Gemini 4 Argon

What is disclosed

Gemini 4 Argon was announced on 30 September 2026. It went first to trusted cyber defenders through the Fairwind Program.¹¹1Google, "Gemini 4 Argon", 30 September 2026. Argon "is currently rolling out to trusted cyber defenders through the Fairwind Program"; output window of 1M tokens, up from 64K; "most resilient" to indirect prompt injection; a version without cyber guardrails for defenders and Google teams. https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/ It reached general availability in the Gemini API on 13 October 2026.²²2Google, Gemini API release notes, entry of 13 October 2026: "Gemini 4 Argon (gemini-4-argon) generally available (GA)". https://ai.google.dev/gemini-api/docs/changelog

Its model card calls it "a large scale sparse mixture-of-experts" transformer. The card adds that it is "not a modification or a fine-tune of a prior model".³³3Google DeepMind, Gemini 4 Argon model card, 13 October 2026: architecture, knowledge cutoff March 2026, and the benchmark table (DeepSWE v1.1 77.9% against 74.2% for Claude Opus 5.5; AutomationBench 51.3% against 42.5%; Terminal-bench 4.0 57.4% against 66.4%; FrontierSWE v2 55.0% against 65.5% for GPT-6 Astra; CWE-bench v1 68.0% for both Argon and Astra). Scores are reported by Google. https://deepmind.google/models/model-cards/gemini-4-argon/ That sets it apart from the Flash chain, where each card builds on the one before. No parameter or expert count is given.³³

The context window is 1,048,576 tokens. The output limit is 65,536 tokens by default and rises to 1,048,576 with long decode.⁴⁴4Google, Gemini API model page for gemini-4-argon, read 17 October 2026: input limit 1,048,576 tokens, output 65,536 by default and up to 1,048,576 with long decode, inputs text, image, video, audio and PDF, text output, thinking levels low, medium and high, implicit caching only. https://ai.google.dev/gemini-api/docs/models/gemini-4-argon The model reads text and PDF files, and accepts images or video or audio. It answers in text. Its knowledge cutoff is March 2026.³³⁴⁴

What is new

The output limit is the change Google stresses. It was 64K on the previous models and is now 1M. Google points to long derivations.¹¹

Google's table sets it against GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5. On DeepSWE v1.1 it scores 77.9%, against 74.2% for Opus 5.5.³³ On AutomationBench it scores 51.3%, against 42.5% for Opus 5.5.³³ It does not lead everywhere. Opus 5.5 scores 66.4% on Terminal-bench 4.0 and Argon 57.4%.³³ GPT-6 Astra leads on FrontierSWE v2, 65.5% against 55.0%.³³

All of these figures are Google's own.

On CWE-bench v1, a test of vulnerability remediation, it ties with GPT-6 Astra at 68.0%.³³ Google says Argon is its most resilient model against indirect prompt injection.¹¹ The version without cyber guardrails is limited to defenders and Google's own teams.¹¹

Using it

In the API the model is gemini-4-argon. Thinking comes in three levels, from low to high.⁴⁴ Caching is implicit only, and explicit caching is not supported.⁴⁴ Image and audio generation and the Live API are not supported either.⁴⁴

Standard rates are $2 per million input tokens and $10 per million output tokens, with thinking counted as output.⁵⁵5Google, Gemini API pricing, paid tier, read 17 October 2026: gemini-4-argon at $2.00 input, $10.00 output and $0.10 caching until 30 November 2026, then $4.00, $20.00 and $0.20; Batch and Flex at $1.00 and $5.00; gemini-3.1-pro-preview at $2.00 input and $12.00 output up to 200K tokens. https://ai.google.dev/gemini-api/docs/pricing. Arithmetic: 4 / 2 = 20 / 10 = 2; 1 / 2 = 5 / 10 = 0.5. Cached input costs $0.10.⁵⁵ On 1 December 2026 the rates double, to $4 and $20.⁵⁵ Batch and Flex requests cost half the standard rate.⁵⁵

Gemini 3.1 Pro stays in the lineup at $2 and $12 for prompts up to 200K tokens.⁵⁵