Gemini 4 Argon ties GPT-6 Astra on the Artificial Analysis Intelligence Index and raises Gemini's output ceiling from 64K to 1M tokens. It became generally available on October 13, 2026 on the paid Gemini API, in Google AI Studio and on Vertex AI, and you can call it on AIHubMix as gemini-4-argon at $2 input / $10 output per million tokens. Google's price doubles on December 1.
The engineering question is no longer when you can call it but what changes when you do. This post covers the numbers that should drive the decision, four failure modes to design for, a client that calls Argon with sensible defaults and a fallback, and what's still open.
TL;DR
-
Access: generally available since October 13. Model ID
gemini-4-argonon Google's Gemini API (paid tier only), Vertex AI and AIHubMix's OpenAI-compatible endpoint. - 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.
- Defaults to know: thinking level medium (every benchmark ran at high); output capped at 65,536 tokens unless you use long decode; temperature and top_p are ignored.
- Behaviour to plan for: it declines when unsure. 15% hallucination rate, but 50% accuracy on the same questions.
- Volume and speed: about 62k output tokens per Artificial Analysis task (GPT-6 Astra: about 27k); about 45 seconds to the first answer token at high.
- Price: $2 / $10 per million tokens through November 30, then $4 / $20. Cached input $0.10, then $0.20.
Where things stand
| Item | Status on Oct 14, 2026 |
|---|---|
| Release | Generally available since Oct 13 |
| Model ID | gemini-4-argon |
| Where to call it | Gemini API paid tier, AI Studio, Vertex AI, AIHubMix |
| Context window | 1,048,576 tokens, shared |
| Max output | 65,536 default, 1M with long decode |
| Thinking levels | Low, medium (default), high |
| Inputs / output | Text, image, video, audio, PDF / text |
| Knowledge cutoff | March 2026 |
| Google price per 1M | $2 / $10, then $4 / $20 from Dec 1 |
| AIHubMix price per 1M | $2 / $10, cached $0.10 |
"Shared" matters: input, thinking and output all come out of the same 1,048,576 tokens, so a 900K-token prompt leaves about 150K for thinking and answer. Google's What's new in Gemini 4 Argon page has the full spec and migration notes. Function calling, structured outputs, code execution, Search and Maps grounding, URL context, file search and computer use (preview) are supported. Explicit context caching, image and audio output and the Live API are not; caching works implicitly, with a 95% discount on repeated prompt prefixes.
Cyber teams in Google's Fairwind Program can still get a version without cyber guardrails. The public model refuses harmful cyber, chemical, biological, radiological and nuclear requests, so security results shared by Fairwind users 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.
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.
| Benchmark | Argon | Best other model |
|---|---|---|
| DeepSWE v1.1 | 77.9% | Opus 5.5, 74.2% |
| FrontierSWE v2 | 55.0% | Astra, 65.5% |
| Terminal-bench 4.0 | 57.4% | Opus 5.5, 66.4% |
| OSWorld-2.0 | 69.2% | Astra, 72.6% |
| Vals Index | 68.9% | Opus 5.5, 67.0% |
| Vals Finance Agent v2 | 65.4% | Fable 5.1, 58.9% |
| Harvey Legal Agent | 19.6% | Fable 5.1, 6.7% |
| GraphWalks 256k–1M | 84.2% | Astra, 71.8% |
| LVBench | 91.7% | Astra, 87.5% |
| CWE-bench v1 | 68.0% | Astra, 68.0% |
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, with one caveat on rank. 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. Claude Opus 5.5 leads that index at 58, and Claude Sonnet 5.5 and Meta's new Muse Wonder Full both score 56. 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: test Argon now. This is where its lead is widest.
- Large code migrations: test Argon. The coding scores are mixed (it leads DeepSWE, trails FrontierSWE), but long decode removes the hand-chunking these jobs needed (next section).
- Front-end work: run your own evals first. Bloomberg reported that some Google engineers with access say Argon struggles on real coding tasks, particularly front-end design (via Startup Fortune).
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 still stops at 65,536 by default. Past that you need long decode, which Google's long decode guide describes:
-
Interactions API: set
background=Trueandmax_output_tokensabove 65,536 (up to 1,048,576 minus your input). Google's servers write the answer in 64K hops for up to 24 hours; you poll the interaction and join itsmodel_outputsteps. -
generateContent API: the client continues. Each hop ends with
finish_reason: CONTINUATIONand acontinuation_token; recent SDKs (google-genai 2.30.1 and later) send the next hop for you. - Billing: every hop is billed, including hops that only think, and each hop re-reads the prompt plus everything written so far as input. That repeated input gets the 95% implicit-cache discount.
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. Most requests don't need it; Google says chat, Q&A and extraction finish well under 32K tokens.
Four failure modes to design for
1. Responses come back empty or cut off
What you'll see: a response with status "incomplete" and truncated or empty text, while you're still billed for the thinking tokens.
Why: max_output_tokens caps thinking and answer together, and Argon can spend more of it on thinking than earlier models. A cap such as 8,192 can run out before the answer starts.
Fix: set the cap to at least 32,768, as Google recommends, or leave it at the default. To spend less, lower the thinking level instead of the cap. For answers over 64K, use long decode.
2. The default thinking level isn't the one in the benchmarks
What you'll see: results that don't match the published scores. Argon defaults to medium, but Google, Artificial Analysis and Vals all ran it at high.
Why: Argon supports low, medium and high and can't turn thinking off. Google recommends medium for most work, including complex code and agents, and says low can lose quality on hard tasks. Nobody has published results below high.
Fix: set the level explicitly and keep it in configuration. On AIHubMix's OpenAI-compatible route, reasoning_effort maps straight to it (minimal falls back to low). Run your own evaluation at all three levels and pick the lowest one that meets your bar.
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. Timeouts, latency and volume
What you'll see: slow first tokens and long answers. 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 averaged about 62k output tokens per Artificial Analysis task, against about 27k for GPT-6 Astra.
Why: Argon thinks at length, and Google warns that hard prompts can take several minutes.
Fix: set HTTP client timeouts to at least 15 minutes, stream long answers, and keep latency-sensitive paths on another model. Budget at the $4 / $20 price that starts on December 1: Artificial Analysis puts Argon at $1.99 per index task now and $3.98 at that price, against $1.30 for Gemini 3.1 Pro Preview.
Your first call on AIHubMix
Argon runs on AIHubMix's OpenAI-compatible endpoint with the standard OpenAI SDK. AIHubMix's current prices match Google's introductory rate:
| Item | AIHubMix price |
|---|---|
| Input | $2.00 per 1M tokens |
| Output, including thinking | $10.00 per 1M tokens |
| Cached input | $0.10 per 1M tokens |
| Video or audio input | $2.00 per 1M tokens |
| Web search | $0.014 per request |
AIHubMix hasn't said what it will charge after December 1.
The client below calls gemini-4-argon with an explicit thinking level, a generous output cap and a long timeout, and falls back to Gemini 3.8 Flash if the call fails:
import os
from openai import OpenAI, APIError
client = OpenAI(
base_url="https://aihubmix.com/v1",
api_key=os.environ["AIHUBMIX_API_KEY"],
timeout=900, # hard prompts can take several minutes
)
MODEL = os.environ.get("ARGON_MODEL_ID", "gemini-4-argon")
EFFORT = os.environ.get("ARGON_EFFORT", "medium") # low, medium (default) or high
FALLBACK = "gemini-3.8-flash"
PROMPT = "List the termination clauses in the contract below. Say so if you are unsure about any of them.\n\n<contract text>"
def ask(model):
return client.chat.completions.create(
model=model,
reasoning_effort=EFFORT,
max_tokens=32768, # thinking and answer share this cap
messages=[{"role": "user", "content": PROMPT}],
)
try:
resp = ask(MODEL)
except APIError as e:
print(f"{MODEL} failed ({e}); falling back to {FALLBACK}")
resp = ask(FALLBACK)
print(resp.model)
print(resp.choices[0].message.content)
print(resp.usage) # real token counts for your budget
Run it:
pip install openai
export AIHUBMIX_API_KEY="your-aihubmix-key"
python argon_client.py
There's no temperature in the request on purpose: Argon ignores temperature, top_p and top_k. On the OpenAI-compatible route, AIHubMix's Gemini guide says function calling needs tool_choice="auto".
Long decode runs through background execution, which AIHubMix's Gemini guide documents for its native Gemini endpoint (https://aihubmix.com/gemini). Whether it works for Argon on AIHubMix hasn't been confirmed yet.
What's still open
- Results below high. Every published benchmark used high; there's no data on what low or medium give up.
- AIHubMix pricing from December 1. AIHubMix matches Google's introductory rate today and hasn't announced the later price.
- Long decode through AIHubMix. Documented by Google and plausible through AIHubMix's native Gemini endpoint, but not yet confirmed for Argon.
- Rival scores in Google's table. Mostly self-reported or taken from public leaderboards.
Keep reading: the Gemini 4 Argon series
- Failure modes 1 and 2 in depth. How to pick a thinking level and when to turn on long decode: Gemini 4 Argon thinking_level and Long Decode: A Developer's Guide
- Failure mode 4 is where the bill comes from. The measured cost per task, before and after the price change: Gemini 4 Argon Cost per Task, Measured: $1.99 Until November 30, $3.98 After
- Moving an existing Gemini 3.x client over? Failure modes and fixes for the switch: Moving Gemini 3.x Code to Gemini 4 Argon: Failure Modes and Fixes
Sources
- Gemini 4 Argon: Built to solve your hardest problems (Google)
- What's new in Gemini 4 Argon (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)
Top comments (0)