CTA keyword: ARGON · from the @compedge.ai x @am_roshh carousel
Google announced Gemini 4 Argon on 30 Sep 2026 (1 Oct, 01:31 IST). Most people can't use it yet, but the API prices are public. Here's how to work out what a job will cost you before you switch.
The formula
Cost = (tokens in ÷ 1,000,000) × input price + (tokens out ÷ 1,000,000) × output price
At the introductory prices (Google's launch pricing):
- Input: $2 per 1M tokens
- Output: $10 per 1M tokens
At the regular prices (after the intro period):
- Input: $4 per 1M tokens
- Output: $20 per 1M tokens
So the intro formula is: (tokens in ÷ 1M) × $2 + (tokens out ÷ 1M) × $10. For the regular price, double both numbers.
Two more things Google listed:
- Cached input is 95% off. Input you send again and again (a long system prompt, the same document) can cost far less.
- Output can run up to 1M tokens in a single response.
Copy it into a sheet
Put tokens in into column A and tokens out into column B, from row 2.
Intro price: =A2/1000000*2 + B2/1000000*10
Regular price: =A2/1000000*4 + B2/1000000*20
Or into code
def argon_cost(tokens_in, tokens_out, regular=False):
"""Gemini 4 Argon API cost in US dollars."""
price_in, price_out = (4, 20) if regular else (2, 10)
return tokens_in / 1_000_000 * price_in + tokens_out / 1_000_000 * price_out
print(argon_cost(100_000, 5_000)) # 0.25 (intro)
print(argon_cost(100_000, 5_000, True)) # 0.5 (regular)
Use the token counts your API responses report, not guesses. The example sizes below are only examples; your jobs will differ.
Three worked jobs
1. A small chat reply
2,000 tokens in, 500 tokens out.
- Intro: (2,000 ÷ 1M) × $2 = $0.004, plus (500 ÷ 1M) × $10 = $0.005. Total $0.009.
- Regular: $0.008 + $0.010 = $0.018.
- A thousand replies like this: $9 at intro prices, $18 at regular prices.
2. A long-document summary
100,000 tokens in (a long report), 5,000 tokens out (the summary).
- Intro: (100,000 ÷ 1M) × $2 = $0.20, plus (5,000 ÷ 1M) × $10 = $0.05. Total $0.25.
- Regular: $0.40 + $0.10 = $0.50.
- A hundred documents: $25 intro, $50 regular.
3. An agent run
An agent re-reads its context on every step, so input adds up fast. Say the whole run uses 1,000,000 tokens in and 50,000 tokens out.
- Intro: (1,000,000 ÷ 1M) × $2 = $2.00, plus (50,000 ÷ 1M) × $10 = $0.50. Total $2.50.
- Regular: $4.00 + $1.00 = $5.00.
- With caching: if 800,000 of those input tokens are cached and the 95% discount applies to the input price, they cost 800,000 ÷ 1M × $2 × 0.05 = $0.08. Add $0.40 for the other 200,000 input tokens and $0.50 for output: about $0.98 at intro prices ($1.96 at regular). Check how caching is billed on Google's pricing page before you rely on it.
The lesson: on agent runs, input is most of the bill. Trim what the agent re-reads, and cache what you can.
Who can use it, and when
Google is rolling Gemini 4 Argon out in steps:
- Trusted cyber defenders first, through Google's Fairwind Program.
- Then paid API customers and Google AI Ultra subscribers, as the start of the wider rollout to developers, enterprises and consumers.
- Everyone else later. Google has not given a date for broad access.
Google says it is also taking part in the US government's voluntary pre-release process for the model.
If you're not in step 1 or 2, you can't run it yet. You can still price your jobs now, so you know what switching would cost the day it opens up.
Google's claims vs the independent check
Google's own benchmark numbers (company claims, from Google's launch post):
- DeepSWE v1.1: 77.9%
- AutomationBench: 51.3% (Google says #1)
- LVBench: 91.7%
- CWE-bench v1: 68% (Google says tied for first)
The independent check (Artificial Analysis, published 1 Oct 2026):
- Gemini 4 Argon (high) matches GPT-6 Astra (max) on the Artificial Analysis Intelligence Index, at about 60% of the cost per task.
- Hallucination rate: 15% for Argon vs 51% for GPT-6 Astra on its test.
- Artificial Analysis also lists the same $2 / $10 per 1M token prices.
The pushback: Bloomberg reported that some Google employees say the model struggles with some real-world coding tasks. Google disputes that. Treat it as a claim, not a finding, and test it on your own work.
Before you switch
- Price three of your real jobs with the formula above, at both intro and regular prices.
- Budget on the regular price. The intro price only runs for the intro period.
- Run your own test set when you get access. Benchmarks, Google's or anyone's, are not your workload.
Sources
- Google DeepMind, Gemini 4 Argon: our next era of frontier intelligence, 30 Sep 2026, 20:01 UTC
- Google blog, Gemini 4 Argon launch post (prices, rollout, benchmarks), 30 Sep 2026
- Artificial Analysis, Gemini 4 Argon model page (listed prices), 30 Sep 2026
- Artificial Analysis results (Intelligence Index, cost per task, hallucination rate), via Techmeme, 1 Oct 2026, 00:35 UTC
- TechCrunch, limited rollout, no broad release date, 30 Sep 2026, 23:43 UTC
- Bloomberg, employee concerns and Google's response, via Techmeme, 30 Sep 2026, 20:25 UTC
Prices and access as published on 30 Sep and 1 Oct 2026. Check Google's pricing page before you commit a budget.