Google Announces Gemini 4 Argon at $2 Input and $10 Output, but Broad Access Is Still Pending

Google announced Gemini 4 Argon on September 30

Google announced Gemini 4 Argon on September 30, 2026 as a frontier model for long-running software engineering, enterprise knowledge work and cybersecurity defense. The company says the first live rollout is to selected trusted cyber defenders in its Fairwind Program. It plans to widen access after gathering feedback and strengthening guardrails, beginning with paid API customers and Google AI Ultra subscribers.

That sequence matters more than the launch language. The reviewed public material does not provide a general-release date, public API identifier, account eligibility page or consumer rollout schedule. Google’s live developer pricing catalog, updated October 1, did not yet contain an Argon entry. Argon is an announced model with narrow early access today—not a generally available endpoint that every builder can call.

The announced price is $2 input and $10 output

Google says Argon will launch at an introductory rate of $2 per million input tokens and $10 per million output tokens. Cached input is described as 95% off the input rate, which corresponds to $0.10 per million cached input tokens. The announcement does not state a free-tier allowance, batch rate, context-based surcharge, cache-storage charge, regional premium or date when billing begins.

Treat those numbers as planned introductory list prices, not a current invoice guarantee. The word “introductory” leaves future duration and replacement pricing unspecified, and the absence of Argon from the public pricing table means there is not yet a complete public billing contract to evaluate. Cost-per-task also depends on reasoning length, unusually large outputs, retries, tool calls and whether cache reuse is attainable.

One million output tokens is not a context-window claim

Google says Argon increases the maximum output-token limit from 64K to 1 million so the model can sustain long reasoning and generation trajectories. That is a major interface claim, but it is specifically an output limit. The reviewed announcement does not state Argon’s maximum input size or total context window.

A larger ceiling does not establish that typical responses should approach it, that long trajectories remain correct, or that an application can safely consume the resulting latency and cost. Teams should separately measure accepted outcomes, completion time, truncation, repetition, tool-state drift and reviewer burden across increasing output bands once a documented endpoint is available.

Google reports leads in coding, enterprise and video evaluations

Google reports Argon at 77.9% on DeepSWE v1.1 for long-horizon software engineering, 51.3% and first place on Zapier’s AutomationBench, 91.7% on LVBench for long-video understanding, and leadership on the Vals Index plus named finance and legal evaluations. These figures come from Google’s launch presentation; AccessAllGPT did not reproduce them or review complete per-task traces and contamination controls.

The announcement also describes internal outcomes: a quantum subroutine example beating a published baseline by 40%, agent-driven memory optimizations that freed more than 300 TiB after rollout, and code-migration work including a 32K-line SIMD replacement for libgav1. These are Google-reported case studies, not general performance guarantees. Buyers still need workload-specific comparisons with fixed tools, stop conditions and accepted-output scoring.

Cyber defenders get a distinct early-access path

Google says Argon can find, validate and patch vulnerabilities and that selected Fairwind defenders receive a configuration without cyber guardrails for authorized defensive use. The Fairwind page says a subset of approved partners can use Argon as a standalone managed model or with CodeMender. It prioritizes governments, critical-infrastructure operators, core technology platforms and academic defensive benchmarking labs.

This is controlled access, not an open cyber model release. Participating organizations agree to user-level authentication, phishing-resistant multifactor authentication, access tracking and restrictions to internal cybersecurity, incident-response or penetration-testing teams. Google prohibits sharing, resale and malicious tasks. The page says zero data retention is supported when Argon is accessed directly as a managed model on Gemini Enterprise; teams must verify that configuration and contract in their own tenant.

The safety evidence is promising but incomplete publicly

Google says Argon ties for first at 68% on CWE-bench v1, improves on 3.8 Flash Cyber in internal vulnerability testing and leads Gray Swan’s indirect prompt-injection benchmark. It describes internal and external red teaming, activation monitoring for misuse, chain-of-thought and action monitoring for misalignment, and hardened isolated environments for high-risk evaluations.

Those are vendor claims and control descriptions. The launch page does not provide a public system card with complete evaluation sets, confidence intervals, subgroup failures, monitor thresholds, bypass rates or deployment incident data. Monitoring model reasoning may help detect failures, but it should not replace deterministic authorization, least privilege, isolated execution, independent logs and tested rollback around consequential actions.

What builders and buyers should do next

Most teams should wait for three concrete artifacts: an account-visible API model identifier and availability notice, a full model or system card, and complete pricing and service terms. At that point, freeze representative coding or knowledge-work cases and compare Argon with the current production model under identical tool permissions, time budgets, stop conditions and review rules. Measure accepted outcomes, severe failures, latency, total tokens, cache behavior, retries and cost per accepted task.

Eligible Fairwind defenders can consider a bounded pilot now, but should record the exact managed route, zero-retention setting, permitted users and tasks, authentication, external tool authority, monitoring, escalation and revocation evidence. Do not schedule a production migration from the headline alone. The useful decision today is wait, or run a tightly governed early-access evaluation if Google has explicitly approved the organization.

Disclosures

AccessAllGPT did not receive Google access, credits, a subscription, briefing, demo, review or compensation for this article. Google did not sponsor, review or endorse it. AccessAllGPT Research is operated by NeuralArc, is independent, and is not affiliated with Google, Google DeepMind or organizations cited. Publication-wide relationships are listed on the disclosures page.