Google announced Gemini 4 Argon on September 30, with initial access restricted to selected cybersecurity partners. Reuters reports that Google provided no public release timeline.

The announcement therefore describes a staged release, not general availability. Google also describes participation in voluntary United States government prerelease access. Neither statement establishes when ordinary users will receive the model.

Claude can be your analyst, editor, and strategist.
But most professionals are using it to fix grammar.

These 200+ Claude prompts take it from grammar tool to your most powerful AI work assistant.

Sign up for Superhuman AI and get:

Claim your free prompts

One technical change concerns output length. Google reports an increase from 64,000 to one million output tokens. This is a company claim about generated output capacity. It should not be confused with an input context limit.

That distinction matters when reading specifications. An output allowance describes how much a model may generate. It does not establish whether a long answer is correct. The announcement offers no universal accuracy guarantee for outputs of that length.

Independent evaluations provide narrower evidence than the release language. The inspected Vals Index table gives Argon a score of 68.90. The index combines finance, coding, legal and tax tasks. Its sector weights reflect selected parts of the United States economy.

This is not a measure of all human knowledge or every occupation and country's economic structure. A composite score depends on the tasks and weights chosen. Different weighting could answer a different evaluation question.

Vals uses eight benchmarks, including six private evaluations. Coding tasks include terminal work, application construction and code migration. The tax component concerns United States corporate questions. These details identify the tested work rather than a universal capability.

The methodology also changed on September 25. Tax entered the index, and a harder terminal benchmark replaced its predecessor. Scores from different versions should not be treated as directly interchangeable. The inspected table and its accompanying leadership prose were also inconsistent.

For that reason, the table supports a dated score only. This issue does not repeat the stale prose ranking. It also makes no claim that small score differences prove reliable superiority.

It comes from someone who’s been there, done that, and has the receipts. Get ready for Return on Influence Festival ‘26. An entire day dedicated to influencer marketing with speakers running some of the best programs in the world. 

October 21, 2026, 10 am - 5 pm ET.  Free and online.

Save your spot now

CWE-bench examines a different problem. Its private set contains 120 defensive code auditing and repair tasks. Agents receive repositories without the identity or number of vulnerabilities. They must find problems and produce repairs. The set covers 73 weakness types across six programming languages. That breadth still leaves evaluation bounded by the selected repositories and tasks.

Programmatic checks test whether exploits are blocked and existing tests still pass. A separate judging panel examines the patched code. These methods distinguish a completed patch from a plausible written explanation.

Argon recorded 68 percent on the programmatic first-attempt measure. GPT-6 Astra and Grok 4.7 recorded the same rounded result. However, the evaluated systems used different agent harnesses. Argon used Antigravity, while Astra used Codex.

The comparison therefore concerns models working through particular tools. It does not isolate the model as the sole cause. Each task allowed four runs, with an hour per run. Additional attempts and different grading tracks answer different questions.

Private held-out tasks also limit public reproduction. The published results are evaluator measurements, not an independent rerun here. They cannot certify every repository, language or operational security setting.

Google describes controls for misuse, prompt injection and environment isolation, alongside monitoring of model reasoning and actions. These remain attributed descriptions of safeguards, not proof of immunity.

The available evidence separates capacity, task performance and access conditions. None establishes that general intelligence or universally safe autonomy has been achieved. This regular check-in keeps those distinctions visible for your own technical decisions.

You don't have to scroll every AI thread, track every new tool, or watch every demo. 

The Rundown AI breaks it all down for you — the latest AI news, tools, and tutorials in one free 5-minute email every morning. 

Trusted by 2M+ professionals at Apple, Google, and NASA.

Sign up to start learning.