Skip to main content
Gemini 4 Argon is the first model in the Gemini 4 family. Google positions it for repository code, finance, law, charts, and video, and for security defense when the job is to fix a flaw, not to exploit it. The announcement is from September 30, 2026. Broad access is still staged: first a group of defenders in the Fairwind program, and Google describes the next opening for paid API customers and Google AI Ultra subscribers. The model id on this page is gemini-4-argon. A second id, gemini-4-argon-cyber-permissive, exists only for cybersecurity workloads approved in that program. The 1,048,576-token window is a single budget, shared by input, thinking, and output. One request stops at 65,536 tokens when streaming, or at 32,768 on a non-streaming HTTP call. Generation can continue up to the window on GenerateContent, with the SDK resuming on its own. Artificial Analysis records text and image input and text output. Agentic video understanding exists and comes turned off.

​
When to use Argon and when to use Opus 5.5

The difference is the kind of task. Argon fits when the material is repository code, a finance or legal file, a chart, or a video. Opus 5.5 stays ahead when the work needs open-ended judgment, a terminal, or a program written from scratch.
Use Argon when the answer depends on code, a PDF, and an image in the same request. For a short classification queue, LOW is enough. For a change across several files, move up to HIGH.

​
What Google measured

The numbers below were published by Google. Where another source uses the same benchmark name, the result is not automatically comparable.

​
Repository code

On DeepSWE v1.1, Google measured 77.9% and called the result state of the art. Meta published 78.8% for Muse Wonder on DeepSWE 1.1. The two numbers come from different publishers. Inside Google, Argon agents were used on C and C++ to Rust migrations, from libraries such as re2 and libgav1 up to more than 800,000 lines of the Fuchsia Zircon kernel. On libgav1, Google reports a memory-safe Rust decoder, 2.7 times faster than the previous Rust version, with the same video on the output.

​
Finance, law, and processes

Google says Argon leads the Vals Index, which weights finance, coding, law, and tax by each sector’s share of U.S. GDP. On Finance Agent v2, the methodology Google published scores 65.4%. On Zapier’s AutomationBench, end-to-end business processes, 51.3%, first place on their tally.

​
Charts and video

On Chartography with no tools, Google measured 71.6%. On LVBench, video understanding, 91.7%, the number they present as state of the art.

​
Defense, not attack

On CWE-bench v1, which measures vulnerability remediation, Argon ties for first at 68%. Wiz reported, in the Scan for Good program, a critical flaw in healthcare software that earlier models had missed. This article does not describe how to reproduce that flaw.

​
What the independent benchmarks show

In Artificial Analysis’s own measurement, the high variant scored 53 on the Intelligence Index, ranking 8th among 216 models in the same class. The median for that price range is 27. In that run the model generated 110 million output tokens on the index, against a median of 85 million. Average cost was 1.99 dollars per index task, calculated at the introductory price. Output speed was not published. On BenchLM, the partial aggregate is 80.88 out of 100, ranking 6th among 897 models. Coverage is still partial (31 benchmarks), so the score is conservative. BenchLM lists the Intelligence Index as 52.6. The Artificial Analysis page scores 53. Use 53.
The table mixes numbers published by Google with GDP.pdf from Artificial Analysis. The Intelligence Index, the amount of text, and the cost per task are from Artificial Analysis. The BenchLM aggregate uses partial coverage. Results depend on the harness, the tools, and the thinking level. Validate in your own workflow before changing a production setup.

​
Where it doesn’t lead

  • On GDP.pdf, Artificial Analysis measured 21.8%. All-pass professional documents are not the strong point of that run.
  • On DeepSWE, Google’s 77.9% sits below the 78.8% Meta published for Muse Wonder. Do not treat the two as the same test until you check the harness.
  • A single request still stops at 65,536 tokens. Going past that requires GenerateContent continuation.
  • temperature, top_p, and top_k are ignored. The effort control is thinking_level.
  • The model on this page is not gemini-4-argon-cyber-permissive.

​
Where it adds the most value

​
A change that crosses several files

The fix does not fit in one file, and the tests already exist. Argon edits, runs the check, and stops if the second attempt fails. Expected result: the diff inside the scope, with the requested tests passing.

​
A brief written from the files

The question is about finance or law, and the answer has to come from the documents, not from general knowledge. The model cites the passage and lists what it did not find. Expected result: every figure or clause has a source, and the gap is explicit.

​
A chart or a video

The figure is in the image or in the length of the video, not in a text field. Argon points to the frame or the moment the answer came from. Expected result: the value, the unit, and the visual point of origin.

​
How to apply it in Tess

1

Pick code, a file, or media

Prefer a repository change, a brief from files, a chart, or a video. For a classification queue, use LOW.
2

Lock what counts as done

State the outcome, the files that may change, and the check that ends the task.
3

Set tools and approvals

Enable only the tools you need. Require confirmation before sending, publishing, or changing data.
4

Choose the thinking level

LOW for classification and summaries. MEDIUM, the default, for general code, tools, and documents. HIGH for several files.
5

Sample before you let it run

Run one full case, check what changed, and only then repeat the flow.

​
Three playbooks to get started

​
1. A repository change with tests

Input: the repository, the expected behavior, and the tests that already exist.
Tools: reading and editing files, running the tests.
Done when: the requested tests pass and the diff stays inside the scope.

​
2. A brief with a citation

Input: the documents and the finance or legal question.
Tools: file reading.
Done when: every fact has a source and missing data is explicit.

​
3. Reading a chart or a video

Input: the chart, or the video with video understanding turned on.
Tools: file reading.
Done when: every value points to a visual source, or the absence is written down.

​
Limitations and best practices

  • The opening for developers, enterprises, and consumers was still staged at the announcement. The first group was Fairwind.
  • The introductory price is 2 dollars of input and 10 dollars of output per million tokens. After that period, Google applies 4 dollars of input and 20 dollars of output. Implicit cache, from a prefix of 6,144 tokens, is discounted 95% off the input price. Explicit cache is not supported.
  • The id gemini-4-argon-cyber-permissive is a different model, for workloads approved in Fairwind. gemini-4-argon refuses harmful cybersecurity requests and harmful chemical, biological, radiological, or nuclear requests.
  • Google describes Argon as its most resistant model so far to indirect prompt injection, and says it monitors reasoning to stop an action that goes past the request.
  • Review a sample before publishing, sending, or automating decisions.
  • Keep human approval for financial, legal, clinical, or high-impact decisions.
See also: Models and Costs · Claude Opus 5.5 · Gemini 4 Argon — Google · Gemini 4 Argon — developer guide.