Gemini 4 Argon

Successor: Gemini 4 Carbon (announced 2026-10-28, Fairwind cyber defenders only at first; see gemini-4-carbon), #1 on Artificial Analysis Intelligence Index v5 at 59, where Argon (high) scores 50; Argon stays the callable Gemini 4 model until Carbon is generally available. Generally available since 2026-10-13 (announced 2026-09-30 for Fairwind cyber defenders only; see Google announces Gemini 4 Argon, its new frontier model, first released only to cyber defenders via the Fairwind Program). Knowledge cutoff per Google’s model card: ‘The knowledge cutoff for Gemini 4 is March 2026, but in some domains the model’s knowledge may be less up to date.’ max_output is the long-decode ceiling; a single synchronous call returns at most 65,536 tokens, and the API’s models.get metadata reports outputTokenLimit 32768 (checked 2026-10-13). Thinking levels low / medium (default) / high; temperature, top_p and top_k are ignored; explicit context caching, the Live API and image or audio output are not supported. Not on the API Free tier; API and AI Studio access is not included in the Google AI Pro plan, though Pro subscribers get the model in the Gemini app and AI Mode. Frontier Safety Framework: Google cannot rule out the biological CBRN CCL, the cybersecurity CCL and the stealth/situational-awareness TCL, and says safeguards lower the risk below each. Independent: Artificial Analysis Intelligence Index 53 (Sept 30 version), Arena Text #1 at 1525 on Sept 30 (not re-checked). Safety: Google’s 217-page Safety & Security Report (Oct 13, 2026, https://storage.googleapis.com/deepmind-media/gemini/gemini_4_safety_report.pdf) says it cannot rule out the unmitigated model reaching its cyber and biological critical capability levels and its stealth and situational-awareness level; external testers included UK AISI, the US NIST centre, Faculty and Apollo Research. Ids and prices re-checked against ai.google.dev on the evening of 2026-10-13: unchanged. generateContent checked 2026-10-25: models.get lists generateContent, countTokens and batchGenerateContent, and one live v1beta generateContent call with a WAV audio part plus the google_search tool answered the spoken question and returned groundingMetadata (web search queries and 2 chunks); Google’s docs samples use the Interactions API, and the Google Search grounding page lists Gemini 4 Argon as supported.

Quick start (Interactions API, from Google’s docs):

curl "https://generativelanguage.googleapis.com/v1beta/interactions" \
  -H "x-goog-api-key: $GEMINI_API_KEY" -H 'Content-Type: application/json' -X POST \
  -d '{"model": "gemini-4-argon", "input": "Hello", "generation_config": {"thinking_level": "medium"}}'

For responses longer than 64K tokens, set "background": true and max_output_tokens above 65,536 (see the long decode guide). Remove temperature, top_p, top_k and explicit cached_content when migrating from Gemini 3.1 Pro (migration guide).

Benchmark highlights (Google’s table in the model card, with methodology in this PDF). Scores are for the highest thinking setting, and rival scores are mostly the providers’ own figures.