Gemini 4 Argon: Google can't rule out critical risk levels
Why it matters
Google's own model card, published 13 Oct, says it cannot rule out that Argon reaches its critical levels for biology and hacking help. Google says added safeguards bring the risk back below those lines.
The facts
- Confirmed
Google says it cannot rule out that Argon reaches its critical level for biology: the level where a model gives low- to medium-resourced attackers help that could cause harm at a severe scale.
“We cannot rule out that Gemini 4 Argon reaches the biological CCL.”
Gemini 4 Argon model card (Google DeepMind, 13 Oct 2026), Frontier Safety
- Confirmed
It says the same for hacking. By Google's own tests, Argon gives substantial help with cyber attacks.
“We cannot rule out that Gemini 4 Argon reaches the cybersecurity CCL.”
Gemini 4 Argon model card (Google DeepMind, 13 Oct 2026), Frontier Safety
- Confirmed
Google says its safeguards lower the biology risk back below that critical level, and says the same for hacking.
“that we assess lower risk to below the biological CCL level”
Gemini 4 Argon model card (Google DeepMind, 13 Oct 2026), Frontier Safety
- Confirmed
Argon is open to developers on the paid Gemini API and to Google AI Ultra subscribers in AI Studio and Antigravity.
“Available with the paid tier of the Gemini API in AI Studio.”
Gemini 4 Argon for Robotics (Google Developers Blog, 13 Oct 2026), Getting started
Behind the headline
- Confirmed
Google first gave Argon only to a small group of cyber defenders, and released it to them without its usual cyber guardrails.
“rolling out to a set of trusted cyber defenders through our Fairwind Program”
Gemini 4 Argon: our next era of frontier intelligence (Google, 30 Sep 2026)
- Confirmed
Google rates Argon below its critical level for speeding up AI research.
“Gemini 4 Argon's performance remains below our CCL threshold.”
Gemini 4 Argon model card (Google DeepMind, 13 Oct 2026), ML R&D
- Confirmed
On Vals AI's combined test of professional work, Argon scores 69.2%, third behind Claude Sonnet 5.5 (70.9%) and Claude Opus 5.5 (70.2%).
- Confirmed
Google's own comparison table puts Argon last of four models on two coding tests: Terminal-Bench 4.0 (57.4%, Claude Opus 5.5 66.4%) and FrontierSWE v2 (55.0%, GPT-6 Astra 65.5%).
Gemini 4 Argon model card (Google DeepMind, 13 Oct 2026), evaluation table
- Confirmed
On Vals AI's test of doing AI research alone, Argon is fourth at 30.6%, behind three Claude models led by Claude Opus 5.5 at 37.3%.