Gemini 4 Argon is Google's flagship reasoning model, released on October 13, 2026 and reachable at the endpoint `gemini-4-argon`. Its distinguishing feature is not a benchmark: it is a single unified budget of 1,048,576 tokens that covers input and output together, with the ability to generate up to that whole budget in one request instead of stopping at a fixed output ceiling.
Three things make it worth understanding even if you will not use it this week. It launched at half its eventual price. It shares one budget between reading and writing, which changes how you size a request. And Google paired it with a coding agent called Antigravity that is bundled into enterprise subscriptions rather than sold on its own.
Everything below is from Google's own pages, read on October 15, 2026.
That last row is deliberate. The pages Google publishes for this model document limits and pricing in detail and do not state a knowledge cutoff, so the honest entry is that it is unpublished rather than a guess from a third-party registry.
Most models give you a context window and a separate maximum output. Argon gives you one pool. Google's model guide describes a unified limit of 1,048,576 tokens that includes both input and output, where output covers thinking tokens, function calls and the final response.
That has a consequence people miss: a large input reduces the output available in the same request, because they draw on the same number. A 900,000-token codebase read leaves roughly 148,000 tokens for the answer, not 1,048,576.
The second consequence is better news. The default single-request ceiling is 65,536 output tokens when you stream, or 32,768 when you call the REST endpoint without streaming, and you can raise it up to the full budget:
# pip install --upgrade "google-genai>=2.30.1" from google import genai from google.genai import types client = genai.Client() response = client.models.generate_content( model="gemini-4-argon", contents="Return a complete migration plan for every module listed below.", config=types.GenerateContentConfig( max_output_tokens=200_000, thinking_config=types.ThinkingConfig(thinking_level="HIGH"), ), ) print(response.candidates[0].finish_reason) print(len(response.text.split()), "words")
Print the finish reason. A response that stopped at the ceiling and one that finished cleanly read the same otherwise, and the finish reason is the documented field that tells them apart.
Over REST, the same request needs streaming above 32,768 output tokens:
curl -N -sS -X POST \ "https://generativelanguage.googleapis.com/v1beta/models/gemini-4-argon:streamGenerateContent?alt=sse" \ -H "x-goog-api-key: $GEMINI_API_KEY" \ -H "Content-Type: application/json" \ -d '{ "contents": [{"parts": [{"text": "Plan a migration for the modules below."}]}], "generationConfig": {"maxOutputTokens": 200000} }'
Thinking level is the other dial, and it has a cost consequence rather than only a quality one. Thinking tokens are billed as output at the output rate. Google's guide notes that `gemini-4-argon` defaults to MEDIUM, while a variant called `gemini-4-argon-cyber-permissive` defaults to HIGH and is intended for trusted defenders.
# cheap exploratory call config = types.GenerateContentConfig(thinking_config=types.ThinkingConfig(thinking_level="LOW")) # hard reasoning call, billed at the output rate for every thinking token config = types.GenerateContentConfig(thinking_config=types.ThinkingConfig(thinking_level="HIGH"))
Before you build anything on a new model, check two things: that your credential reaches it, and which region serves it.
Google's model page lists Argon under a Gemini 4 heading, with previous generations below it. The names that matter when you are choosing:
The practical reading of that table is that two models answer most questions. Argon for work where the reasoning is the bottleneck and the answer is long. Flash for work where volume is the bottleneck, because a Flash-class model at a fraction of the price will do the same classification or extraction task.
You can check what your own project can reach rather than trusting a list:
curl -sS "https://generativelanguage.googleapis.com/v1beta/models" \ -H "x-goog-api-key: $GEMINI_API_KEY" | head -c 600
Without a credential that call returns this, captured on October 15, 2026:
{ "error": { "code": 403, "message": "Method doesn't allow unregistered callers (callers without established identity). Please use API Key or other form of API consumer identity to call this API.", "status": "PERMISSION_DENIED" } }
Google serves this model from three locations, and the client form differs between the API key path and the enterprise path:
from google import genai from google.genai import types # enterprise path, authenticated with Application Default Credentials client = genai.Client(enterprise=True, project="PROJECT_ID", location="eu") response = client.models.generate_content( model="gemini-4-argon", contents="List the data residency questions to ask before using this in production.", config=types.GenerateContentConfig(thinking_config=types.ThinkingConfig(thinking_level="LOW")), ) print(response.text)
The documented locations are `global`, `us` and `eu`. Residency decisions usually pick the region before the model, so check it against your own commitments instead of inheriting a default.
Argon is a model, and Google's agentic coding story runs through a product called Antigravity, described on the launch pages as an agentic development platform. It reaches you through three doors, and none of them is a standalone purchase.
The enterprise post from August is the useful one for context. It added IDE extensions so developers can use Antigravity inside editors including VS Code, alongside Antigravity 2.0 and an Antigravity CLI, and it put administrative and spend controls plus pooled usage into the subscription. So the answer to "how do I buy Antigravity" is that you buy a Gemini Enterprise licence, and the answer to "how do I try it on a personal plan" is that the macOS app interface needs Ultra.
The launch post states that Argon is available globally to developers and enterprises through the API in AI Studio and the Gemini platform, to Google AI Ultra subscribers in Antigravity, and to Ultra and Pro subscribers in the Gemini app and AI Mode in Search. Google Cloud followed on October 14 with its own availability post. The API pricing page adds the restriction that matters for budgeting: the model is not available on the free tier or on the Google AI Pro plan.
Pricing has a date attached, which is unusual enough to plan around:
Vendor benchmark claims are worth reading with the source in view. Google's launch post states that Argon sets a new state of the art on DeepSWE v1.1 at 77.9%, ranks first on the Vals Index across finance, coding, legal and tax work, and scores 91.7% on LVBench for long video understanding. Those are Google's own numbers on Google's own pages, and no independent replication is cited.
The more interesting material is the engineering write-up, because it describes work on a real codebase rather than a benchmark. Google Cloud engineers published an account of using Argon to optimize the Agent Development Kit, a codebase of roughly 200,000 lines, overnight. The published result was a 22-file change backed by 17 new test functions that landed in main with the commit note "Argon generated these changes", after a run of 158 internal turns of tool calling.
The same launch post describes a Chrome refactor that separated core page handling from a legacy folder of roughly 6 million lines inside a 40 million line codebase, and a vLLM optimization of Gemma 4 E2B that produced a 2.85x end-to-end speedup on an A100, moving throughput from about 37,900 to over 108,000 tokens per second. Read those as case reports from the vendor, not as guarantees, and note that all three are the kind of task where a long output budget is the enabling feature.
Two operational details sit alongside the benchmarks. Agentic video understanding is generally available and off by default, so it has to be turned on. And Google states that for trusted defenders and its own internal teams, Argon is being released without cyber guardrails so they can use full frontier-level cybersecurity capability, which is a distribution decision worth knowing about if you work in security.
Worth stating as plainly as the claims above:
The practical follow-ups are already written: how to use Gemini 4 Argon with opencode for the harness setup, Gemini 4 Argon vs Claude Code for the model decision with both vendors' price tables, and how to use it free for what the tiers actually require. On the tooling side, the Jev opencode guide and the complete Jev guide cover the layer around the model, and how much an AI engineer costs covers the hiring version of the same question.
If the interesting part for you is turning a model like this into a workflow a team can run, that is the work we do: we embed AI engineers inside your existing engineering team, spec first, with the review gates and the exit written down. Send a written intake or see what we do.
Is Gemini 4 Argon better than Gemini 3.8 Flash?
They are for different jobs. Argon is the flagship reasoning model with a unified budget up to 1,048,576 tokens. Flash is positioned by Google for long-horizon software engineering and autonomous agents at Flash cost, which makes it the cheaper default for volume work.
What makes the unified token limit different?
Input and output share one 1,048,576-token budget, so reading more leaves less room to write in the same request. Models with a separate context window and output ceiling do not behave that way.
When does the price change?
December 1, 2026. Input moves from $2 to $4 per million tokens, output from $10 to $20, and cache reads from $0.10 to $0.20.
Can I use it for free?
Not on the API. The published pricing page states it is unavailable on the free tier and on the Google AI Pro plan. The app path needs Pro or Ultra, and Antigravity needs Ultra.
What is `gemini-4-argon-cyber-permissive`?
A variant Google's model guide lists with a HIGH default thinking level, intended for trusted defenders and internal teams working on cybersecurity tasks.
How long an answer can I get in one request?
Up to 1,048,576 tokens minus your input, provided you raise `max_output_tokens` and stream above the default single-request ceiling of 65,536, or 32,768 for non-streaming direct HTTP calls.
[1] Google, Introducing Gemini 4 Argon, October 13, 2026. Availability across the API, Antigravity and the Gemini app, pricing, the DeepSWE v1.1, Vals Index and LVBench claims, the Antigravity internal figures, the Chrome and vLLM case reports, and the cyber guardrail decision.
[2] Google, Gemini 4 Argon: our next era of frontier intelligence, October 2026. The DeepSWE state-of-the-art claim and the companion responsible-development link.
[3] Google, Gemini models, read October 15, 2026. The model ID and endpoint, the unified 1M context window and long decodes, the Gemini 4 and Gemini 3 headings, and the Antigravity Agent, embedding and robotics endpoints.
[4] Google, Gemini Developer API pricing, read October 15, 2026. Introductory and December rates, batch rates, cache pricing, and the free tier and Pro plan restriction.
[5] Google, Gemini 4 Argon model guide, read October 15, 2026. The unified token budget, default and maximum output, thinking levels including the cyber-permissive variant, implicit caching, agentic video understanding being off by default, regions, and SDK versions.
[6] Google Cloud, Gemini 4 Argon available on Google Cloud, October 14, 2026. Enterprise availability.
[7] Google Cloud, Expanding Google Antigravity for enterprise customers, August 21, 2026. Antigravity in Gemini Enterprise, the licence types, IDE extensions including VS Code, Antigravity 2.0, the CLI, and the administrative and spend controls.
[8] Google, Gemini 4 Argon comes to the Gemini app, October 13, 2026. The macOS app interface for Ultra subscribers and the app availability windows.
[9] Google, Optimizing a 200,000-line Python codebase overnight with Gemini 4 Argon, October 13, 2026. The ADK case study: the 22-file change, 17 new test functions, the commit note and the 158-turn run.
Maxpertise is an AI-native engineering company. We embed native AI engineers inside your team, live in about 10 days. Please enable JavaScript to view the site, or email contact@maxpertise.net.