Gemini 4 Argon: Frontier Scores, a 1M-Token Output Limit, and How to Call It Now

Gemini 4 Argon is Google's first Gemini 4 model and its strongest so far. Independent testers put it level with GPT-6 Astra, it tops the Vals Index for finance, legal and tax work, and it can write up to 1M tokens in a single answer. It also costs less per token than Astra, even after its introductory price ends.

Argon became generally available on October 13, 2026. Google's announcement opened it to developers and enterprises through the paid Gemini API, Google AI Studio and Vertex AI, and to Google AI Pro and Ultra subscribers in the Gemini app. It's also live on AIHubMix as gemini-4-argon. This post covers what changed, where Argon wins and where it doesn't, and how to make your first call.

The short version

• Argon replaces Gemini 3.1 Pro Preview at the top of Google's lineup. It is Google's first new model above the Flash tier in more than seven months.
• On the Artificial Analysis Intelligence Index it scores 53, the same as GPT-6 Astra at max reasoning and 23 points above Gemini 3.1 Pro Preview. Claude Opus 5.5 (58), Claude Sonnet 5.5 (56) and Meta's Muse Wonder Full (56) score higher.
• It has the lowest hallucination rate of any top-scoring model on Artificial Analysis' test (15%), but it also answers fewer of those questions correctly than Astra.
• It trails Claude Opus 5.5 and GPT-6 Astra on terminal-heavy coding and computer use.
• Price: $2 input and $10 output per million tokens through November 30, 2026, then $4 and $20. AIHubMix currently charges the same $2/$10.
• The default thinking level is medium, but every published benchmark ran Argon at high.

What changed from Gemini 3.1 Pro Preview

Google frames Argon around "long-horizon" work: real-world software engineering, enterprise knowledge work such as legal and finance, and cyber defense. The biggest concrete change is how much it can write in one answer.

3.1 Pro Preview prices are for prompts up to 200k tokens ($4 / $18 above that). Argon has a single price tier with no long-context surcharge; the $2 / $10 rate runs through November 30, 2026.

A 1M-token output limit, with a switch. A normal Argon call still stops at 65,536 output tokens. To go further you turn on long decode: in the Interactions API you set background execution and a max_output_tokens above 65,536, and Google's servers write the answer in 64K hops for up to 24 hours. Google's own examples show why this matters: its agents have used Argon for C and C++ to Rust migrations, including more than 800K lines of the Fuchsia Zircon kernel. Those are jobs where a 64K cap forced you to chunk the work by hand.

One shared context window. Argon's context window is 1,048,576 tokens, and input, thinking and output all come out of it. A 900K-token prompt leaves room for only about 150K tokens of thinking and answer.

Adjustable thinking, set to medium by default. Argon supports low, medium and high thinking levels and can't turn thinking off. Google recommends medium for most work, including complex code and agents, and high for the hardest reasoning. Every result in the next section was measured at high, and nobody has published results at the lower levels. [TBD: accuracy and tokens by thinking level]

A new naming scheme. There's no Pro or Flash label, just a generation number and a codename. Google had planned a Gemini 3.5 Pro earlier in 2026 and moved past it.

Cyber guardrails for most users. Argon was trained for defensive cyber work, and Google says it can "autonomously find, validate, and patch critical software vulnerabilities." The public model refuses harmful cyber, chemical, biological, radiological and nuclear requests. Security teams that need the version without cyber guardrails still apply through Google's Fairwind Program.

What it supports. The Gemini API model page lists the details. Its knowledge cutoff is March 2026.

Context caching works only implicitly: repeated prompt prefixes are discounted 95% automatically, and the cachedContents calls you may use with Gemini 3.x aren't supported.

How it compares with GPT-6 Astra, Claude Opus 5.5 and Fable 5.1

The table below comes from Google DeepMind's benchmark table. Google ran Argon at its highest thinking setting and mostly used rivals' self-reported or public-leaderboard numbers, so treat it as vendor data. Artificial Analysis and Vals AI have published independent results, quoted after the table.

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

Three patterns stand out:

• Knowledge work is where the gap is widest. On Harvey's legal agent benchmark Argon scores 19.6% against at most 6.7% for the others. On Finance Agent v2 the lead is more than six points.
• Very long inputs hold up. On GraphWalks between 256k and 1M tokens, Argon keeps 84.2% where the others fall to around 65–72%.
• Coding is mixed. Argon leads DeepSWE but trails on FrontierSWE (Astra) and Terminal-bench 4.0 (Opus 5.5).

The independent numbers broadly agree, with one caveat on rank. Artificial Analysis scores Argon (high) at 53 on its Intelligence Index, tied with GPT-6 Astra (max) and one point ahead of GPT-6.1 Sol (max). Claude Opus 5.5 leads that index at 58, and Claude Sonnet 5.5 and Meta's new Muse Wonder Full both score 56. Its Terminal Bench 4 run puts Argon at 57%, behind Claude Sonnet 5.5, Claude Opus 5.5 and GPT-6 Astra. Vals ranks Argon first of the models on its index, with its best single result on Finance Agent v2, and lists computer use (CUA-bench) as a weak spot.

The trade behind "low hallucination"

Artificial Analysis' Omniscience test asks hard factual questions and penalises confident wrong answers. Argon's hallucination rate is 15%, the lowest of any model scoring 45 or more on the index; GPT-6 Astra's is 51%. But Argon's accuracy on the same questions is 50%, against 63% for Astra. Argon gets its low hallucination rate by declining to answer when it isn't sure, not by knowing more.

For legal and financial review that is often the right trade: a refusal is easy to spot and route to a person or another model, while a confident wrong figure can slip past a reviewer. It only pays off if your application has somewhere to send the questions Argon declines.

Where Argon is not the best choice

• Terminal-heavy coding agents. Google's own table has Opus 5.5 at 66.4% on Terminal-bench 4.0 against Argon's 57.4%, and Artificial Analysis' run agrees on the order.
• Computer use. Astra leads OSWorld-2.0, and Vals ranks Argon seventh of eight on CUA-bench.
• Front-end work. Bloomberg reported that some Google engineers with access say Argon struggles on real coding tasks, particularly front-end design (via Startup Fortune). That's anecdotal, but it fits the mixed coding results above.
• Questions that need an answer every time. A 50% Omniscience accuracy means many "I'm not sure" responses on hard factual recall.
• Latency-sensitive work. At high thinking, Artificial Analysis measured about 45 seconds to the first answer token and 99.5 to 116.5 output tokens per second, depending on the provider. Argon also averaged about 62k output tokens per task, against about 27k for Astra. Google itself warns that hard prompts can take several minutes.

Who should switch, wait, or stay

"Test Argon now" means: run your own evaluation at medium and at high, budget at the $4/$20 price that starts on December 1 and Artificial Analysis' average of about 62k output tokens a task, and keep your current model as the fallback until the results are in.

Calling Argon on AIHubMix

Argon is on AIHubMix's OpenAI-compatible endpoint, so the standard OpenAI SDK works. AIHubMix's current prices match Google's introductory rate:

AIHubMix hasn't said what it will charge after Google's price doubles on December 1. [TBD: AIHubMix price from December 1]

import os
from openai import OpenAI

client = OpenAI(
    base_url="https://aihubmix.com/v1",
    api_key=os.environ["AIHUBMIX_API_KEY"],
    timeout=900,  # hard prompts can take several minutes
)

resp = client.chat.completions.create(
    model="gemini-4-argon",
    reasoning_effort="medium",  # Argon's default; "high" is what the benchmarks used
    max_tokens=32768,           # thinking and answer share this cap
    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)
print(resp.usage)  # check real token counts against your budget

On the OpenAI-compatible route, reasoning_effort maps to Argon's thinking levels: low, medium and high map directly, minimal falls back to low, and leaving it out gives you medium. Don't send temperature or top_p; Argon ignores sampling parameters. Keep max_tokens generous, because thinking and the answer share it and a request that runs out can come back empty.

Answers longer than 64K tokens need long decode through background execution, which AIHubMix's Gemini guide documents for its native Gemini endpoint. Whether that works for Argon on AIHubMix hasn't been confirmed yet. [TBD: long decode through AIHubMix]

FAQ

Is Gemini 4 Argon available to everyone?
It became generally available on October 13, 2026 for paid Gemini API users, in Google AI Studio and Vertex AI, in GitHub Copilot (Pro+, Max, Business and Enterprise), and for Google AI Pro and Ultra subscribers in the Gemini app. It isn't on the Gemini API free tier. AIHubMix also serves it.

What is Gemini 4 Argon's model ID?
gemini-4-argon, both on Google's Gemini API and on AIHubMix.

How much does Gemini 4 Argon cost?
$2 per million input tokens and $10 per million output tokens through November 30, 2026, then $4 and $20. Cached input is $0.10 per million, rising to $0.20. Output prices include thinking tokens.

What thinking level does Gemini 4 Argon use by default?
Medium. You can choose low, medium or high, but you can't turn thinking off. Published benchmarks were run at high.

How big is Gemini 4 Argon's context window?
1,048,576 tokens, shared by input, thinking and output. A normal call writes up to 65,536 tokens; long decode raises that to 1M minus your input.

Is Gemini 4 Argon better than GPT-6 Astra?
On Artificial Analysis' index they tie. Argon leads on knowledge work, long-context retrieval and video understanding; Astra leads on FrontierSWE, terminal science tasks and computer use.

Why is the hallucination rate low but the accuracy only 50%?
Argon declines to answer more often when it isn't sure. That lowers confident wrong answers but also lowers the share of questions it answers correctly.

Keep reading: the Gemini 4 Argon series

• Every benchmark above was run at high thinking, but the default is medium. How to pick a level and when to use long decode: Gemini 4 Argon Thinking Levels and Long Decode: Defaults, Costs and When to Use Each
• The per-token price looks low until you count tokens per task. The measured cost per task is here: Gemini 4 Argon Pricing: Cheap Per Token, Not Per Task
• Moving a Gemini 3.1 Pro or 3.8 Flash workload over? What breaks and how to fix it: Moving to Gemini 4 Argon from Gemini 3.1 Pro or 3.8 Flash: What Breaks and How to Fix It

Sources

• Gemini 4 Argon: Built to solve your hardest problems (Google)
• Gemini 4 Argon model page (Google AI for Developers)
• 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)
• Gemini 4 Argon on AIHubMix (AIHubMix)