Gemini 4 Argon is not broadly available yet: access, pricing and what Google announced

What Google announced

Google describes Argon as the flagship model anchoring the Gemini 4 generation. It is positioned above the company’s previous Pro line and is intended to sustain longer, multi-step work rather than only answer short prompts.

The company highlights four broad areas: real-world software engineering, enterprise knowledge work such as finance and legal research, visual and multimodal analysis, and cybersecurity defense. Google also says the model can produce as many as one million output tokens in one trajectory, up from a previous 64,000-token limit.

A larger output ceiling can be useful for long code migrations, extensive research or multi-stage agent work, but it is not a guarantee of accuracy. It can also increase latency and cost. Teams evaluating Argon should still divide high-risk work into reviewable stages, use tests and source checks, and set explicit spending and action limits.

Who can use Argon now?

The only clearly available path at announcement is the Fairwind Program. Google DeepMind says the program is aimed primarily at governments and national cyber authorities, critical-infrastructure operators, core technology platforms and qualified academic security labs.

Applicants are vetted. Participating organizations must use controls such as individual authentication, phishing-resistant multifactor authentication and access tracking. The program permits authorized defensive work such as threat simulation, reverse engineering and malware analysis for research or protection; it does not permit malicious use or resale of access.

Fairwind partners receive a cybersecurity-capable version of Argon and can also use it with CodeMender, Google’s specialized code-security agent. Google says direct managed-model access through Gemini Enterprise can support zero data retention, but that statement applies to that managed access path and should not be assumed for every future Argon product.

For everyone else, the answer is to wait for a broader release. Google’s announcement says rollout will begin with paid API customers and Google AI Ultra subscribers. It does not name a day, list supported countries or confirm that every account in those groups will receive access at once.

Pricing is announced; public availability is not

Google’s introductory commercial price is $2 per million input tokens and $10 per million output tokens. Cached inputs are slated to receive a 95% discount from the regular input rate.

A simple example illustrates the difference between headline price and actual project cost. One million fresh input tokens plus one million generated output tokens would total $12 at the announced rates before any other platform charges. Reusing one million eligible cached input tokens would reduce that input portion to $0.10, making the same illustrative combination $10.10.

Real bills can be very different. A model designed for long trajectories may generate far fewer or far more than one million tokens, and agent systems can make repeated calls, use tools and retry failed steps. Google has not yet published every production detail that buyers need, including final regional availability, rate limits and service-level commitments. Google has not said how long the introductory period will last. It has said that after that period, prices will rise to $4 per million input tokens and $20 per million output tokens.

What the performance claims do — and do not — show

Google reports a 77.9% score on DeepSWE v1.1, a benchmark for long-horizon software engineering, and a 51.3% score on Zapier’s AutomationBench. It says Argon ranks first on the Vals Index, which covers finance, legal, coding and tax tasks, and reports a 91.7% result on LVBench for long-video understanding.

For defensive security, Google reports a 68% score on CWE-bench v1, tied with other leading models in the chart shown by the company. Google also says Argon improved over Gemini 3.8 Flash Cyber on internal Google and Wiz vulnerability-discovery evaluations.

These are useful signals, not independent proof of every workload. Google’s own evaluation document explains that several Argon results were self-computed, while scores for competing models often came from providers or public leaderboards. Harnesses, tool access, reasoning settings and time budgets can differ. Reuters also noted that Argon led selected competitors on some company-reported measures while trailing on other coding metrics included in Google’s release.

A sensible buyer should reproduce a small set of real tasks with private test data, fixed scoring rules and human review before making a deployment decision.

Why the cybersecurity release is restricted

Argon’s ability to find, validate and patch vulnerabilities creates both defensive value and misuse risk. Google says it is improving refusal systems for cyber and chemical, biological, radiological and nuclear threats, hardening sandboxes, testing against indirect prompt injection and monitoring agents for actions that exceed a user’s intent.

The company is also participating in a voluntary U.S. government process for pre-release model access. Google says it will collect feedback from early testers and refine safeguards before expanding availability.

This phased approach means the announcement should not be read as a normal consumer-model launch. The most capable cybersecurity configuration is initially being provided to controlled defenders, and broader products may have different guardrails.

What developers and buyers should do next

1. Do not redesign a production system around Argon until access appears in the relevant Google account and the final API documentation is published.

2. Treat the $2 input and $10 output rates as announced introductory pricing, not as a complete cost estimate.

3. Prepare a representative evaluation set now: long code fixes, research tasks, document workflows and security cases that reflect the organization’s real risk.

4. Record tool permissions, network access, data-retention settings and human approval points separately from model quality.

5. Compare Argon with existing models under the same harness. Public benchmark tables rarely create a fair purchasing comparison by themselves.

6. If an organization qualifies for Fairwind, review its governance and security requirements before applying; the program is not a shortcut to unrestricted general access.

What is still unknown

Google has not given a firm date for public or standard developer availability. It has not published a full list of countries, account tiers, rate limits or final service terms for the commercial rollout. The company also has not established through independent production evidence that the model’s strongest benchmark results will transfer to every coding, legal, finance or security environment.

The confirmed picture is narrower: Gemini 4 Argon was announced on September 30; approved cyber defenders are receiving controlled early access; broader access is planned to begin with paid API customers and AI Ultra subscribers; and introductory pricing has been announced. Until Google publishes the next rollout step, most people can evaluate the documentation, but they cannot yet assume they can run the model.