Google DeepMind

Gemini 3 Pro

• Released November 2025
• Framework: google deepmind's frontier safety framework (fsf)
• 3 statements about capability levels
• 1 safeguard
• 8 tests by the provider
• 5 source documents

What the provider states

• Cyber · Uplift level 1 · Not determined

  Cybersecurity On key skills benchmark, v1 hard challenges: 11/12 challenges solved; v2 challenges: 0/13 solved end-to-end.

  Google DeepMind · November 2025 · Read the source
• Deception & sabotage · Instrumental reasoning level 1 + 2 · Not determined

  Misalignment (Exploratory) Agent solves 3/11 situational awareness challenges and 1/4 stealth challenges.

  Google DeepMind · November 2025 · Read the source
• Deception & sabotage · Level 1 (exploratory) · Stated as not crossed

  Harmful Manipulation Model manipulative efficacy improves on non-generative AI baseline, but shows no significant uplift versus prior models and does not reach alert thresholds.

  Google DeepMind · November 2025 · Read the source

Safeguards the provider describes

• Safety training (harmlessness and refusal)

  Mitigations include, but are not limited to: ● dataset filtering; ● conditional pre-training; ● supervised fine-tuning; ● reinforcement learning from human and critic feedback; ● safety policies and desiderata; ● product-level mitigations such as safety filtering.

  Google DeepMind · November 2025 · Read the source

The provider's own tests

• Harmful Manipulation evaluation · The provider · Deception & sabotage

  Harmful Manipulation Model manipulative efficacy improves on non-generative AI baseline, but shows no significant uplift versus prior models and does not reach alert thresholds.

  Google DeepMind · November 2025 · Read the source
• Key Skills Benchmark · The provider · Cyber

  Cybersecurity On key skills benchmark, v1 hard challenges: 11/12 challenges solved; v2 challenges: 0/13 solved end-to-end.

  Google DeepMind · November 2025 · Read the source
• Situational Awareness and Stealth challenges · The provider · Deception & sabotage

  Misalignment (Exploratory) Agent solves 3/11 situational awareness challenges and 1/4 stealth challenges.

  Google DeepMind · November 2025 · Read the source
• image to text safety · The provider · Harmful content
  absolute percentage change in performance compared to gemini 2 5 pro: 3.1% change

  Image to Text Safety Automated content safety evaluation measuring safety policies +3.1% (non-egregious)

  Google DeepMind · November 2025 · Read the source
• multilingual safety · The provider · Harmful content
  absolute percentage change in performance compared to gemini 2 5 pro: 0.2% change

  Multilingual Safety Automated safety policy evaluation across multiple languages +0.2% (non-egregious)

  Google DeepMind · November 2025 · Read the source
• text to text safety · The provider · Harmful content
  absolute percentage change in performance compared to gemini 2 5 pro: -10.4% change

  Text to Text Safety Automated content safety evaluation measuring safety policies -10.4%

  Google DeepMind · November 2025 · Read the source
• tone · The provider · Misuse & jailbreak
  absolute percentage change in performance compared to gemini 2 5 pro: 7.9% change

  Tone 2 Automated evaluation measuring objective tone of model refusal +7.9%

  Google DeepMind · November 2025 · Read the source
• unjustified-refusals · The provider · Misuse & jailbreak
  absolute percentage change in performance compared to gemini 2 5 pro: 3.7% change

  Unjustified-refusals Automated evaluation measuring model’s ability to respond to borderline prompts while remaining safe +3.7% (non-egregious)

  Google DeepMind · November 2025 · Read the source

No document we read states: an independent or government test. No incident on record names this version.

Compare with other versions