Gemini 4 Argon for Developers: Benchmarks, a 1M-Token Output Cap, and a Model ID You Can't Call Yet

Gemini 4 Argon ties GPT-6 Astra on the Artificial Analysis Intelligence Index, raises Gemini's output cap from 64K to 1M tokens, and launched at $2 input / $10 output per million tokens. You can't call it. As of October 13, 2026 Google hasn't published a model ID, it isn't in the Gemini API model list, AIHubMix lists it as gemini-4-argon but marks it coming soon, and calls to that ID through Google's own API have come back 404.

So the engineering question for most teams isn't how to call Argon. It's what to build today so that switching to Argon is a config change, and what will behave differently after the switch. This post covers the numbers that should drive that decision, four failure modes you can design for now, a client that runs today on Gemini 3.8 Flash and moves to Argon by setting one environment variable, and the things nobody outside the Fairwind Program can verify yet.

TL;DR

• Access: Google's Fairwind Program (cyber defenders) only. Paid API customers and Google AI Ultra subscribers are next, with no date.
• Strong at: knowledge-work agents (Harvey Legal Agent 19.6% vs at most 6.7% for rivals), very long inputs (GraphWalks 256k–1M at 84.2%), DeepSWE.
• Weak at: terminal-heavy coding (Terminal-bench 4.0: 57.4% vs 66.4% for Claude Opus 5.5), computer use, FrontierSWE.
• Behaviour to plan for: it declines when unsure. 15% hallucination rate, but 50% accuracy on the same questions.
• Volume: about 62k output tokens per Artificial Analysis task, against about 27k for GPT-6 Astra.
• Price: $2 / $10 promo, $4 / $20 standard, no end date for the promo. Budget at the standard price.

Where things stand

Google's launch post doesn't state the input limit; the 1M context figure comes from third-party testers. [TBD: official context window] Artificial Analysis also lists speech as an input.

One more thing that will differ between today's users and you: Fairwind members and Google's internal teams get a version without cyber guardrails. The wider release will refuse harmful cyber, chemical, biological, radiological and nuclear requests, so results shared by Fairwind users on security tasks may not carry over.

Argon also drops the Pro / Flash label (just a generation number and a codename) and replaces Gemini 3.1 Pro Preview at the top of the lineup. It's Google's first new model above the Flash tier in more than seven months.

Benchmarks, sorted by what you're building

Google's numbers come from the Google DeepMind benchmark table. Google ran Argon at its highest thinking setting and mostly used rivals' self-reported or public-leaderboard numbers, so read them as vendor data. The "best other model" column compares against GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5, the three rivals in Google's table.

OSWorld-2.0 is the offline subset, partial score; GraphWalks is F1. Google's full table has 19 rows, and Argon leads or ties 14 of them.

Independent results line up with that picture. Artificial Analysis scores Argon (high) at 53 on its Intelligence Index: tied with GPT-6 Astra (max), one point ahead of GPT-6.1 Sol (max), and 23 points above Gemini 3.1 Pro Preview. On its own Terminal Bench 4 run Argon scores 57%, behind Claude Sonnet 5.5, Claude Opus 5.5 and GPT-6 Astra. Vals AI ranks Argon first on the Vals Index, with Finance Agent v2 as its best single result, and puts it seventh of eight on CUA-bench (computer use).

Reading that by workload:

• Agents that live in a terminal or drive a desktop: stay on Opus 5.5 or Astra. Google's table and the independent runs both have Argon trailing here.
• Agents that read contracts, filings or long logs: plan for Argon. This is where its lead is widest.
• Large code migrations: plan for Argon. The coding scores are mixed (it leads DeepSWE, trails FrontierSWE), but the 1M-token output cap removes the hand-chunking these jobs needed (next section).

What a 1M-token output cap changes

Earlier Gemini models stopped at 64K output tokens, so a large migration or a long report had to be chunked by hand. Argon can write up to 1M tokens, through a new Gemini API feature called Long Decode Continuation that pauses a long response and resumes it across follow-up calls so it doesn't time out.

Google's own examples are the kind of job this is for: its agents used Argon for C and C++ to Rust migrations, including more than 800K lines of the Fuchsia Zircon kernel, and for one library replaced 32K lines of SIMD code with safe Rust that runs 2.7 times faster than the earlier Rust port.

For your client, that means a long answer arrives across several calls, not one. Google hasn't published the request parameters or response fields for Long Decode Continuation yet, so there is nothing concrete to code against today. Keep your chunking logic until there is.

Four failure modes to design for now

1. A guessed model ID returns 404

What you'll see: calls to guessed IDs such as gemini-4-argon through Google's own API have returned 404 errors.

Why: Google hasn't published an ID. AIHubMix lists Argon as gemini-4-argon, but the listing is marked coming soon.

Fix: keep the model ID in configuration with a default that exists today (the client below falls back to gemini-3.8-flash), and switch to gemini-4-argon when the AIHubMix listing opens.

2. You don't know which reasoning levels Argon accepts

What you'll see: no documented list of thinking levels for Argon. Every published score was run at the top setting: Google used its highest thinking setting, and Artificial Analysis and Vals tested at high.

Why: the Gemini API docs don't list Argon at all yet.

Fix: keep reasoning_effort in configuration next to the model ID. On AIHubMix's OpenAI-compatible route it maps to Gemini's thinking settings. [TBD: Argon reasoning levels on AIHubMix]

3. "I'm not sure" answers have nowhere to go

What you'll see: more declined answers than you're used to. On Artificial Analysis' Omniscience test (hard factual questions, confident wrong answers penalised), Argon's hallucination rate is 15% against 51% for GPT-6 Astra, but its accuracy is 50% against 63%. It gets the low rate by declining when unsure, not by knowing more.

Why: that's the trade the model makes. For legal and financial review it's usually the right one: a refusal is easy to spot, a confident wrong figure can slip past a reviewer.

Fix: ask the model to say when it's unsure (the prompt below does), detect those answers, and route them to a person or another model. If your pipeline expects an answer every time, Argon is the wrong fit.

4. Output volume inflates latency and cost

What you'll see: long responses. Argon averaged about 62k output tokens per Artificial Analysis task, against about 27k for GPT-6 Astra. Google hasn't published speed figures. [TBD: output speed and time to first token]

Why: Argon produces more output per task, and output is the expensive side of the price ($10 against $2 per million tokens at the promo).

Fix: budget at the standard price. Artificial Analysis puts Argon at $1.99 per index task at the promo price and $3.98 at the standard price, against $1.30 for Gemini 3.1 Pro Preview. Size timeouts for long runs, and keep latency-sensitive paths on another model until speed figures are out.

A client that runs today and switches later

Argon has a model page on AIHubMix under the ID gemini-4-argon, but it's marked coming soon and has no price yet. [TBD: AIHubMix availability date]

The client below runs today on Gemini 3.8 Flash through AIHubMix's OpenAI-compatible endpoint, with the standard OpenAI SDK. The AIHubMix Gemini guide covers this route and the native Google SDK.

import os
from openai import OpenAI

client = OpenAI(
    base_url="https://aihubmix.com/v1",
    api_key=os.environ["AIHUBMIX_API_KEY"],
)

# AIHubMix lists Argon as gemini-4-argon (coming soon). Prototype on
# Gemini 3.8 Flash, then set ARGON_MODEL_ID=gemini-4-argon once it opens.
MODEL = os.environ.get("ARGON_MODEL_ID", "gemini-3.8-flash")

resp = client.chat.completions.create(
    model=MODEL,
    reasoning_effort="high",
    messages=[
        {"role": "user", "content": "List the termination clauses in the contract below. Say so if you are unsure about any of them.\n\n<contract text>"},
    ],
)
print(resp.choices[0].message.content)

Run it, and switch later without touching the code:

pip install openai
export AIHUBMIX_API_KEY="your-aihubmix-key"
python argon_client.py   # runs on gemini-3.8-flash today

# Once Argon opens on AIHubMix:
export ARGON_MODEL_ID="gemini-4-argon"
python argon_client.py

When Argon opens on AIHubMix, its price and supported parameters will show on that model page.

What can't be verified yet

• Google's model ID. Not published; AIHubMix lists gemini-4-argon as coming soon.
• The context window. Third-party testers report 1M tokens; Google's launch post doesn't say.
• Thinking levels. No list, no default, and no published results below the top setting.
• Long Decode Continuation. Described, but its request parameters and response fields aren't public.
• Speed. No output speed or time-to-first-token figures from Google.
• The promo end date. Google hasn't said when $2 / $10 becomes $4 / $20.
• Rival scores in Google's table. Mostly self-reported or taken from public leaderboards.

Keep reading: the Gemini 4 Argon series

• Failure mode 4 is where the bill comes from. The measured cost per task, at both prices: Gemini 4 Argon Cost per Task, Measured: $1.99 Now, $3.98 at Standard Prices

Sources

• Gemini 4 Argon: our next era of frontier intelligence (Google)
• Gemini models and benchmarks (Google DeepMind)
• Gemini 4 Argon: Google is back as one of the top three labs in intelligence achieved (Artificial Analysis)
• Gemini 4 Argon benchmarks, cost and capabilities (Vals AI)
• AIHubMix Gemini guide (AIHubMix)