Gemini 4 Argon: pricing, limits and what changes on December 1

1. What Google published#

Google introduced Gemini 4 Argon on October 13, 2026 and is rolling it out globally for developers and enterprises through the Gemini API (AI Studio) and its enterprise platform. The table gathers the data from the announcement and from the model page in the API documentation.

Google also states that the model leads the Vals Index, which measures finance, coding, legal and tax tasks. This is the company's own figure: check it against your tasks before deciding.

Documentation: Google: introducing Gemini 4 Argon ↗ · Google AI: gemini-4-argon model page ↗

2. The price changes on December 1#

According to Google's announcement, the rate is introductory: $2 per million input tokens and $10 per million output tokens until November 30, 2026. After that, for the publicly available model, it rises to $4 and $20, double. Cached input tokens carry a 95% discount.

To see the effect, the code calculates a month with 150 million input tokens and 30 million output tokens, an example volume. The pesos use the TRM of October 17, 2026.

from decimal import Decimal as D

RATES = {                                    # USD per million tokens (input, output)
    "until Nov 30, 2026": (D("2"), D("10")),
    "from Dec 1, 2026": (D("4"), D("20")),
}
TRM = D("3174.67")                             # TRM of Oct 17, 2026

INPUT_M, OUTPUT_M = D(150), D(30)            # one month: 150 M input tokens and 30 M output tokens

for period, (t_in, t_out) in RATES.items():
    usd = INPUT_M * t_in + OUTPUT_M * t_out
    print(f"{period}: ${usd:,.2f} a month | COP {usd * TRM:,.0f}")

Output weighs more than input in this example: 30 million output tokens cost as much as 150 million input tokens. Since a reasoning model's output includes thinking tokens (as Codifly's catalog records from the pricing page; the model page does not detail how it is billed), it is worth measuring how many your real case generates.

Documentation: Google AI: Gemini API pricing ↗

3. One window, not two#

The model page describes the 1,048,576-token limit as unified: input and output share the same budget. The default maximum output is 65,536 tokens in synchronous calls and can reach 1,048,576 with long decode. If you send a document of one million tokens, only 48,576 are left for the answer.

WINDOW = 1_048_576          # unified input and output limit
DEFAULT_OUTPUT = 65_536  # default maximum output in synchronous calls

def fits(input_tokens: int, output_tokens: int = DEFAULT_OUTPUT) -> bool:
    """Input and output share the same window."""
    return input_tokens + output_tokens <= WINDOW

print(WINDOW - DEFAULT_OUTPUT)           # maximum input if you reserve the default output
print(fits(1_000_000), fits(983_040), fits(983_041))

4. Caching and inference modes#

The 95% discount applies to cached input tokens, but the model page says it supports only implicit caching and that explicit caching is not available. That means you do not control what gets cached: calculate your savings with a realistic hit fraction, not the maximum. The code shows the cost of 150 million input tokens at $2 with different fractions served from the cache.

from decimal import Decimal as D

def input_cost(m_tokens: D, price: D, cached_fraction: D, discount: D = D("0.95")) -> D:
    """Input cost if part of it is served from the cache at a 95% discount."""
    normal = m_tokens * (1 - cached_fraction) * price
    cached = m_tokens * cached_fraction * price * (1 - discount)
    return normal + cached

for fraction in (D("0"), D("0.5"), D("0.8")):
    print(fraction, input_cost(D(150), D(2), fraction))

The page also lists Batch, Flex and Priority modes. It does not include their prices, so check the pricing page before using them.

5. What to do before November 30#

• Recalculate your budget with the December rate ($4 and $20), not the introductory one.
• Measure how many input, output and reasoning tokens each type of call consumes.
• Compare alternatives at the rate that will apply in your project, not the promotional one: use Codifly's AI model comparator.
• Reserve the output when calculating what fits in the unified window.
• If your savings depend on caching, measure real hits before counting on them.

Go deeperWhat an AI API really costs in Latin American currencies (September 2026)

6. Keep your numbers current#

The figures are those Google published between October 13 and 14, 2026. The rate and the inference modes may change; confirm on the Gemini API pricing page before committing a budget.

• Check the pricing page every month while the introductory rate lasts.
• Note the reading date of each rate.
• Recalculate in pesos with the TRM of the day you budget.