Gemini 4 Argon Goes GA: What Product Teams Should Check

Google made Gemini 4 Argon generally available on October 13, 2026, which turns a model you could demo into one you can put on a roadmap. Gemini 4 Argon GA covers the Gemini API, Gemini Enterprise Agent Platform and BigQuery's generative AI functions.

What Gemini 4 Argon GA actually changes

The Gemini API changelog lists gemini-4-argon as GA on October 13. The Google Cloud model page shows launch stage GA and a retirement date of "October 13, 2027 or later", so you have at least a year before a forced migration. BigQuery's release notes say its generative AI functions now support the model.

The Gemini API model page lists one stable version, so there is no snapshot string to pin.

One shared window, two output ceilings

Input, thinking and output share a single 1,048,576-token window. A normal request returns up to 65,536 tokens (32,768 for non-streaming REST), and max_output_tokens counts thinking and answer together. Set it too low and the model can spend the whole budget thinking and return no text.

Long decode raises the ceiling to 1,048,576 tokens minus your input. The model works in hops of up to 64K tokens, and recent SDKs chain the hops for you on GenerateContent:

from google import genai
from google.genai import types

client = genai.Client()  # Reads GEMINI_API_KEY

response = client.models.generate_content(
    model="gemini-4-argon",
    contents="Write a complete, rigorous proof of the prime number theorem.",
    config=types.GenerateContentConfig(
        max_output_tokens=262_144,  # Total thinking and output across all hops
        thinking_config=types.ThinkingConfig(thinking_level="high"),
        automatic_continuation=True,  # Default in current SDKs
    ),
)

finish = response.candidates[0].finish_reason
print(response.text)
print("Finish reason:", finish)
if finish == types.FinishReason.CONTINUATION:
    print("Cut at the per-request cap: check SDK and API versions.")

On the Interactions API you set background: true instead (see the long decode guide), and without it a value above 65,536 returns a 400 error today. Background runs can last up to 24 hours, and you pay for every token in every hop, including hops spent only on thinking.

Long decode is a job, not a chat reply. It takes minutes, the docs say it isn't suited to latency-sensitive requests, and the model doesn't follow token or word-count targets. Don't promise "a 50,000-word report" in your UI copy.

The known issue: long decode works only on v1beta1

On Gemini Enterprise Agent Platform, the developer guide says long decoding on GenerateContent is supported only in the v1beta1 API, not v1. The Google Gen AI SDK uses v1beta1 by default unless you override it, so the trap is a platform team that pinned v1 for stability. The same guide says the OpenAI compatibility layer and older SDK versions can't retrieve long completions with continuation tokens.

The failure is quiet: output stops at the per-request cap with a CONTINUATION finish reason instead of an error. On the Gemini API side, automatic continuation needs google-genai 2.30.1 or later (@google/genai 2.29.1 for JavaScript), and the OpenAI compatibility endpoint doesn't support long decode there either.

Find your pins before launch (adjust the directories):

grep -rnE "api_version|v1beta1|/v1/" src/ config/

This lists every place an API version is set. Any long-output path on Google Cloud should resolve to v1beta1.

Implicit caching only, and no Free tier

Explicit caching (cachedContents / caches.create) is not supported on Gemini 4 Argon and returns an error. Implicit caching is on by default and discounts cached input tokens by 95%. The Gemini API docs give a 4,096-token minimum, and the Cloud model page lists 6,144, so test against the platform you ship on.

You can no longer create a cache on purpose, so watch total_cached_tokens in the Interactions usage object instead of assuming a hit. Sampling parameters (temperature, top_p, top_k) are ignored.

The model isn't available on the Free tier or the Google AI Pro plan, so prototypes need a billed project. Paid-tier prices double on December 1, 2026, after introductory pricing ends November 30. Put that date in your budget; the pricing page has the table.

Monday morning checklist for Gemini 4 Argon

1. Swap the model ID in staging and rerun your evals at low, medium and high thinking. medium is the default.
2. Delete explicit-cache code and sampling parameters.
3. Raise any max_output_tokens below 32K to at least 32,768.
4. For paths that need more than 64K output, confirm SDK versions, v1beta1 and no OpenAI-compatible gateway in between, and alert on a CONTINUATION finish reason.
5. Decide which features become background jobs, and set client timeouts to 20 minutes for them.
6. Add December 1 to the cost model.

A stable model ID with a year-plus runway is easy to plan around, but only if the plumbing between your product and it (API version, SDK, gateway) lets the long outputs through.