Google announced Gemini 4 Argon on September 30, 2026, describing it as a model for software engineering, enterprise work and cybersecurity defense. The headline number is a 1-million-token output limit—not a stated 1-million-token context window. Google has begun a phased rollout to selected cyber defenders, but has not announced a general-release date.
What does the 1-million-token limit mean?
Google says Argon’s maximum output has increased from 64,000 tokens to 1 million. Output is what the model generates; it is not the amount of text the model can take in. Neither Google’s detailed announcement nor its October 2 roundup specifies Argon’s input limit or total context-window size. Those figures should not be inferred from the output limit.
Google says the larger output allowance is intended to support longer reasoning and generation trajectories. It does not establish how much input a user can provide in one request.
What is Gemini 4 Argon designed to do?
Google positions Argon as a frontier model for complex software engineering, enterprise knowledge work—including legal and finance tasks—and cybersecurity defense. The company says the model can find, validate and patch critical software vulnerabilities. These are Google’s descriptions of its capabilities, not independently verified results.
Quick wins for a faster PC:
Clear out junk files and repair common Windows errorsFree Scan →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Google also reports results on several evaluations. The scores below are company-reported; the announcement does not provide independent replication or establish that the results predict performance on every real-world task.
| Evaluation | Google-reported result | What Google says it measures or indicates |
|---|---|---|
| DeepSWE v1.1 | 77.9% | Real-world, long-horizon software-engineering tasks |
| AutomationBench | 51.3% | Google says Argon ranks first |
| LVBench | 91.7% | Long-video understanding; Google says the result is state of the art |
| CWE-bench v1 | 68% | Vulnerability remediation; Google says Argon ties for first |
Google also says Argon leads the Vals Index and reports strong results on Vals Finance Agent v2 and Harvey’s Legal Agent Benchmark, but gives no numeric scores for those claims in the announcement. It says a Wiz internal, black-box penetration-testing benchmark found Argon ahead of Gemini 3.8 Flash Cyber in attack-surface discovery, vulnerability identification and proof-of-concept validation; no score is provided in the announcement.
What can Argon do for cybersecurity?
Google says Argon can autonomously find, validate and patch critical vulnerabilities. The company describes Wiz using the model for defense through its Scan for Good initiative and says Argon helped discover a critical vulnerability affecting healthcare software. Google’s announcement does not name the software or vulnerability, so the account does not establish further details about the incident.
Google describes safeguards intended to limit misuse and keep the model’s actions aligned with a user’s intent. Its stated measures include:
- Refusing harmful cyber and chemical, biological, radiological and nuclear requests while allowing legitimate dual-use work.
- Testing safeguards with internal and external red teams.
- Training and testing against indirect prompt injection.
- Monitoring reasoning and actions to stop behavior beyond the user’s intent.
- Isolating sandbox environments before high-risk training or evaluation.
Google says trusted defenders and its internal teams will receive Argon without cyber guardrails. These descriptions reflect Google’s account of its safeguards; they are not an independent safety certification.
Who can use Gemini 4 Argon, and is it generally available?
The initial rollout is limited to selected trusted cyber defenders through Google’s Fairwind Program. Google says it will gather feedback and iterate on guardrails before expanding availability to developers, enterprises and consumers. It names paid API customers and Google AI Ultra subscribers as starting audiences for broader access, but gives no date for that expansion or for general availability.
Fairwind is a limited-access program for governments, trusted Google Cloud customers and cybersecurity partners, according to Google’s September 2 description. Its defensive work includes finding, verifying and fixing vulnerabilities. Participating organizations agree to operational restrictions, including limiting access to internal cybersecurity, incident-response or penetration-testing teams and using protections such as multifactor authentication. Membership in Fairwind alone does not establish that an organization has Argon access.
How much does Gemini 4 Argon cost?
Google announced the following API token rates. The announcement does not say when billing begins or how long the introductory period lasts.
| Token type | Introductory rate announced by Google | Rate after the introductory period |
|---|---|---|
| Input | $2 per million tokens | $4 per million tokens |
| Output | $10 per million tokens | $20 per million tokens |
| Cached input | 95% below the introductory input rate | Not stated in Google’s announcement |
The figures are announced API rates, not a total cost estimate: actual charges depend on token use and the applicable rate when a request is billed.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.What to take from Google’s announcement
Argon’s million-token headline refers to generated output, while its input and context-window limits remain unspecified. Google is pitching the model for long-horizon coding, enterprise work and cyber defense, with an initial, controlled defender rollout. Its benchmark scores, capability descriptions and safeguards should be read as Google’s claims; the announcement does not provide independent confirmation.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
