What Google is introducing
Google DeepMind positions Argon around extended software engineering tasks, enterprise knowledge work and cybersecurity defense. Its model page describes reasoning across multiple steps and finding, validating and patching software vulnerabilities. These are Google’s capability claims, not results from a LOCATEWISE product test. [2]
The practical promise is continuity: can a model keep track of a difficult job through investigation, changes and checking? For a software team, a useful evaluation might be a contained maintenance task with known requirements. For an analyst, it might be a brief assembled from a fixed set of documents. In either case, the finished artifact matters more than a persuasive account of what the model says it did.
Google announces a one-million-token output limit. [1] Do not read that as a guarantee about every consumer chat. More room for an answer is useful only if the extra work improves a result you can inspect.
Release date, API access and AI Ultra availability
| Audience | Announced position |
|---|---|
| Trusted Fairwind cyber defenders | Initial rollout underway. |
| Paid API customers and AI Ultra subscribers | First wider-release groups; no date given. |
| Other developers, businesses and consumers | Wider availability remains a stated plan. |
The announcement describes a phased release while safeguards are refined. [1] Being named in a future access group does not establish that your account has access now.
Announced introductory API rates are $2 per million input tokens and $10 per million output tokens, later rising to $4 and $20. The introductory period has no stated end date. [1]
Before budgeting, look for the actual service’s pricing and usage terms at release. Keep subscription access and metered API use as separate purchasing questions. A rate per token alone cannot tell you the cost of completing your particular task.
The Gemini app has its own updates
Google’s app release notes dated 30 September describe reusable skills rolling out globally to users aged 18 and over. Skills save instructions for recurring tasks, and Google says they will eventually replace Gems. That entry concerns an app feature; it does not announce general Argon access. [3]
Check the feature you need in the product you use. A new shortcut, reusable instruction or interface feature can be valuable without answering which underlying model is available. Conversely, a model announcement does not establish that a specific app feature has reached your account.
Prepare a comparison that can change your decision
Our suggested next step is to keep a small task and its current result as a baseline. Define success before trying the new model: correct conclusions, usable code, traceable references or fewer necessary corrections. Record elapsed time and your own review effort alongside the output.
For example, give each model the same short document pack and ask for a comparison with references to the relevant passages. Save the prompts and original files. Count unsupported statements and missing exceptions, then decide whether the improvement justifies switching. This is a proposed test design, not evidence that Argon has passed it.
DeepMind publishes a benchmark table covering several kinds of work, with results varying by evaluation. [2] Use that as a reason to choose a relevant test, not as a substitute for one. When access reaches you, rerun the same task and inspect the differences. Until then, make subscription decisions on the features you can confirm today.
Read the underlying guidance.
- Gemini 4 Argon: our next era of frontier intelligenceGoogle · checked 2026-10-02. 30 September announcement, initial Fairwind rollout, phased future access, announced output allowance and future introductory API prices; no general-access date or introductory-period end date stated.
- Gemini — Gemini 4 ArgonGoogle DeepMind · checked 2026-10-02. Vendor-described task focus and capabilities; published benchmark table spanning different evaluations. Does not establish actual account availability or independent performance.
- Gemini Apps’ release updates & improvementsGoogle · checked 2026-10-02. 30 September app release entry for reusable skills, adult-user rollout and planned replacement of Gems; a product-feature announcement distinct from Argon model access.
Researched and written with AI. The producing AI checked English, source attribution, announcement-versus-availability wording and JSON structure. A separate AI reviewer independently read the official sources and reviewed the English text on 2 October 2026. Human English and technical review have not been completed. human English and technical review; actual account-access verification and product evaluation; launch-date, regional eligibility and final service-term checks when wider access begins.
