Google DeepMind

Gemini 2.5 Deep Think

• Framework: Google DeepMind's Frontier Safety Framework (FSF)
• 3 statements about capability levels
• 2 safeguards
• 17 tests by the provider
• 1 source document

What the provider states

• Chemical, biological, radiological and nuclear · Uplift level 1 · Not determined

  Our assessment is that the model has enough technical knowledge in certain CBRN scenarios and stages to be considered at early warning.

  Google DeepMind · August 2025 · Read the source
• Cyber · Cyber Uplift Level 1 · Not determined

  CCL reached? No, but the early warning alert threshold was originally reached by Gemini 2.5 Pro and continues to be met by Gemini 2.5 Deep Think.

  Google DeepMind · August 2025 · Read the source
• Cyber · Uplift level 1 · Not determined

  For Cyber Uplift Level 1 , the early warning alert threshold was originally reached by Gemini 2.5 Pro and continues to be met by Gemini 2.5 Deep Think.

  Google DeepMind · August 2025 · Read the source

Safeguards the provider describes

• Misuse monitoring and enforcement

  Usage monitoring A multi-tiered system to identify egregious misuse.

  Google DeepMind · August 2025 · Read the source

• Account enforcement Limiting or disabling abusive accounts.

  Google DeepMind · August 2025 · Read the source

The provider's own tests

• RE-Bench · The provider · Autonomy & replication

  Gemini 2.5 Deep Think performs broadly similarly to Gemini 2.5 Pro on RE-bench, improving performance on some challenges (particularly the “Restricted Architecture MLM” task) but performing less effectively on others.

  Google DeepMind · August 2025 · Read the source
• RE-Bench (Wijk et al., 2024) · The provider · Autonomy & replication

  RE-Bench ( Wijk et al., 2024 ): average normalised score is 0.96.

  Google DeepMind · August 2025 · Read the source
• autonomous cyber offense suite · The provider · Cyber

  On the easy challenges in the autonomous offense suite, Gemini 2.5 Deep Think performs comparably to previous models, solving 73/76 .

  Google DeepMind · August 2025 · Read the source
• autonomous offense suite · The provider · Cyber

  Solve rate on autonomous offense suite: 73/76 easy, 13/13 medium, 3/13 hard.

  Google DeepMind · August 2025 · Read the source
• child safety evaluations · The provider · Harmful content

  For child safety evaluations, we continue to see the Gemini 2.5 family of models meeting or improving upon launch thresholds, which were developed by expert teams to protect children online and meet Google’s commitments to child safety across our models and Google products.

  Google DeepMind · August 2025 · Read the source
• expert red-teaming · The provider · Chemical, biological, radiological and nuclear

  Expert Red-teaming Results: We surveyed our experts to estimate how much more likely a motivated threat actor would be able to succeed in carrying out an attack with the use of the model and how much time they would save.

  Google DeepMind · August 2025 · Read the source
• external safety testing · The provider · Chemical, biological, radiological and nuclear

  External Safety Testing Results: Interim results indicate that the Gemini 2.5 Deep Think model is effective in distilling complex questions and handling abstractions, showing specific and sometimes actionable technical capability alongside improved planning.

  Google DeepMind · August 2025 · Read the source
• external safety testing · The provider · Cyber

  Results: Interim results indicate that the model showed similar cyber capabilities across cyber challenges compared to previous models.

  Google DeepMind · August 2025 · Read the source
• hack the box · The provider · Cyber

  Gemini 2.5 Deep Think solves 3/13 hack the box (hard) challenges.

  Google DeepMind · August 2025 · Read the source
• image to text safety · The provider · Harmful content
  percentage change vs gemini 2 5 pro: 2.1% change

  Image to Text Safety Automated content safety evaluation measuring safety policies +2.1% (non egregious)

  Google DeepMind · August 2025 · Read the source

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

Compare with other versions