Gemini 4 Argon Coding Benchmarks: Long Decode, Limits and Use Cases

Quick summary

Google’s model card reports high Gemini 4 Argon scores on DeepSWE v1.1 and Vibe Code Bench, while other models lead on FrontierSWE v2 and Terminal-Bench 4.0. Its 1M-token window is shared by input, thinking and output; long decode also requires background execution.

Why the window and benchmark conditions matter

Large coding tasks can involve long context and many files, so output length may affect whether a model can return a complete patch in one response. A long answer still needs compilation, tests and human review. Benchmark scores apply to their published evaluation conditions, not automatically to your project.

Key takeaways

• The 1,048,576-token limit is a shared context window, not a fixed amount of final code.
• Gemini 4 Argon can decode beyond 65,536 tokens with background execution, while input still uses part of the window.
• Google’s model card reports 77.9% on DeepSWE v1.1 and 55.0% on FrontierSWE v2.
• Before using Argon in DSH, check whether the installed Google model catalog lists gemini-4-argon.

Gemini 4 Argon’s 1M-token limit is shared

The Gemini API model ID is gemini-4-argon. Google positions it for complex coding, reasoning and multi-step work. Its context window is 1,048,576 tokens shared across input, thinking and output. This is not a promise of one million final-code tokens for every request. Gemini 4 Argon API documentation

The normal synchronous request has a 65,536-token output limit. To produce a longer single response, Google’s long-decode flow requires background execution and a higher max_output_tokens value. The service completes one response through continuation steps of up to 64K tokens; the available output also depends on tokens used by the input and thinking. Long decode and background execution

That can help with a large refactor: the model has room to return a longer cross-file patch, tests and explanation. It does not remove repository permissions, compilation, tests or human review. Reading more code also does not mean the model has inferred every hidden dependency.

What do Google’s published coding benchmarks show?

Google DeepMind’s model card reports the results below. The comparison scores are from the same model card; they remain results published by Google. Gemini 4 Argon model card and evaluation methods

Across these results, Argon scores strongly on DeepSWE v1.1 and Vibe Code Bench, while other models score higher on FrontierSWE v2 and Terminal-Bench 4.0. A benchmark can help shortlist candidates. It cannot establish that a model will perform better in your repository.

The interactive panel below shows selected scores from Google’s model card and estimates API token costs from adjustable inputs, cache-hit rate and output volume. The costs are scenarios, not measured bills; the benchmark scores are not tests run by this site.

Choose a coding model by the task

For work across several files, long context or a larger patch, Argon’s long-decode capability is worth evaluating. Define an inspectable result first: the diff, acceptance conditions, test output and allowed directories.

For terminal commands, dependency failures or a narrow algorithm problem, check whether a benchmark matches that task. Google’s model card reports a higher Terminal-Bench 4.0 score for Claude Opus 5.5 and a higher FrontierSWE v2 score for GPT-6 Astra. Do not turn a lead in one test into a claim that a model is better at all coding.

A repeatable local task is closer to your work than a public leaderboard. Keep the task, model version, context, tools and acceptance checks fixed. Record how many attempts pass and how much manual revision is needed. The site’s coding-model test template can help record those conditions; it does not publish rankings from tests run by this site.

Google’s code-migration examples are still under review

Google says Argon agents are working on C/C++ to Rust migrations across codebases ranging from core libraries with tens of thousands of lines to the Fuchsia Zircon kernel, which has more than 800,000 lines. Google also says these rewrites are undergoing automated and manual audits, emulation tests and code review before production rollout. The source does not describe all of this work as already merged into production. Google’s engineering examples

Google also describes an optimization in libgav1: Argon agents replaced about 32,000 lines of SIMD code in an existing Rust port, then used repeated profiling experiments to produce safe Rust that the compiler could vectorize. Google reports that this version runs 2.7× faster than the existing Rust port with identical video output. The comparison is with the Rust port, not with the optimized C++ implementation.

These examples show how a model can participate in long engineering workflows. They also show why generated code still needs testing, auditing and review before deployment.

Check the DSH model catalog before choosing Gemini 4 Argon

DSH’s Google provider uses Google’s native Generative AI protocol. The Google provider setup guide documents the current setup path. Google’s API model ID is gemini-4-argon, but the DSH model picker is supplied by the installed version’s model catalog. Check that your picker lists Argon and that your DSH version and Google API project can use it.

If Argon is missing from the catalog, do not force the Google-native API into a custom provider form for another protocol. Check the DSH version and catalog first. This article summarizes public documentation; it does not claim that this site has sent a live Gemini 4 Argon request through DSH.

Common questions

Does Gemini 4 Argon have a 1M-token output limit?

Not for final text alone. The API documentation defines a 1,048,576-token window shared by input, thinking and output. Long-decode output must also leave room for input and other tokens. Requests above 65,536 output tokens require background execution.

Is Gemini 4 Argon better at coding than Claude Opus 5.5?

There is no universal answer. Google’s model card reports a higher Argon score on DeepSWE v1.1, while Claude Opus 5.5 scores higher on Terminal-Bench 4.0. Match the benchmark and acceptance checks to your task.

Do public benchmarks prove it will work in my repository?

No. Results apply to specific test sets, model settings and tool conditions. Run the same repeatable task against your codebase and checks before deciding whether it fits your workflow.

What if DSH does not list Gemini 4 Argon?

Check the installed DSH version and its provider catalog. The built-in Google route uses Google’s native protocol; the custom-provider form has a different set of protocols. If the model is not listed, confirm version support instead of assuming that copying the model ID will make the configuration work.

Decide with your own acceptance task

Google’s published results make Gemini 4 Argon a candidate for long-context, multi-file engineering tasks, while other models lead on some coding benchmarks. Run one small comparison with the same repository, tests and tools before setting it as the default model in DSH.

Model and workflow guides