---
type: "post"
id: "what-s-new-in-gemini-4-argon-05c6a3"
slug: "what-s-new-in-gemini-4-argon-05c6a3"
title: "What's new in Gemini 4 Argon"
description: "Google’s Gemini API documentation specifies how Gemini 4 Argon handles long context and long-form output: a shared 1,048,576-token window and background execution for responses beyond 64K tokens."
retrieval_nugget: "Google’s Gemini API documentation specifies how Gemini 4 Argon handles long context and long-form output: a shared 1,048,576-token window and background execution for responses beyond 64K tokens. It also sets introductory API prices through November 30, 2026, giving teams concrete limits and costs to evaluate before integrating the model."
published_at: "2026-10-31"
updated_at: "2026-10-31"
record_date: "2026-10-31"
date_kind: "published_at"
topics: ["gemini","google"]
entities: ["Google","Gemini API","Gemini 4 Argon","Google AI Studio","Antigravity","Interactions API"]
related_patterns: ["harness-architecture-outlives-model-choice"]
novelty_label: "restatement"
what_changed: null
prior_refs: ["memory:e847ee2c-2999-86ef-a2e5-f24effce1f0f","memory:290c73b4-7aac-8074-a379-ee7d04b9eafa","memory:ee28b738-8f75-8009-a495-a020df089691"]
source_url: "https://ai.google.dev/gemini-api/docs/latest-model"
source_urls: ["https://ai.google.dev/gemini-api/docs/latest-model"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:what-s-new-in-gemini-4-argon-05c6a3"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/what-s-new-in-gemini-4-argon-05c6a3/","markdown":"https://newruntime.com/posts/what-s-new-in-gemini-4-argon-05c6a3.md","json":"https://newruntime.com/posts/what-s-new-in-gemini-4-argon-05c6a3.json"}
---

# What's new in Gemini 4 Argon

## Retrieval answer

Google’s Gemini API documentation specifies how Gemini 4 Argon handles long context and long-form output: a shared 1,048,576-token window and background execution for responses beyond 64K tokens. It also sets introductory API prices through November 30, 2026, giving teams concrete limits and costs to evaluate before integrating the model.

## Summary
Google’s Gemini API documentation specifies how Gemini 4 Argon handles long context and long-form output: a shared 1,048,576-token window and background execution for responses beyond 64K tokens. It also sets introductory API prices through November 30, 2026, giving teams concrete limits and costs to evaluate before integrating the model.

## What happened
Google’s Gemini API documentation identifies Gemini 4 Argon, model ID `gemini-4-argon`, as its latest and most capable model. The page says it is intended for complex coding, agentic, math, science, software engineering, and knowledge-work tasks. It is available through the Gemini API and Google AI Studio. The documented default thinking level is `medium`, with `low`, `medium`, and `high` available to control how deeply the model reasons before answering.

The model has a unified context window of 1,048,576 tokens. Input, thinking, and output all share that single window, rather than having separate allowances. The documentation also describes long decode: a response can exceed 64K tokens and extend up to 1M tokens minus the input. That is not simply a larger ordinary synchronous response. For output above 65,536 tokens, Google recommends background execution. A request can set `max_output_tokens` above that threshold; the server then generates continuation hops of up to 64K tokens until completion. Each hop appears as a `model_output` step, so clients must join the text from those steps after polling the interaction to its completed status.

The quickstart uses the Interactions API and shows examples in Python, JavaScript, and REST. The documentation’s longer-generation example sets `background` to true, configures `max_output_tokens` to 262,144 and uses `thinking_level` set to `high`. It advises setting the output limit high enough because thinking and generated output together count against the hard maximum; generation stops when they reach it. The API examples require a Gemini API key.

The introductory price is $2 per million input tokens and $10 per million output tokens through November 30, 2026. The page says prices thereafter are $4 per million input tokens and $20 per million output tokens. The documentation says the model is not available on the Free tier. It does not establish access for every paid consumer subscription, nor does it spell out all operational limits on this page. It also notes that agentic video understanding is generally available and that the Antigravity agent in Gemini Managed Agents supports Argon.

## What the archive adds
[Gemini 4 Argon is now available for developers.](https://newruntime.com/posts/gemini-4-argon-is-now-available-for-developers-fbfa82/) covered the broadening of developer access after a narrower cybersecurity launch. This is a new fact on that known story: the current API documentation adds specific context, long-decode mechanics, and introductory pricing. A separate archive note on AI Gateway is related by product availability, but does not establish the API behavior described here.

## Trend signal
This supports [harness architecture outlives model choice](https://newruntime.com/patterns/harness-architecture-outlives-model-choice/) only weakly: the document makes the execution pattern for very long generation explicit, including background polling and stitching output steps, but says little about production harness design or model portability. The more immediate signal is operational: teams adopting long outputs need client-side orchestration, not just a larger token limit.

## Why now
The documentation sets a dated introductory price window ending November 30, 2026, followed by a stated doubling of both input and output rates on December 1. That makes cost evaluation time-sensitive for teams considering sustained API usage. The source does not explain why Google chose these dates or whether the introductory period will be extended.

## Why it matters
A million-token context and a million-token response are not interchangeable features. The unified window means teams must budget input, reasoning, and output together, while long decode requires a background interaction that can remain in progress across multiple continuation steps. Applications that assume one response payload or synchronous completion may need changes before they can use the full output allowance.

The price schedule also makes workload shape important: output tokens cost five times as much as input tokens during the introductory period, and the listed rates double afterward. Teams evaluating Argon should test representative prompts, measure both token categories, and account for the work of waiting for and assembling long responses. The model’s advertised task scope is broad, but the documentation alone does not demonstrate that it will outperform alternatives on a particular team’s tasks.

## What happens next
- **Teams that rely on the introductory rates will either absorb higher token costs or shift workloads before the stated price change.** (3 months, 75%). Watch for: updated pricing on Google’s Gemini API pricing page after November 30, 2026.

## Limits and uncertainty
This is product documentation, not an independent evaluation of model quality. It does not provide benchmark results, latency, reliability, or evidence that Argon improves outcomes on a particular coding or agent workflow. The page describes a maximum context and output capacity, not typical performance or a guarantee that every request can use the full limit. It also does not establish all account, regional, rate-limit, or paid-tier eligibility details; it points elsewhere for complete pricing and model specifications. The introductory price is stated to end on a date, but the source does not say whether Google could change or extend the schedule.

## Skill to build now
Build a small background-generation test before routing real workloads to Argon. Send a representative long task with `background: true`, poll until completion, and concatenate text from every `model_output` step. Record input and output token usage separately, and verify that your configured maximum leaves room for both thinking and response text. This tests the API mechanics and produces a workload-specific cost estimate.

## Reader takeaway
Gemini 4 Argon’s headline limits come with implementation and cost details: input, thinking, and output share a 1M-token window, while very long responses require background execution and step assembly. Benchmark the workflow you intend to run, and account for the documented price increase after November 30, 2026.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Gemini](https://newruntime.com/topics/gemini.md): Explore the gemini topic hub.
- topic hub: [Google](https://newruntime.com/topics/google.md): Explore the google topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Source Ledger](https://newruntime.com/source-ledger.json): Inspect the public source evidence graph.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)