Gemini 4 Argon for Coding: Test Results and Benchmarks

👁 114 reads

Gemini 4 Argon passed all 51 unit tests across my four coding tasks. It cost $0.028 and took 25 seconds in total. No other flagship model I tested did the job more cheaply or quickly. Coding benchmarks, however, paint a more mixed picture: Google says Argon leads on DeepSWE but trails on FrontierSWE and Terminal-Bench 4.0, while it ranks eighth in the independent Arena Code Arena: WebDev leaderboard. Below are my test results, the benchmark figures and the coding tools where you can use Argon. For a broader look at the model, see the Gemini 4 Argon guide.

Four tasks, 51 tests

The tasks were short, but each included a trap that can trip up a model. The tests used Polish-specific material, including Polish tax ID numbers, public holidays and currency wording; they were run in Polish, not retested in English.

• Polish tax ID validation: handle hyphens, spaces and the PL prefix; apply the weighted checksum; reject a checksum remainder of 10.
• Merge intervals: merge intervals whose end and start points touch, and reverse intervals written in descending order, such as (5, 2).
• Polish working days: cover 2024-2030, movable holidays and Christmas Eve, which became a day off only from 2025.
• Writing an invoice amount in words: use the correct Polish forms for the currency and its subunit, round to the nearest grosz, and handle amounts up to one billion.

What Argon’s code looks like

Argon returned just the code, without an introduction, using type annotations and comments in Polish. The most interesting part was the holiday list in the working-days task: Gemini 3.8 Flash got this wrong the day before.

def get_holidays(year: int) -> set[datetime.date]:
    easter = calculate_easter(year)
    holidays = {
        datetime.date(year, 1, 1),   # New Year's Day
        datetime.date(year, 1, 6),   # Epiphany
        easter,                      # Easter Sunday
        easter + datetime.timedelta(days=1),   # Easter Monday
        datetime.date(year, 5, 1),   # Labor Day
        datetime.date(year, 5, 3),   # Constitution Day
        easter + datetime.timedelta(days=49),  # Pentecost Sunday
        easter + datetime.timedelta(days=60),  # Corpus Christi
        datetime.date(year, 8, 15),  # Assumption of Mary
        datetime.date(year, 11, 1),  # All Saints' Day
        datetime.date(year, 11, 11), # Independence Day
        datetime.date(year, 12, 25), # Christmas Day
        datetime.date(year, 12, 26), # Second Day of Christmas
    }
    # Under the law in force as of 2026 (a 2024 act),
    # Christmas Eve has been a statutory day off since 2025.
    if year >= 2025:
        holidays.add(datetime.date(year, 12, 24))
    return holidays

The Easter date is calculated with the Meeus/Jones/Butcher algorithm, using only the standard library, as requested. Gemini 3.8 Flash treated Christmas Eve as a day off in every year and failed the December 2024 test. For the amount-in-words task, Argon also added a ValueError for out-of-range amounts and correctly rendered 12.13 in Polish words as “twelve units of the currency, thirteen subunits” - a translation of the tested Polish output. In the original Polish test, Muse Wonder produced a nonexistent word instead.

Argon vs. rivals on the same tasks

Among the models with no errors, Argon cost almost three times less than Opus 5.5 and 4.6 times less than GPT-6 Astra. These were small functions, each only a few dozen lines long. The test shows whether a model can handle Polish-specific requirements without making basic mistakes, but it does not show how it will perform on a large repository, across multiple files or during a long-running terminal agent task. That is where benchmarks come in.

Google’s coding benchmarks

In a table dated September 30, Google reports Argon’s results at its highest thinking setting. Its performance across three coding benchmarks is mixed:

In the same table, Claude Fable 5.1 scores 67.4% on DeepSWE, 56.3% on FrontierSWE and 57.9% on Terminal-Bench. Across Google’s full table, Argon leads on 13 of 19 tests. But it trails its competitors on two of the three coding benchmarks, and is 9 percentage points behind Opus 5.5 on Terminal-Bench 4.0.

What independent benchmarks show

• Artificial Analysis, Terminal-Bench 4.0: Argon scores 57%, compared with 64% for Claude Sonnet 5.5, 60% for Claude Opus 5.5 and 59% for GPT-6 Astra. This independent result also puts Argon behind Anthropic and OpenAI models on terminal work.
• Arena, Code Arena: WebDev: Argon ranks eighth with 1,679 points. By contrast, the same Argon model at High is first in Text Arena with 1,525 points, a preliminary result.

Bloomberg: Google employees reportedly had doubts

On September 30, Bloomberg, as cited by Techmeme, reported that Google employees thought the model performed worse on some real-world coding tasks than its benchmarks suggested. Google disputed that account. My small test does not settle the question: Argon handled the short functions without a problem. But the independent Terminal-Bench result and its eighth-place finish in WebDev are more consistent with Bloomberg’s account than with the strongest figures in Google’s table.

Where you can code with Argon

• Google Antigravity (desktop app and CLI): Argon is available only on the Google AI Ultra plan. The Google AI Pro plan includes Antigravity but not Argon. Researchers can get the highest Antigravity tier free for six months. For setup instructions, see How to get started with Antigravity.
• Gemini API: use the model ID gemini-4-argon; it is available to any paid account. Through November 30, 2026, pricing is $2/$10 per million tokens; from December 1, it is $4/$20. The combined input and output limit is 1 million tokens. According to 9to5Google, using Argon in the AI Studio interface and in Android Studio (Canary) requires Ultra.
• OpenRouter: use google/gemini-4-argon. Pricing is the same, with a context window of 1,048,576 tokens. I ran my test through OpenRouter, where Argon was accessible from Poland on October 13, 2026. Availability may vary by country, so check with the provider. OpenRouter also works with any tool that accepts an OpenRouter API key.

Verdict

For individual functions, scripts and code fixes, Argon is the cheapest flagship model I have tested so far, and it made no mistakes. For long-running agent work in a repository or terminal, the benchmarks, including independent results, point more toward Claude Opus 5.5. If you code through an API, try Argon on your own code before November 30, while it costs half as much as its later price. For other options, see the AI coding tools guide.

// read moreGemini 4 Argon: Pricing, Benchmarks, and Hands-on TestNewsGemini 4 Argon vs GPT-6 Astra vs Claude Opus 5.5 BenchmarksBenchmarksGPT-6 Astra for Coding: What the Benchmarks Show (and Hide)Benchmarks