Gemini 4 Argon's Coding Scores Defy a Single Ranking

Google released Gemini 4 Argon on September 30, and its coding results offer a useful warning to anyone shopping for an AI assistant: the winner depends on what you ask it to do. In four reported tests of coding agents, Argon ranks first twice and fourth among four models twice. An agent is a model set up to work through a task in steps, rather than answer a single prompt. If your job involves building an app from a description, fixing difficult software problems or working in a terminal, those differences are considerably more useful than one giant trophy marked best at code.

Argon's arrival followed a wait. Google had promised Gemini 3.5 Pro for June, then spent the summer releasing smaller Flash models before introducing Argon. That sequence explains some of the interest in this flagship release: it put Google's claims about ambitious, multi-step work back in view after a summer focused on smaller models. It doesn't tell us that Argon is a renamed 3.5 Pro, or how much any one developer's workflow has improved. For that, the tasks need a closer look.

Four Coding Tests, Four Different Jobs

The reported coding results put Argon at 77.9% on DeepSWE v1.1, a test of longer software-engineering tasks, and 91.9% on Vibe Code Bench, which tests building apps from plain-language requests. It leads the four listed models on both. On FrontierSWE v2, a test of hard software-engineering problems, Argon scores 55.0%, behind GPT-6 Astra's 65.5%. On Terminal-Bench 4.0, which tests work in a command-line terminal, Argon scores 57.4%, behind Claude Opus 5.5's 66.4%. It ranks fourth among the four listed models on each of those last two tests.

The strongest case for Argon is substantial: it leads tests of extended engineering work and of turning an idea into an app. A builder who mostly describes an application and asks a model to make it could reasonably start with the Vibe Code Bench result. Someone whose agent spends its day navigating terminal tasks has a reason to scrutinize Terminal-Bench instead. “Coding” covers both jobs, much as “cooking” covers making a sandwich and running a dinner service. One score won't tell you who can handle your kitchen.

The comparisons have a measurement wrinkle. The four rows come from a table of Google-published results with different score sources: a third party supplied the Vibe Code Bench and FrontierSWE results, while Google ran Argon for DeepSWE and Terminal-Bench and drew rival scores from elsewhere. Even a shared public leaderboard doesn't guarantee every model used matching settings. Treat the standings as leads for choosing what to test, rather than a controlled contest in which only the model changed.

It shouldn't erase Argon's wins simply because the scores appear in a vendor's table. Nor should a first-place finish on DeepSWE settle a question about terminal work when the reported Terminal-Bench row points the other way. Google also has encouraging reported results beyond coding: Argon leads all four knowledge-work rows in that table, with scores supplied by outside parties. Those tasks provide another reason to pay attention to the model, while measuring a different set of jobs from a developer's codebase.

The Price Tag Belongs to a Test

A second set of numbers looks like an answer to a developer's budget question. On its Gemini 4 Argon (High) model page, evaluator Artificial Analysis gives this reasoning version a 53 on its Intelligence Index and places it eighth among 223 models in the page's comparison class. Its listed prices are $2 per million input tokens and $10 per million output tokens. A token is a piece of text a model processes or produces; input is what you send, output is what it returns. The page also gives a $1.99 cost per Intelligence Index task.

That last figure needs its full label. Artificial Analysis combines the costs of running its evaluations, accounting for input, cached material, reasoning and answers, then weights tasks by their place in its index. The index itself incorporates ten evaluations, including Terminal-Bench 4.0, SciCode and Humanity's Last Exam. The resulting $1.99 describes a weighted average task in that mix. A developer sending short prompts, asking for long code changes or repeatedly retrying failed tasks will use a different mix of tokens and work. The page gives a way to compare models on a defined course, rather than a quote for completing your ticket.

There's another clue on the same page: Argon generated 110 million output tokens during the Intelligence Index evaluation, against a listed median of 82 million for comparable models. That total doesn't say how long a response to your prompt will be. It does show why an output-token price can't answer the whole cost question by itself: how much a model produces while working affects the bill. A cheaper rate can lose its appeal if a workload needs far more output or repeated attempts; a higher rate can be worthwhile if the task finishes with less work. The reported numbers don't decide either scenario for a reader's own project.

Argon supports a context window of one million tokens. That's the limit on combined input and output the model can handle, useful to know if you want to bring a lot of material into a task. The same page notes that output commonly has a lower limit that varies by model. A roomy context window therefore tells you about capacity to take in material, while the coding tests ask whether the model can do the requested work with it. Those are separate questions for anyone contemplating a large repository or a long-running agent.

The release history and the scorecards lead to a practical way to read Argon's debut. After Google's summer of smaller Flash releases, this is a flagship model with reported strengths worth testing. Start with the benchmark closest to your job, inspect who ran it and under what conditions, then look at token use alongside the rate. If you're deciding whether Argon belongs in a coding workflow, the most revealing question is narrower than “Which model is number one?”: what happens when it tries to finish your task?