Google Announces Gemini 4 Argon: Coding Gains and a 1M-Token Output Limit, but Access Is Limited

Google announced Gemini 4 Argon on September 30, positioning it as a frontier model for long-running software engineering, enterprise work and defensive cybersecurity. The immediate caveat is access: **this is not a broadly available developer API release**. The first rollout is to selected cyber defenders in Google's Fairwind Program; Google says wider availability will follow after further safety testing.

What changed for coding?

Google says its engineers are already using Argon for everyday debugging, algorithm design and large codebase migrations. Its examples include agents helping move C/C++ systems toward Rust, from core libraries to work on the Fuchsia Zircon kernel. Google explicitly says these critical rewrites are undergoing automated and manual auditing, emulation testing and review before production rollout. That is a useful distinction: the announcement describes serious engineering work, not a claim that an agent can safely merge a kernel rewrite unattended.

Another notable specification is a **1-million-token output limit**, increased from 64,000 tokens. This is the maximum response or reasoning trajectory Google describes; it is **not** a claim that the model has a new 1-million-token input context window.

What do the benchmarks show?

On Google's published model table, Argon scores 77.9% on DeepSWE v1.1, 55.0% on FrontierSWE v2, 91.9% on Vibe Code Bench and 57.4% on Terminal-bench 4.0. The pattern is more informative than any single headline number: results vary by task and setup. Google's methodology notes that some comparison scores come from public leaderboards or model providers, while Argon's own runs use specified harnesses. **We have not benchmarked Argon ourselves**, and these figures should not be treated as an apples-to-apples verdict on everyday coding agents.

For security teams, Google also says Argon can find, validate and patch vulnerabilities, and that selected defenders will receive access to its full defensive cybersecurity capabilities. Those are Google and partner reports from limited early use, not evidence of general availability.

Availability and price

Google's rollout begins with trusted cyber defenders. It says developers, enterprises and consumers will gain access later, starting with paid API customers and Google AI Ultra subscribers. The company has not announced a general-release date.

For that future API rollout, Google lists an introductory price of **$2 per million input tokens and $10 per million output tokens**, with cached input tokens 95% cheaper. Its footnote says the price will become **$4 input / $20 output per million tokens** after the introductory period. These figures are a pricing announcement, not a live public purchase option today.

Bottom line

Argon is worth watching for long-horizon coding and defensive-security workflows, especially if the output limit proves useful in real projects. For now, the practical decision for most developers is to wait for access and test it on representative repositories, with review quality, latency and total cost measured alongside benchmark scores.