Gemini 4 Argon is Google's new flagship after Gemini 3.1 Pro, and the company calls it the start of the "next era" of its models. In Google's announcement table it beats GPT-6 Astra and Claude Opus 5.5 on twelve tests of eighteen; independent measurements put it in the top five, but not first. Below: what the model can do, where it really is ahead, where it trails, what the benchmarks on the Kumo page show, and what it costs — $1.1 per million input tokens and $5.5 per million output tokens on Kumo.
Context
1M tokens
1,048,576 tokens; up to 32K tokens per response on Kumo
Price on Kumo
$1.1 / $5.5
input / output per 1M — 55% of Google's introductory price
Intelligence Index
52.6
fifth place; the leader, Claude Opus 5.5, scores 57.6
Coding agents
5 of 42
Antigravity CLI with Argon, best DeepSWE at 78.8%
What Gemini 4 Argon is
Google DeepMind introduced Gemini 4 Argon on September 30, 2026. It is a reasoning model built for long multi-step work: programming, documents and data, cybersecurity and long-form writing. Before it, Google's flagship was still Gemini 3.1 Pro: the promised Gemini 3.5 Pro never shipped, and all summer the company only moved its Flash line forward.
The context window is 1,048,576 tokens, and that million covers the prompt, the hidden reasoning and the answer together. Google raised the response limit from 64 thousand to a million tokens, but only in a special continuation mode; a plain request to Google's API returns up to 65,536 tokens. The model reads text, images, video, audio and PDF and reasons at three levels — low, medium and high.
Inside Google, Argon already works as an agent. The company says the model rewrote 32 thousand lines of SIMD code in the libgav1 video decoder into safe Rust — the new version runs 2.7 times faster than the previous Rust port with identical output — and is helping move the Zircon kernel of the Fuchsia operating system, over 800 thousand lines, to Rust. These are vendor claims, but they show well what the model is built for.
Where and when Gemini 4 Argon is available
The release came in stages. For the first two weeks only cyber defenders in the Fairwind Program and US government customers had it. Since October 13, Argon is open to Google AI Pro and AI Ultra subscribers in the Gemini app and in AI Mode in Search. Developers get it through the paid Gemini API.
Google's API has an introductory price: $2 per million input tokens and $10 per million output tokens, with cached reads 95% cheaper than input. After the introductory period the price doubles to $4 and $20. Google gave no exact date; according to TechTimes, the introductory price holds until the end of 2026.
On Kumo, Gemini 4 Argon is live since October 20 through the same OpenAI-compatible gateway and the same key as every other model in the catalogue. Two suppliers serve the model.
Google's benchmarks: where Argon is ahead
The announcement table has eighteen tests. By VentureBeat's count, Argon is first on twelve and shares first place on one more. The widest leads come where the model has to work with a lot of data for a long time: Harvey's legal agent (19.6% against 5.4% for GPT-6 Astra and 3.8% for Claude Opus 5.5), graph search in long context GraphWalks (84.2% against 71.8% and 66.8%), the Vals Finance Agent v2 (65.4% against 53.5% and 58.6%) and long-video understanding LVBench (91.7%).
In programming the picture is mixed. On DeepSWE v1.1, where the model solves long engineering tasks in real repositories, Argon sets a record — 77.9% against 74.2% for Opus 5.5 and 74.1% for GPT-6 Astra. On vulnerability remediation CWE-bench v1 it shares first place with Astra (68%). In agentic terminal work it trails: on Terminal-Bench 4.0 Opus 5.5 scores 66.4% against Argon's 57.4%, and on FrontierSWE v2 GPT-6 Astra leads by 10.5 points.
What independent measurements show
The vendor picks the tests in its own announcement, so we compared the same three models on Artificial Analysis and OpenRouter data from the Kumo benchmarks page. All runs use the top reasoning setting.
Argon clearly leads on the GPQA Diamond science questions (96.5%) and on the τ²-Bench Airline support agent with tools (78.7%). Claude Opus 5.5 is ahead on long-document reasoning AA-LCR, on scientific code SciCode and on the hardest questions of Humanity's Last Exam. On Terminal-Bench v4 the three run almost level, and Argon is last of the three at 57.1% — matching Google's own figure.
Place in the overall ranking
Artificial Analysis's Intelligence Index rolls about ten tests into one number. Argon scores 52.6 — fifth place, behind Claude Opus 5.5 (57.6), Claude Sonnet 5.5 (56.0), Claude Fable 5.1 (53.4) and GPT-6 Astra (52.7), with one tenth of a point between Argon and Astra. The headline "Google retakes the lead" holds for the announcement table, not for the independent index.
The jump within Google's own line is huge, though: the previous flagship Gemini 3.1 Pro scores 29.7, the strongest Flash model, Gemini 3.8 Flash, 40.9. Argon adds 23 and 12 points to them.
Argon as a coding agent
Artificial Analysis separately measures agent-plus-model pairs on three tests: DeepSWE v1.1, SWE-Atlas-QnA and Terminal-Bench v4. Antigravity CLI — Google's own agent — with Gemini 4 Argon is fifth of 42 with an index of 63.8. Only Claude Code with Sonnet 5.5 and Opus 5.5, Devin Fusion CLI and Muse Code are ahead.
The parts tell the same story as the announcement. On DeepSWE v1.1 this pair has the best result in the whole ranking — 78.8%. On SWE-Atlas-QnA (56.5%) and Terminal-Bench v4 (56.1%) it trails the leaders noticeably. An average task takes it about 35 minutes and costs about $5.8 at vendor prices — less than Claude Code with Opus 5.5 ($13) or Sonnet 5.5 ($14.2).
Makes things up less, writes a lot
In AA-Omniscience, Artificial Analysis checks how often a model confidently answers wrong when it could have said "I don't know". Argon's hallucination rate is 15%, the lowest among models scoring above 45 on the index; GPT-6 Astra's is 51%, GPT-6.1 Sol's 54%. The flip side is lower accuracy on the same test: 50% correct answers against 63% for Astra. Argon admits not knowing more often — for fact-finding and document work that is mostly a plus.
The second trait is verbosity. On the index run Argon spends about 62 thousand tokens per task on average, GPT-6 Astra about 27 thousand. So compare models by cost per task, not per million tokens: the Artificial Analysis index run cost Argon about $2 per task at Google's introductory prices, Astra $3.3, Claude Opus 5.5 $6. Mind the reasoning level too: every independent figure was taken at high, while Gemini subscribers get medium by default.
What Gemini 4 Argon costs
On Kumo, Argon costs $1.1 per million input tokens and $5.5 per million output tokens — 55% of Google's introductory price and a little over a quarter of what Google will charge after the introductory period. Cached reads are $0.055 per million. Cache writes are billed as regular input.
| Where | Input | Output | Cached read |
|---|---|---|---|
| Google, introductory | $2 | $10 | $0.1 |
| Google, after the introductory period | $4 | $20 | $0.2 |
| Kumo | $1.1 | $5.5 | $0.06 (0.055) |
In practice: a coding agent sends about 20 million input tokens a working day, 80% of them re-reads of the same context from cache, and gets about a million tokens back. On Kumo such a day costs about $10.8; at Google's introductory price about $19.6, after the introductory period about $39.2.
The model is also available in the Pro, Max and Ultra token packages: a package is charged the sum of input and output tokens times ×3.
Level against price: what to compare Argon with
At the same Kumo price, $5.5 per million output tokens, there are two more strong models: Claude Sonnet 5.5 with an index of 56.0 and GPT-6.1 Sol with 51.8. Argon sits between them. Claude Opus 5.5 costs twice as much and scores five points higher; GPT-6 Astra, at almost the same level, costs $30 per million output tokens on Kumo.
What Gemini 4 Argon is good for
The model's strengths are long tasks over a lot of data, factual accuracy and a well-protected agent. Try it first where those matter more than speed.
- Long engineering tasks in a real repository: changes from a description, cross-language migrations, vulnerability fixes — Argon has the best DeepSWE v1.1 both in the announcement and among independent agents.
- Analysis of large documents: legal and financial material, reports, spreadsheets — leading on Harvey Legal Agent and Vals Finance Agent v2.
- Agents that read other people's pages and files: a low prompt-injection success rate lowers the risk of the agent following someone else's instruction.
- Search and fact-checking, where an honest "I don't know" beats a confident invention.
- Science questions and calculations: 96.5% on GPQA Diamond, the best of the three flagships.
Connect Gemini 4 Argon in five minutes
- Sign up in the Kumo console and top up your balance.
- Create an API key and give it a spend limit right away: Argon reasons a lot, and a limit protects you from a surprise bill.
- Set the gateway address https://api.kumorouter.com/v1 in any OpenAI-compatible client or SDK.
- Use the model gemini-4-argon (full name google/gemini-4-argon) and send your first request.
- Check the console's request log to see the charge went to the right model.
curl -s https://api.kumorouter.com/v1/chat/completions \
-H "Authorization: Bearer $KUMO_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "gemini-4-argon",
"messages": [
{"role": "user", "content": "Find the vulnerability in this function and propose a fix: ..."}
],
"max_tokens": 8192
}'
Python example
The official OpenAI SDK works unchanged — just swap the address and the key. On Kumo, streaming and tool calling are confirmed for Argon over both Chat Completions and Responses.
from openai import OpenAI
client = OpenAI(
base_url="https://api.kumorouter.com/v1",
api_key="KUMO_API_KEY",
)
stream = client.chat.completions.create(
model="gemini-4-argon",
messages=[{"role": "user", "content": "Compare these two contracts and list the differences: ..."}],
stream=True,
)
for chunk in stream:
print(chunk.choices[0].delta.content or "", end="")
Argon in Cursor, Cline and other editors
In any tool with an "OpenAI Compatible" setting Argon connects the same way: the address https://api.kumorouter.com/v1, your Kumo key and the model name gemini-4-argon. That is how it works in Cline, Roo Code, Continue, OpenCode and in Cursor with your own key. Step-by-step setups are in the guides for Cline and VS Code, OpenCode and Cursor.
Claude Code speaks the Anthropic protocol, so use Claude models with it — the setup is in the Claude Code guide.
FAQ
- How much does Gemini 4 Argon cost via API?
- On Kumo — $1.1 per million input tokens and $5.5 per million output tokens. Google's introductory price is $2 and $10, after the introductory period $4 and $20. Current rates are always on the models page.
- Is Gemini 4 Argon better than Claude Opus 5.5 and GPT-6 Astra?
- Depends on what for. In Google's table Argon leads most tests, especially long tasks with documents, video and data. On the Artificial Analysis composite index it is fifth: Opus 5.5 is five points higher, Astra one tenth. In terminal work Argon trails both.
- How is Argon different from Gemini 3.8 Flash?
- Argon is the flagship for hard and long tasks: an index of 52.6 against 40.9 for 3.8 Flash and much better agentic terminal work. Gemini 3.8 Flash is faster and 2.7 times cheaper on output on Kumo — take it for high-volume tasks of medium difficulty.
- What is Gemini 4 Argon's context window?
- 1,048,576 tokens for the prompt, reasoning and answer together. On Kumo one response is capped at 32,768 tokens.
- Is it safe to send code to Gemini 4 Argon?
- Requests go through the Kumo gateway to the model's supplier. How we handle data is described on the security page. For secrets, keys and personal data the general rule applies: don't send them to any cloud model.
Sources
- Google's announcement: Gemini 4 Argon, September 30, 2026.
- The full announcement benchmark table and lead count: VentureBeat, September 30, 2026.
- Subscriber rollout, reasoning levels, AA-Omniscience and introductory price timing: TechTimes, October 15, 2026.
- Independent benchmarks, indices and coding agents: Kumo benchmarks page — Artificial Analysis and OpenRouter data, October 20, 2026.
- Model prices and specs on Kumo: Kumo models page, price list v28 of October 20, 2026.
- Gemini API pricing: ai.google.dev.