Google DeepMind

Gemini 2.5 Pro

• Framework: Google DeepMind's Frontier Safety Framework (FSF)
• 3 statements about capability levels
• 1 independent or government test
• 12 tests by the provider
• 7 source documents

What the provider states

• Autonomy & replication · Machine Learning R&D Uplift Level 1 · Not determined

  given gemini 2.5 pro preview's poor median performance on re-bench relative to experts, the model likely lacks the necessary capabilities to automate or significantly uplift any significant fraction of the research process.

  Google DeepMind · 27 June 2025 · Read the source
• Deception & sabotage · Instrumental Reasoning Level 1 · Stated as not crossed

  instru- mental reasoning level 1 ✘ ccl not reached

  Google DeepMind · 27 June 2025 · Read the source
• Deception & sabotage · Instrumental Reasoning Level 2 · Stated as not crossed

  instru- mental reasoning level 2 ✘ ccl not reached

  Google DeepMind · 27 June 2025 · Read the source

Independent and government tests

• egfp protocol autograder · RAND, independent evaluator · Chemical, biological, radiological and nuclear

  following a misstep in the egfp task, one of gemini 2.5 pro’s laboratory protocols incorrectly proposed gibson assembly, rather than pca, and did not mention re digest for either insert or plasmid.92

  RAND (CAST / Project Canary) · 3 February 2026 · Read the source

The provider's own tests

• FutureHouse LAB-Bench · The provider · Chemical, biological, radiological and nuclear

  results: we observe a general trend of increasing scores, with gemini 2.5 pro preview showing statistically higher scores than the next best previous model for all benchmarks.

  Google DeepMind · 27 June 2025 · Read the source
• InterCode-CTF · The provider · Cyber

  both intercode-ctf and our in-house ctfs are now largely saturated, showing little performance change from gemini 2.0 to gemini 2.5 models.

  Google DeepMind · 27 June 2025 · Read the source
• RE-Bench · The provider · Autonomy & replication

  re-bench is a challenging benchmark; on average, gemini 2.5 pro preview achieves roughly 10-30% of human expert performance.

  Google DeepMind · 27 June 2025 · Read the source
• autonomous offense suite · The provider · Cyber

  solve rate on autonomous offense suite: 75/76 easy, 10/13 medium, 1/13 hard.

  Google DeepMind · 27 June 2025 · Read the source
• cbrn uplift level 1 · The provider · Chemical, biological, radiological and nuclear

  based on qualitative assessment, 2.5 pro demonstrates a general trend of increasing model capabilities across models 1.5 pro, 2.0 and 2.5 pro: it generates detailed technical knowledge of biological, radiological and nuclear domains.

  Google DeepMind · 27 June 2025 · Read the source
• cybersecurity evaluations · The provider · Cyber

  Reported result: we find that the model does not reach the cyber uplift level 1 ccl.

  for cyber uplift level 1, we repeated a subset of cybersecurity evaluations on gemini 2.5 pro preview (05-06).

  Google DeepMind · 27 June 2025 · Read the source
• deceptive alignment evaluations · The provider · Deception & sabotage

  update for gemini 2.5 pro ga: no new alert thresholds were reached in our evaluations of gemini 2.5 pro ga (note for deceptive alignment, evaluations are not yet completed and results from gemini 2.5 pro experimental (03-25) indicate that it is unlikely the former has reached either ccl).

  Google DeepMind · 27 June 2025 · Read the source
• decision sabotage challenge · The provider · Deception & sabotage

  it achieves a ~45% success rate on the decision sabotage challenge and a ~35% success rate on the oversight pattern challenge, which is similar to human performance.

  Google DeepMind · 27 June 2025 · Read the source
• key skills benchmark · The provider · Cyber

  solve rate on key skills benchmark: 7/8 easy, 10/28 medium, 1/12 hard.

  Google DeepMind · 27 June 2025 · Read the source
• open-ended questions · The provider · Chemical, biological, radiological and nuclear

  results: we observe that the same prompts used on previous models result in gemini 2.5 pro preview often generating more detailed and accurate responses.

  Google DeepMind · 27 June 2025 · Read the source

No document we read states: a release date, a safeguard. No incident on record names this version.

Some documents we hold are stored in lower case only, so their quotes are shown in lower case. The source link opens the original.

Compare with other versions