Gemini 4 Argon

Google's flagship model for coding, legal and finance work, math and science, with a 1M-⁠token context window

About Gemini 4 Argon

Gemini 4 Argon is Google's flagship model for software engineering, legal and finance work, math and science. It reads text, images, video, audio and PDF files, takes up to 1 million tokens in one conversation and thinks before it answers

Gemini 4 Argon scores 77.9% on DeepSWE v1.1, ahead of Claude Opus 5.5 at 74.2% and GPT-⁠6 Astra at 74.1%, and has the top results on AutomationBench, Vals Finance Agent v2 and Harvey's Legal Agent Benchmark among the models it is compared with. It also leads on LVBench, at 91.7%, for questions about long videos

On Overchat AI, Gemini 4 Argon shares one chat with Claude, GPT-⁠6 and Grok, so you can send the same question to another model without leaving the conversation

Developer
Google
Released
Sep 30, 2026
Context window
1M tokens
Max output
943,718 tokens
Input
Text, images and files
Output
Text
Reasoning
Yes, with adjustable effort
Intelligence Index
50, #8 of 54 models
License
Proprietary
Available in
Gemini app, Gemini API, Google AI Studio, Google Cloud, Antigravity, OpenRouter, Overchat AI
Modalities
Intelligence Index?
50
Context window?
1M tokens
Released
Sep 30, 2026

Gemini 4 Argon vs other models

Gemini 4 Argon ranks #20 on the Artificial Analysis Intelligence Index out of 54 models

Artificial Analysis Intelligence Index50#8 of 54 modelsBetter than 62% of 54 models
  1. 65Claude Fable 5.2
  2. 60Gemini 4 Carbon
  3. 59Claude Opus 5.5
  4. 57GPT-6 Astra
  5. 55GPT-6.1 Sol
  6. 50Gemini 4 Argon
  7. 50Muse Wonder Full
  8. 46Kimi K3.1 Pro
  9. 44Muse Wonder
  10. 35Grok 4.7
  • #1of 63 modelsAA-Omniscience Non-Hallucination RateHow often the model says it doesn't know instead of making up an answer95.7%
  • #4of 344 modelsHumanity's Last ExamExpert-written questions across many subjects57.1%
  • #4of 136 modelsSciCodePython code for scientific computing61.8%
  • #6of 130 modelsTerminal-Bench 4.0Hard tasks an agent completes in a terminal57.1%

Source: Artificial Analysis

Compare Gemini 4 Argon

Gemini 4 Argon vs GPT-6 Astra

Gemini 4 Argon scores higher than GPT-⁠6 Astra in 13 of 17 benchmarks. The widest lead is in Harvey's Legal Agent Benchmark: 19.6% against 5.4%. GPT-⁠6 Astra is ahead in FrontierSWE v2: 65.5% against 55.0%. Gemini 4 Argon writes 91 tokens per second, GPT-⁠6 Astra writes 48

Benchmarks won
Gemini 4 Argon: 13
GPT-6 Astra: 4
Average score
Gemini 4 Argon: 68.8%
GPT-6 Astra: 65.7%
Output speed, tokens/s
Gemini 4 Argon: 91
GPT-6 Astra: 48
Time to first token, s (log scale)
Gemini 4 Argon: 6.0
GPT-6 Astra: 3.2
GPT-6 Astra aheadGemini 4 Argon ahead
  • Harvey's Legal Agent Benchmark+14.2
  • GraphWalks, 256K to 1M+12.4
  • Vals Finance Agent v2+11.9
  • AutomationBench+9.9
  • Vals Index+5.8
  • LVBench+4.2
  • RiemannBench+4.0
  • DeepSWE v1.1+3.8
  • LABBench 2+3.4
  • Vibe Code Bench+2.3
  • PostTrainBench+1.0
  • GraphWalks, up to 128K+1.0
  • Chartography+0.6
  • CWE-bench v10.0
  • Terminal-Bench 4.0−0.8
  • FrontierSWE v2−10.5
  • Terminal-Bench Science−10.5

Gemini 4 Argon vs Claude Fable 5.1

Gemini 4 Argon scores higher than Claude Fable 5.1 in 15 of 17 benchmarks. The widest lead is in Chartography: 71.6% against 46.2%. Claude Fable 5.1 is ahead in FrontierSWE v2: 56.3% against 55.0%

Benchmarks won
Gemini 4 Argon: 15
Claude Fable 5.1: 2
Average score
Gemini 4 Argon: 68.8%
Claude Fable 5.1: 58.9%
Claude Fable 5.1 aheadGemini 4 Argon ahead
  • Chartography+25.4
  • LABBench 2+20.2
  • AutomationBench+19.9
  • GraphWalks, 256K to 1M+19.2
  • Harvey's Legal Agent Benchmark+12.9
  • LVBench+12.0
  • DeepSWE v1.1+10.5
  • RiemannBench+10.4
  • CWE-bench v1+10.0
  • GraphWalks, up to 128K+8.3
  • Vals Finance Agent v2+6.5
  • PostTrainBench+5.1
  • Terminal-Bench Science+5.0
  • Vals Index+3.1
  • Vibe Code Bench+1.6
  • Terminal-Bench 4.0−0.5
  • FrontierSWE v2−1.3

Gemini 4 Argon vs Claude Opus 5.5

Gemini 4 Argon scores higher than Claude Opus 5.5 in 13 of 17 benchmarks. The widest lead is in GraphWalks, 256K to 1M: 84.2% against 66.8%. Claude Opus 5.5 is ahead in Terminal-Bench 4.0: 66.4% against 57.4%. Gemini 4 Argon writes 91 tokens per second, Claude Opus 5.5 writes 88

Benchmarks won
Gemini 4 Argon: 13
Claude Opus 5.5: 4
Average score
Gemini 4 Argon: 68.8%
Claude Opus 5.5: 64.4%
Output speed, tokens/s
Gemini 4 Argon: 91
Claude Opus 5.5: 88
Time to first token, s (log scale)
Gemini 4 Argon: 6.0
Claude Opus 5.5: 1.9
Claude Opus 5.5 aheadGemini 4 Argon ahead
  • GraphWalks, 256K to 1M+17.4
  • Harvey's Legal Agent Benchmark+15.8
  • LABBench 2+15.7
  • GraphWalks, up to 128K+9.1
  • AutomationBench+8.8
  • LVBench+8.0
  • Vals Finance Agent v2+6.8
  • RiemannBench+6.4
  • Chartography+5.3
  • DeepSWE v1.1+3.7
  • Vals Index+1.9
  • Vibe Code Bench+1.6
  • CWE-bench v1+1.0
  • PostTrainBench−4.0
  • Terminal-Bench Science−5.7
  • FrontierSWE v2−7.3
  • Terminal-Bench 4.0−9.0

Best model by use case

Average benchmark scores of Gemini 4 Argon and 3 other models by use case

  • Knowledge work

    ?
    Average of 4 tests
    1. 1Gemini 4 Argon51.3%
    2. 2Claude Opus 5.543.0%
    3. 3GPT-6 Astra40.9%
    4. 4Claude Fable 5.140.7%
  • Coding

    ?
    Average of 6 tests
    1. 1Claude Opus 5.568.3%
    2. 2GPT-6 Astra66.6%
    3. 3Gemini 4 Argon65.9%
    4. 4Claude Fable 5.161.7%
  • Reasoning

    ?
    Average of 5 tests
    1. 1Gemini 4 Argon81.3%
    2. 2GPT-6 Astra79.2%
    3. 3Claude Opus 5.572.7%
    4. 4Claude Fable 5.168.6%
  • Vision

    ?
    Average of 2 tests
    1. 1Gemini 4 Argon81.7%
    2. 2GPT-6 Astra79.3%
    3. 3Claude Opus 5.575.0%
    4. 4Claude Fable 5.163.0%

Gemini 4 Argon benchmarks

Compare Gemini 4 Argon with GPT-⁠6 Astra, Claude Fable 5.1 and Claude Opus 5.5 in 17 benchmarks measuring knowledge work, coding, reasoning and vision

Order

Knowledge work

  • Vals Index?Gemini 4 Argon68.9%GPT-6 Astra63.1%Claude Fable 5.165.8%Claude Opus 5.567.0%68.9%+1.9
  • AutomationBench?Gemini 4 Argon51.3%GPT-6 Astra41.4%Claude Fable 5.131.4%Claude Opus 5.542.5%51.3%+8.8
  • Vals Finance Agent v2?Gemini 4 Argon65.4%GPT-6 Astra53.5%Claude Fable 5.158.9%Claude Opus 5.558.6%65.4%+6.5
  • Harvey's Legal Agent Benchmark?Gemini 4 Argon19.6%GPT-6 Astra5.4%Claude Fable 5.16.7%Claude Opus 5.53.8%19.6%+12.9

Coding

  • DeepSWE v1.1?Gemini 4 Argon77.9%GPT-6 Astra74.1%Claude Fable 5.167.4%Claude Opus 5.574.2%77.9%+3.7
  • FrontierSWE v2?Gemini 4 Argon55.0%GPT-6 Astra65.5%Claude Fable 5.156.3%Claude Opus 5.562.3%55.0%−10.5
  • Vibe Code Bench?Gemini 4 Argon91.9%GPT-6 Astra89.6%Claude Fable 5.190.3%Claude Opus 5.590.3%91.9%+1.6
  • Terminal-Bench 4.0?Gemini 4 Argon57.4%GPT-6 Astra58.2%Claude Fable 5.157.9%Claude Opus 5.566.4%57.4%−9.0
  • PostTrainBench?Gemini 4 Argon45.3%GPT-6 Astra44.3%Claude Fable 5.140.2%Claude Opus 5.549.3%45.3%−4.0
  • CWE-bench v1?Gemini 4 Argon68.0%GPT-6 Astra68.0%Claude Fable 5.158.0%Claude Opus 5.567.0%68.0%+0.0

Reasoning

  • Terminal-Bench Science?Gemini 4 Argon57.6%GPT-6 Astra68.1%Claude Fable 5.152.6%Claude Opus 5.563.3%57.6%−10.5
  • LABBench 2?Gemini 4 Argon88.8%GPT-6 Astra85.4%Claude Fable 5.168.6%Claude Opus 5.573.1%88.8%+3.4
  • RiemannBench?Gemini 4 Argon76.0%GPT-6 Astra72.0%Claude Fable 5.165.6%Claude Opus 5.569.6%76.0%+4.0
  • GraphWalks, up to 128K?Gemini 4 Argon99.7%GPT-6 Astra98.7%Claude Fable 5.191.4%Claude Opus 5.590.6%99.7%+1.0
  • GraphWalks, 256K to 1M?Gemini 4 Argon84.2%GPT-6 Astra71.8%Claude Fable 5.165.0%Claude Opus 5.566.8%84.2%+12.4

Vision

  • Chartography?Gemini 4 Argon71.6%GPT-6 Astra71.0%Claude Fable 5.146.2%Claude Opus 5.566.3%71.6%+0.6
  • LVBench?Gemini 4 Argon91.7%GPT-6 Astra87.5%Claude Fable 5.179.7%Claude Opus 5.583.7%91.7%+4.2

Sources: Google, Gemini 4 Argon launch post (Sep 30, 2026), Google DeepMind, Gemini 4 Argon model evaluation

Gemini 4 Argon speed

Output speed is how fast the model writes (tokens per second, higher is better). Time to first token is how long you wait before the answer starts (lower is better)

Output speed and time to first token of 35 models

?Percentile
Time to first token, s (log scale)
Output speed, tokens/s (log scale)
Output speed and time to first token of 35 models
ModelsOutput speed, tokens/s (log scale)Time to first token, s (log scale)
GPT-6 Luna1432.4
GPT-6.1 Terra754.3
GPT-6.1 Sol582.9
GPT-6 Astra483.2
Claude Opus 5.5881.9
Claude Opus 5542.6
Claude Opus 4.8591.3
Claude Fable 5.2842.3
Claude Sonnet 5.51021
Claude 5 Sonnet21510.5
Claude Sonnet 4.622010.8
Claude Haiku 5.51672.1
Gemini 4 Carbon16929.5
Gemini 4 Argon916
Gemini 3.8 Flash2037.5
Gemini 3 Pro27011.2
Grok 4.7818.9
Grok 4.629032
Qwen 3.8 Max8130.6
DeepSeek V4 Pro1441.1
DeepSeek V4.1 Flash4222.2
Mistral Large 42268.6
Hy421437.7
ERNIE 5.5 Flash4052.7
Muse Wonder9808.7
Muse Spark11296.2
Kimi K3.1 Pro3407.4
Kimi K33385.9
GLM 5.31971.2
Flint19412.3
Llama 4810.7
Claude 4.5 Haiku1620.9
MiniMax M31694.5
MiniMax M3.1 Flash6567.2
Kimi K2.7 Code4553.6

Typing test

Every model writes the same answer. The test waits the model's time to first token, then prints the text at the model's output speed and shows the total time

Answer length
  • Gemini 4 Argon91 tokens/s6.0 s to first token
  • GPT-6 Astra48 tokens/s3.2 s to first token
  • Claude Opus 5.588 tokens/s1.9 s to first token

Source: Overchat AI, 2026-11-11

See how Gemini 4 Argon answers

Examples of Gemini 4 Argon answers to 3 prompts: Analyze bank data, Review a contract and Solve a math problem. Each answer is the unedited output of the model

Write a Python script that reads a CSV of bank transactions (date, description, category, amount), groups spending by month and category and prints the top three categories for each month. Keep it under 40 lines.

Gemini 4 Argon

Here is a complete, dependency-free Python script that achieves this in under 30 lines using the built-in csv and collections modules.

It assumes dates are in YYYY-MM-DD format and uses abs() on the amount so it works regardless of whether spending is listed as positive or negative numbers.

python
from collections import Counter, defaultdict
import csv


def top_spending_categories(csv_file_path):
    # Dictionary mapping each 'YYYY-MM' to a Counter of categories
    monthly_spending = defaultdict(Counter)

    with open(csv_file_path, mode="r", newline="", encoding="utf-8") as file:
        reader = csv.DictReader(file)
        for row in reader:
            # Extract 'YYYY-MM' from 'YYYY-MM-DD'
            month = row["date"][:7]
            category = row["category"]
            amount = abs(float(row["amount"]))

            monthly_spending[month][category] += amount

    # Sort chronologically by month and print the top 3 categories
    for month in sorted(monthly_spending.keys()):
        print(f"--- {month} ---")
        for category, total in monthly_spending[month].most_common(3):
            print(f"  {category}: ${total:,.2f}")


if __name__ == "__main__":
    # Replace 'transactions.csv' with your actual file path
    top_spending_categories("transactions.csv")

Expected CSV Format (transactions.csv)

csv
date,description,category,amount
2023-10-02,Grocery Store,Food,150.25
2023-10-05,Gas Station,Transport,45.00
2023-10-12,Apartment Rent,Housing,1200.00
2023-10-18,Restaurant,Food,85.50
2023-10-25,Electric Bill,Utilities,95.00

Example Output

text
--- 2023-10 ---
  Housing: $1,200.00
  Food: $235.75
  Utilities: $95.00

Review this clause from a freelance contract: "The Contractor shall indemnify and hold harmless the Client against any and all claims, losses and damages arising from the Services, without limitation." What are the risks for me as the contractor, and how would you rewrite it in a fairer way?

Gemini 4 Argon

The clause you shared is what lawyers often call a "nuclear" indemnification clause. It is heavily one-sided and exposes you to massive financial risk.

Here is a breakdown of why this clause is dangerous for you as the contractor, followed by fairer ways to rewrite it.


The Risks for You (The Contractor)

  1. "Without limitation" (Uncapped Liability): This is the biggest red flag. If you are getting paid $2,000 for a project, and something goes wrong that costs the client $500,000, you are legally on the hook for the entire $500,000. Your financial risk is completely disproportionate to your reward.
  2. "Arising from the Services" (No Fault Required): Notice that the clause doesn't say "arising from the Contractor's negligence or breach of contract." It just says "arising from the Services." If you follow the client’s exact instructions, or use materials they provided, and a third party sues the client as a result, you could still be forced to pay for the client's legal defense and damages—even though you did nothing wrong.
  3. "Any and all losses and damages" (Includes Consequential Damages): This broad phrasing includes indirect damages, such as "lost profits" or "loss of business reputation." For example, if you are building a website, and a bug causes it to go down for three hours, the client could claim they lost $100,000 in potential sales during that window and demand you pay it.
  4. First-Party vs. Third-Party Claims: Indemnification is primarily meant to protect the client if a third party (like one of their customers, or a copyright troll) sues them because of something you did. Because this clause isn't restricted to "third-party claims," the client can use it to force you to pay for their internal losses and their own attorney's fees in a direct dispute with you.
  5. Uninsurable Risk: If you have Professional Liability (Errors & Omissions) insurance, your policy will generally only cover claims resulting from your actual negligence. By signing a contract where you agree to cover "any and all claims... without limitation," you are taking on liability that your insurance company will likely refuse to cover.

How to Rewrite It (From Most Fair to Compromise)

To make this fair, you need to narrow the trigger (only things that are actually your fault), limit the scope (third-party claims only), and cap the financial damage (usually to the amount they paid you).

Option 1: The Fair & Balanced Standard (Recommended)

This version limits your liability to third-party claims caused by your actual bad acts or IP infringement, makes the protection mutual (so they have to protect you if their materials get you sued), and adds a financial cap.

Indemnification: "The Contractor shall indemnify and hold harmless the Client against any third-party claims, losses, and direct damages to the extent directly caused by the Contractor’s gross negligence, willful misconduct, or infringement of third-party intellectual property rights in performing the Services. Conversely, the Client shall indemnify and hold harmless the Contractor against any third-party claims arising from materials, instructions, or intellectual property provided by the Client."

Limitation of Liability (Add as a separate sentence/clause): "Notwithstanding anything to the contrary, neither party shall be liable for any indirect, incidental, special, or consequential damages (including lost profits). The Contractor’s total aggregate liability under this Agreement, including its indemnification obligations, shall not exceed the total fees paid by the Client to the Contractor under this Agreement."

Option 2: The Short & Simple Revision (Minimal Redlines)

If the client dislikes heavy legal edits and you just want to tweak their exact sentence to make it safe, change it to this:

"The Contractor shall indemnify and hold harmless the Client against any third-party claims and direct damages solely to the extent caused by the Contractor's gross negligence or willful misconduct in performing the Services, up to a maximum amount equal to the total fees paid to the Contractor under this Agreement."


How to Explain the Change to Your Client

Clients often pull templates off the internet without realizing how aggressive the language is. When sending back your edits, you can keep the tone collaborative with a note like this:

“I reviewed the agreement and everything looks great, with the exception of the indemnification clause. As currently written, it carries uncapped liability even for things outside of my control, which my business insurance won't allow me to sign. I’ve updated the language to standard industry terms—making sure you are fully protected if I am negligent or infringe on anyone's IP, while capping the overall liability at the total value of the contract. Let me know if this works for you!”

A ball is dropped from 20 m and each bounce takes it back up to 60% of the height it fell from. What total distance does it travel before it comes to rest? Show the steps.

Gemini 4 Argon

To find the total distance the ball travels before coming to rest, we need to calculate the initial drop distance and then add the distances of all the subsequent bounces (both going up and coming down).

Here is the step-by-step breakdown:

Step 1: Visualize the motion

  1. Initial Drop: The ball falls 20 m to hit the ground the first time.
  2. First Bounce: It rebounds to 60% of 20 m, which is 20 × 0.6 = 12 m going up, and then falls another 12 m going down.
  3. Second Bounce: It rebounds to 60% of 12 m, which is 12 × 0.6 = 7.2 m going up, and then falls 7.2 m going down.
  4. Subsequent Bounces: This pattern continues infinitely, with each bounce height being 0.6 times the previous height.

Step 2: Set up the mathematical series

We can separate the total distance (D) into the initial drop plus twice the height of each bounce (since the ball travels the same distance up and down on every bounce):

D = 20 + 2(20 × 0.6) + 2(20 × 0.6²) + 2(20 × 0.6³) + …

Let's simplify the terms inside the parentheses:

  • 1st bounce height = 12 m
  • 2nd bounce height = 7.2 m
  • 3rd bounce height = 4.32 m

So the equation looks like this: D = 20 + 2(12 + 7.2 + 4.32 + …)


Step 3: Sum the infinite geometric series

The numbers inside the parentheses (12 + 7.2 + 4.32 + …) form an infinite geometric series, where:

  • The first term (a) = 12
  • The common ratio (r) = 0.6 (or 60%)

The formula for the sum (S) of an infinite geometric series (when |r| < 1) is: S = a / (1 − r)

Plug in our values for a and r: S = 12 / (1 − 0.6)
S = 12 / 0.4
S = 30 m

(Note: 30 m is the total distance the ball travels strictly moving upward).


Step 4: Calculate the total distance

Since the ball travels 30 m going up, it must also travel 30 m coming down from those bounces. Now, substitute S = 30 back into our total distance equation from Step 2:

D = 20 + 2(S)
D = 20 + 2(30)
D = 20 + 60
D = 80 m

Answer:

The ball travels a total distance of 80 meters before coming to rest.

More prompts to try with Gemini 4 Argon

What Gemini 4 Argon is best for

  • Software engineering

    Gemini 4 Argon fixes bugs and builds features across a real codebase. It scores 77.9% on DeepSWE v1.1 and 91.9% on Vibe Code Bench, the top results among the models it is compared with

  • Legal and finance work

    Gemini 4 Argon reviews contracts, drafts legal documents and runs financial research. It scores 19.6% on Harvey's Legal Agent Benchmark, against 3.8% for Claude Opus 5.5, and 65.4% on Vals Finance Agent v2

  • Charts, scans and videos

    Gemini 4 Argon reads charts, screenshots, scanned pages and long videos. It scores 71.6% on Chartography and 91.7% on LVBench, the best results among the models it is compared with

  • Math and science

    Gemini 4 Argon solves research-level math and biology lab tasks, with 76.0% on RiemannBench and 88.8% on LABBench 2, ahead of GPT-⁠6 Astra and Claude Opus 5.5

How to use Gemini 4 Argon on Overchat AI

Start in the chat at the top of this page or in the Overchat AI app on the web, iOS or Android

  1. The Models menu with Gemini 4 Argon picked
    01

    Pick Gemini 4 Argon

    It is already picked in the chat on this page. In the Overchat AI app, open the model list next to the message field and pick Gemini 4 Argon

  2. Message field with a file attached and a request
    02

    Attach your code or a file

    Add code, a document, a spreadsheet or a screenshot, and write what you need: the result, the format and what must stay the same

  3. Gemini 4 Argon's answer and a follow-up
    03

    Keep going in the same chat

    Ask a follow-up, ask for tests or a shorter version, or switch to another model without losing the conversation

Why use Gemini 4 Argon on Overchat AI

One account for chat, images and video, on every device

Superchat picked over Gemini 4 Argon: one request done by Claude Opus 5.5 (the plan) and Nano Banana 2.1 (the poster)

Superchat AI agent

Superchat Agent automatically routes your message to the model best suited for the task. It can generate images and videos, use the AI tools built into Overchat AI, and learn from your interactions to get better over time

Message field with a PDF, a spreadsheet and a dashboard screenshot attached, Gemini 4 Argon picked and a voice message being recorded

Chat with images and files

Upload PDFs, spreadsheets, code and screenshots and ask Gemini 4 Argon about them. You can also dictate a message instead of typing it

An image made with Nano Banana 2.1 and a surfing video made with Seedance 2.5

Go beyond text

Generate images with Nano Banana and GPT Image and videos with Seedance and Kling in the same account, with the same credits as chat

Overchat AI on the web and iPhone, and the Superchat screen of the iPhone app

On the web and iPhone

Start a chat on your laptop and finish it on your phone. The Overchat AI app has the same models, and your chats stay in one account on every device

Download on the App Store

Gemini models on Overchat AI

Compare 4 Gemini models on Overchat AI by output speed, time to first token and the price of one message

Models
Output speed, tokens/s
First token, s
Credits per message?
169
29.5
1
Gemini 4 ArgonThis page
91
6.0
1
Gemini 3.8 Flash
203
7.5
30
Gemini 3 Pro
270
11.2
1

Source: Overchat AI, 2026-11-11

Explore other models

Models marked Soon are coming to Overchat AI; the rest are available now

Learn more about AI models

All articles

Ready to try Gemini 4 Argon?

Ask your first question or attach a file, and Gemini 4 Argon answers in the Overchat AI chat

Chat with Gemini 4 Argon

Gemini 4 Argon FAQ

If you have any questions,
reach out to us at hello@overchat.ai

Gemini 4 Argon is Google's flagship AI model for software engineering, legal and finance work, math and science. It reads text, images, video, audio and PDF files and takes up to 1 million tokens in one conversation

Google announced Gemini 4 Argon on September 30, 2026, and it became available to everyone on October 13, 2026: in the Gemini app on Google AI Pro and Ultra, in the Gemini API and in Antigravity

Gemini 4 Carbon is Google's model for the hardest coding and agent tasks: it spends longer checking its work and scores 60 on the Artificial Analysis Intelligence Index against 50 for Argon. Gemini 4 Argon starts answering in a few seconds, while Carbon often thinks for half a minute first, so Argon is the everyday pick. On Overchat AI both are in the same chat

At legal, finance and long-video work, yes: Gemini 4 Argon scores 19.6% on Harvey's Legal Agent Benchmark against 3.8% for Opus 5.5, 65.4% on Vals Finance Agent v2 against 58.6% and 91.7% on LVBench against 83.7%. Opus 5.5 is stronger in a terminal and on long engineering projects, with 66.4% on Terminal-Bench 4.0 against 57.4% and 62.3% on FrontierSWE v2 against 55.0%

Yes. Gemini 4 Argon reads video, audio, images and PDF files along with text, and answers questions about any part of them. It scores 91.7% on LVBench, a test of questions about long videos

Gemini 4 Argon reads up to 1 million tokens in one conversation, about 750,000 words, and writes up to about 940,000 tokens in one answer

Type your message in the chat at the top of this page, attach a file if you need to and press Send. After you sign in, Gemini 4 Argon answers in the Overchat AI chat, where you can keep the conversation going on the web, iOS or Android and switch to another model at any time