; requests are routed automatically.
- 8 October 2026:
gemini-3.5-flash deprecated and replaced by gemini-3.6-flash, also routed automatically.
- 23 October 2026:
deep-research-pro-preview-12-2025 shuts down.
Automatic routing is convenient and a quiet cost risk: your code still names the old model while the invoice reflects the new one. Check which model your logs say actually served each request, and pin an explicit model name where the price matters.
Frequently asked questions
When exactly does the Gemini 4 Argon price go up?
Google's pricing page lists the lower rates through 30 November 2026 and the doubled rates from 1 December 2026[1]. The page does not state a time zone for the cutover, so treat the boundary day as uncertain and do not schedule large jobs for the last hours of 30 November.
Does the price increase apply to Batch, Flex and Priority too?
Yes. Batch and Flex go from $1 / $5 to $2 / $10 per 1M input / output tokens, and Priority from $3.60 / $18 to $7.20 / $36[1]. Cache-read prices double on every tier as well.
Are thinking tokens billed separately?
No separate line. The output price is labeled as including thinking tokens[1], so thinking is charged at the output rate: $10 per 1M now, $20 per 1M from 1 December on Standard.
Can I use Gemini 4 Argon on the free tier?
No. Google's pricing page says it is not available on the Free tier[1]. Google AI Ultra subscribers can try it in AI Studio and Antigravity, per the launch post[3], but that is a subscription feature, not API access.
How does Argon compare with Claude Opus 5.5 on cost?
From 1 December the price cards match: $4 input, $20 output and $0.20 cache read per 1M tokens on both[1][5]. Until 30 November Argon is half the price. Real cost then depends on tokens used per task and, for Anthropic, cache-write charges.
Is there a cheaper way to run it?
Batch and Flex list half of Standard's price on every line[1]. Cache hits cost one twentieth of the input price ($0.10 against $2.00 now). Shorter thinking and shorter prompts cut the volume the rate applies to.
Sources
- Gemini API pricing, Gemini 4 Argon section (accessed 18 October 2026)
- Gemini API changelog, October 2026 entries (accessed 18 October 2026)
- Google: Build with Gemini 4 Argon (accessed 18 October 2026)
- Gemini API model page: gemini-4-argon (accessed 18 October 2026)
- Anthropic: Claude pricing (accessed 18 October 2026)
- OpenAI: API pricing, GPT-6.1 Sol (accessed 18 October 2026)
- DeepSeek API: models and pricing (accessed 18 October 2026)
- Google Cloud Billing: create, edit or delete budgets and budget alerts (accessed 18 October 2026)
Comments · 0
Beta: comments are stored locally on your device and not visible to other readers.
No comments yet. Be the first to share your thoughts.