Gemini 4 Argon: What It Is, What's New and How to Use It

Gemini 4 Argon is Google's new flagship reasoning model, broadly released on 13 October 2026. It takes text, images, video, audio and PDFs, writes text, and has a 1,048,576-token window shared by input and output, with "long decode" that lets one answer run far past the usual 64K-token cap. You can use it in the Gemini app (Google AI Pro and Ultra), in Google AI Studio and the Gemini API (paid tier only), and on Google Cloud. API pricing as of October 2026 is $2 in and $10 out per million tokens until 30 November, then $4 and $20.

Key facts

What's new in Gemini 4 Argon

Google pitches Argon at long, multi-step work: software engineering, finance, legal, cybersecurity and science. The announcement by Koray Kavukcuoglu lists several concrete changes.

A 1M unified input-output limit and long decode

Most models cap a single reply at tens of thousands of tokens. Argon shares one 1,048,576-token budget across input, thinking and output. With background mode and a max_output_tokens above 65,536, the Gemini API generates in hops of up to 64K tokens, each continuing from the last, and you join the text from every model_output step at the end. That matters for whole-codebase rewrites, long reports or big data transforms.

Better visual and spatial reasoning, more direct answers

Google says Argon is stronger at visual and spatial reasoning, follows style instructions better, and in the Gemini app "answers questions more directly with less hedging". On DeepMind's model page it reports leading scores on long-context retrieval (GraphWalks BFS 256K to 1M: 84.2% F1) and on agentic coding benchmarks such as DeepSWE v1.1 (77.9%). Those are Google's own numbers; our Gemini 4 Argon vs GPT-6 comparison puts them next to OpenAI's.

Prompt-injection resistance and agent behaviour

For anyone building agents, the safety notes are the most useful part. Google reports an attack success rate of 0.7% on the Gray Swan indirect prompt injection benchmark, and says Argon was trained to respect system boundaries and pause for human confirmation before irreversible steps. That doesn't replace your own controls; see AI agent security for the layers you still need.

New Gemini app features

Alongside the model, the Gemini app gets a chat-to-tasks toggle (Gemini can keep working when you're away), connections to more than 40 apps, and the Antigravity agentic coding interface inside the Gemini macOS app.

How to use Gemini 4 Argon

In the Gemini app

Google AI Pro and Ultra subscribers can pick Argon in the Gemini app, and it's also in AI Mode in Search for those plans. Google hasn't said that free accounts get it.

In Google AI Studio and the Gemini API

Argon needs the paid tier. The model ID is gemini-4-argon and the default thinking level is medium. Google's own quickstart uses the Interactions API:

from google import genai

client = genai.Client()

interaction = client.interactions.create(
    model="gemini-4-argon",
    input="Analyze this payment pipeline for race conditions during retries.",
    generation_config={"thinking_level": "medium"},
)

print(interaction.output_text)

Migration notes from Google's latest-model guide that will bite if you skip them:

• temperature, top_p and top_k are ignored. Remove them.
• Explicit context caching is not supported; implicit caching is on by default (Gemini API minimum 4,096 tokens).
• max_output_tokens covers thinking and output. If your app sets a limit below 32K, raise it to at least 32,768.
• Raise client timeouts; Google suggests at least 15 minutes for hard prompts.

Thinking levels are low, medium and high. Google's advice is to evaluate at all three and pick the lowest one that meets your quality bar, which is also the cheapest.

On Google Cloud (Vertex AI / Gemini Enterprise Agent Platform)

Google Cloud lists Argon as GA from 13 October 2026, with a retirement date no earlier than 13 October 2027. It runs in the global endpoint and the us and eu multi-regions, supports Provisioned Throughput, batch and pay-as-you-go (Standard and Priority), and offers data residency, CMEK and VPC-SC. Per-prompt limits include up to 3,000 images, 10 videos and roughly 8.4 hours of audio. Tuning, RAG Engine and the Live API are not supported.

In Google Antigravity

Antigravity is Google's agentic coding product. Argon is available there for Google AI Ultra subscribers and enterprises. If you're comparing coding agents, our Claude Code vs Codex vs Cursor guide covers the other side of the field.

Gemini 4 Argon pricing (as of October 2026)

From Google's Gemini API pricing page, per 1M tokens:

There is no separate long-context price. Grounding with Google Search gives 5,000 free requests a month shared across Gemini 3 and newer models, then $14 per 1,000. Note that output price includes thinking tokens, so a high thinking level on a simple task costs real money.

The practical takeaway: the introductory price halves your bill until the end of November. If you're evaluating Argon for production, budget at the December rates, not the launch rates.

Limits and trade-offs

• No free API tier. You need billing enabled to try it through the API.
• Text output only. No image generation, no audio generation and no Live API.
• No explicit caching. You can't pin a large shared prefix the way you could on some earlier Gemini models; you rely on implicit caching.
• No sampling control. If your pipeline depends on low temperature for repeatable output, you'll need to rethink it, for example with structured outputs and validation.
• Long decode is asynchronous. Big outputs need background mode and polling, so they don't fit a simple request-response UI.
• Cyber-permissive variant is gated. gemini-4-argon-cyber-permissive is allowlist-only through Google's Fairwind Program.
• Benchmarks are vendor-reported. Google publishes a methodology document, but independent results will take time.

Should you switch?

If you already run on Gemini and your workload is long-context (large documents, video, whole repositories), Argon is the obvious model to evaluate first, because the unified 1M budget and long decode remove real constraints. If you're cost-sensitive, run the same eval at low and medium thinking before choosing high. And if you're choosing between frontier models in general, start with the best LLM guide and test on your own tasks; a benchmark table can't tell you how a model handles your data.