Frontier models

Gemini 4 Argon

Gemini 4 Argon is a Google DeepMind model announced 30 September 2026. First access is trusted cyber defenders through the Fairwind Program, not the public. Google says it is in the U.S. government’s voluntary process for pre-release model access, and that broader access will start later with paid API customers and Google AI Ultra subscribers. No date is given. There is no public API id. Introductory price is $2 input and $10 output per million tokens. Cached input is 95% off the input price. After the introductory period the post lists $4 and $20. The post does not say when the intro ends.

Status

Limited access, trusted cyber defenders first

AA score

53 (high effort)Artificial Analysis Intelligence Index, one independent score across many tests, fetched 1 Oct 2026

Price

$2 in, $10 out (introductory)US dollars per million tokens (word pieces) in and out, unless noted

Released

30 Sep 2026 (announced)

Last checked

1 Oct 2026newest source date on this page

Training

Google says it trained Argon for cybersecurity defense, and that prompt-injection robustness included adversarial training. Google says it monitored training runs, sent alerts to an incident-response team, and did not feed those findings back into training so the model would not learn to hide from the monitor. Parameter count, training compute, dataset, and knowledge cutoff are not in the 30 September post. No DeepMind model card was up on 1 October 2026.

Inference

Google does not publish tokens per second, rate limits, regions, or thinking-level controls in the announce. Artificial Analysis lists speed as unknown for Gemini 4 Argon (High), and says that page is the reasoning version. Google says the model is multimodal. Google does not publish an input and output modality list.

Context

Google says the output token limit is 1 million tokens, up from 64,000. Google does not state an input context limit in the announce. Artificial Analysis lists a 1 million token context window for Gemini 4 Argon (High). That figure is Artificial Analysis’s, not a second Google sentence.

Methods

For trusted defenders and Google’s own teams, Google says it is releasing Argon without cyber guardrails. The broader model is designed to refuse harmful cyber and chemical, biological, radiological, and nuclear requests, under Google’s Frontier Safety Framework. Google says it monitors internal activations for misuse, and that misalignment checks watch the chain of thought and the actions and can stop execution. Google says Argon is its most resilient model yet against indirect prompt injection, and that it is leading on Gray Swan’s Indirect Prompt Injection benchmark. The post does not give that score. Public tool names and an API method are not in the post.

Benchmarking

Artificial Analysis scores Gemini 4 Argon (High) at 53 on its Intelligence Index, fetched 1 October 2026. That is an independent score, not a Google chart number.

The cells below are Google’s 30 September post. They are lab claims.

The post names Vals, Harvey’s Legal Agent Benchmark, and Gray Swan without a number. Those stay off this card.

Access and pricing

Now: trusted cyber defenders through the Fairwind Program, plus Google’s own internal use. Not the public. No Gemini API model page was up on 1 October 2026. Broader access, when it comes, starts with paid API customers and Google AI Ultra subscribers. No date.

The post does not say when the introductory price ends.

Sources

• Gemini 4 Argon
• Fairwind Program
• Frontier Safety Framework
• Artificial Analysis, Gemini 4 Argon (High)