Google launched Gemini 4 Argon on 13 October, publishing a Google Cloud engineering account of an overnight coding test for the model. Argon is rolling out to Pro and Ultra subscribers in the Gemini app and AI mode in Search, with access on Antigravity for Ultra subscribers, The Verge reported.
Key points
- Google claims Argon beats rival models across coding, cybersecurity, science, finance and law benchmarks.
- Introductory prices are $2 per million input tokens and $10 per million output tokens.
- In Google’s overnight coding test, two proposed optimisations were shipped and four were rejected in review.
Argon moves beyond the Fairwind rollout
The subscriber release widens access after a limited rollout to trusted cyber defenders for testing potential risks. AI Affairs covered that restricted launch, and separately reported employee doubts about Argon’s coding. This is Google’s first major frontier-model release since Gemini 3 in late 2025, according to The Verge.
Google says Argon improves reasoning across complex workflows and reduces hedging, repetition and familiar chatbot phrasing. It also claims the model outperforms GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5 on benchmarks spanning coding, cybersecurity, science, finance and law. Those are Google’s comparisons of benchmark performance, rather than measurements of how the newly released model responds to every kind of question in its subscriber products.
Looking up an answer to a question could involve the new model for a paying subscriber, given its rollout to a search mode. A question requiring several connected steps could put the claimed reasoning improvement to use, though the performance comparisons concern benchmarks rather than that particular answer.
Google is also offering Argon through Antigravity, its development platform for tasks in which a model can keep working towards a goal. That route matters for the coding test Google published: it put the model inside a running workflow, where proposed edits, tests and review could follow an initial instruction. Access there is limited to Ultra subscribers, while Pro and Ultra subscribers get the model in the Gemini app and Search.
An 11-hour run through Google’s Python code
Google Cloud engineers tested Argon on the Agent Development Kit, Google’s open-source framework for building and deploying agents. Its Python repository contains roughly 205,000 lines across nearly 800 source files and nearly 17,000 unit tests. The task was to find performance bottlenecks that crossed files while preserving what the model received and the software’s public interface, and to support changes with regression tests.
The engineers first ran Argon for about two hours without custom repository instructions. It found slowdowns in session copying and package imports, but its changes also altered a public JSON schema, missed edge cases involving mock events and ran only 216 nearby unit tests. Some code from that attempt went into the final commit, according to the engineering account.
For the overnight attempt, the team added maintainer guidelines to the repository and used Antigravity’s /goal mode, which keeps the goal in view as earlier context is compressed. It works like leaving the job instructions on top of a folder while filing away the notes beneath them: the agent can continue from a shorter account of its work without losing the original aim. The engineers asked Argon to open an unmerged pull request with before-and-after measurements, then let it run unattended for 11 hours.
Reviewing a long piece of work before it is accepted could still matter if much of the preparation happened overnight. A similarly organised task could arrive as a proposal to check rather than a completed change, as the Google run ended with proposed optimisations, some of which were rejected in review.
Argon measured a baseline on package startup, a 200-step conversation, a task-handoff test involving 20 sub-agents and a 12-node workflow, with each workload run 30 times. Google says the overnight run produced six optimisation proposals: two shipped after code review and four were rejected. The team measured a cut of about 40% in in-memory session preparation time in long sessions, at both 50 and 150 turns, and a 21% improvement in cold package imports. The model also ran all 16,892 unit tests before stopping.
Argon’s token prices rise in November
Argon’s introductory price is $2 per million input tokens and $10 per million output tokens. The Verge put those rates at around a fifth of the cost of GPT-6 Astra and Fable 5, while cautioning that a lower price per token need not make a task cheaper when models consume different numbers of tokens to complete it. For companies building on Argon, the size of a request and the length of its response would therefore matter alongside the advertised rates.
Google says it is already using Argon to find and patch cybersecurity vulnerabilities, reduce the computing power needed for YouTube’s comments section by half and debug Wing’s drone physics simulator. Those are Google’s accounts of work inside its own products and systems. Google’s introductory token prices will double at the end of November, The Verge reported.
