Gemini 4 Argon Could Roll Out More Widely Next Week

Argon was officially introduced on September 30, 2026, with limited access for cybersecurity researchers. The current reports concern a broader rollout, potentially through Antigravity and developer APIs.

Reported signals:

• Argon configurations are appearing in Antigravity.
• Selected users are reportedly seeing A/B tests.
• Older models may be removed from Antigravity.
• Gemini 4 Flash could launch around the same time.

Google has not confirmed a public release date.

Technical Specifications

The biggest change is the output limit. Gemini 3.1 Pro supports roughly 64K output tokens. Argon supports up to 1 million tokens per trajectory. For developers, that matters mainly for long-running agents.

The key upgrade is not “more text.” It is the ability to keep analysis, coding, testing and debugging inside a longer continuous run.

What Changes for Developers

Gemini 3.1 Pro already supports large context windows, tool use and coding. Argon is more focused on continuous execution.

Instead of splitting a task into separate requests:

1. inspect repository;
2. generate patch;
3. run tests;
4. inspect failures;
5. generate another patch;

Argon is designed for:

analyze → edit → test → debug → repeat

This is the main difference for coding agents.

Coding Performance

Google reports 77.9% on DeepSWE v1.1. Argon has also been tested on large internal codebases. One project involved C/C++ to Rust migration across codebases exceeding 800,000 lines, including work related to Fuchsia's Zircon kernel.

Another involved Google's libgav1 AV1 decoder. Argon assisted with roughly 32,000 lines of SIMD-related code. Google says the resulting Rust implementation was about 2.7× faster than the previous Rust port. The target use case is repository-scale engineering, not isolated code generation.

Cybersecurity

Argon is also optimized for defensive security tasks:

• vulnerability discovery;
• exploitability analysis;
• patch generation;
• remediation.

Google reports 68% on CWE-bench v1. Cybersecurity researchers were among the first external users to receive access.

Benchmarks

The more important real-world metrics will be tool-call reliability, latency, cost per completed task and stability over long agent runs.

Argon in Antigravity

Reported Antigravity configurations include:

• 256K context;
• 512K context;
• 900K context.

Reports also mention higher quota usage for larger modes, around 1.3× for 512K and 1.8× for 900K. These figures are not confirmed as final API limits.

Argon, Barium-B and Carbon

Several internal checkpoint names have appeared around Gemini 4. Barium-B is reportedly the checkpoint behind the current Argon model. Carbon is described as a newer internal checkpoint still being tested at Google. Carbon has not been announced as a commercial model.

What About Gemini 4 Flash?

Gemini 4 Flash has not been officially detailed. If Google follows its previous product structure, Flash would likely target lower latency, lower cost and high-volume API workloads. Specifications, pricing and release timing remain unconfirmed.

Gemini 4 Argon vs Gemini 3.1 Pro

Could It Roll Out Next Week?

Possibly, but Google has not confirmed the timing. The strongest signals are Antigravity testing, A/B access and reported cleanup of older model configurations.

The important distinction is that Gemini 4 Argon has already launched in limited form. The next step would be broader developer access.