Gemini 4 Argon

Overview

Gemini 4 Argon is a new Gemini model announced on September 30, 2026. As of October 2, initial access is restricted; this announcement does not establish general availability in the Gemini app or public API.

For developers and enterprise teams, the decision is whether to prepare an evaluation or keep an existing deployment. The release has distinct output, workflow, and access implications.

What Changed

More output for sustained reasoning

Google raises the output ceiling from 64K to 1M tokens. This is an output limit, not a claim about input context capacity. A larger ceiling gives longer tasks more room, but does not guarantee lower latency or cost.

Coding and enterprise workflow results

DeepMind positions Argon for sustained coding, professional knowledge work, and multimodal tasks. Its performance table reports:

These are published evaluation results, not our own tests. The methodology generally uses the highest thinking settings. Google computes the DeepSWE result; AutomationBench uses Zapier's private evaluation set. Treat these as evaluation signals, then measure completion quality, latency, and spend on your own tasks.

Defensive cybersecurity beyond Flash Cyber

Argon can find, validate, and patch vulnerabilities. Google reports improved vulnerability discovery over 3.8 Flash Cyber, including internal and Wiz evaluations. This builds on an existing defensive capability; vulnerability patching was already part of Flash Cyber.

Compared With Previous Version

Use Gemini 3.8 Flash as the general-purpose baseline and 3.8 Flash Cyber for defensive workloads. These are different deployment variants, so Argon should not be treated as an automatic replacement for every 3.8 model.

The clearest output delta is Google's stated 64K-to-1M increase. For cost, 3.8 Flash's announced introductory API rates are $0.75 input and $3.75 output per million tokens. Argon's announced introductory rates are higher. The Cyber comparison concerns vulnerability discovery, rather than a universal speed improvement.

Recommendation: keep a working Flash deployment until Argon access and workload economics justify a change. A newer generation alone is insufficient reason to switch.

Availability & Access

Fairwind gives approved defenders early access to advanced models for protecting critical infrastructure. The program accepts access applications; applying does not guarantee approval.

The Argon announcement identifies paid API customers and Google AI Ultra subscribers as the starting groups for broader rollout. It does not give a firm public rollout date. An Ultra subscription alone should not be read as proof of current Argon access.

Announced API Pricing

Google also announces a 95% discount on cached input. The introductory period's end date is unspecified. These are announced launch rates, not confirmation that a public endpoint is available to purchase today.

For budgeting, compare total tokens per completed task as well as unit rates. Longer reasoning can make a successful workflow more expensive even when its output is useful.

Who Should Upgrade / Who Should Wait

• Evaluate if eligible: approved defensive teams with an authorized testing environment and a concrete vulnerability-remediation workload.
• Prepare an evaluation: developers whose long engineering tasks hit output limits, or enterprise teams with measurable document and business workflows.
• Wait before migrating: teams that require immediate public API access, predictable latency, or confirmed integration instructions.

Before switching production traffic, verify the available endpoint, supported tools, quotas, and access terms. This release announcement alone does not establish API compatibility with an existing Flash integration.

Sources

• Google launch announcement, September 30, 2026
• DeepMind capabilities and performance table
• Argon evaluation methodology
• Fairwind access program
• 3.8 Flash and Flash Cyber announcement