Rankings / Google DeepMind
Gemini 3 Pro Preview
released 2025-11-18
BenchAtlas Index
as of 2026-07-14
high effort
rank #45
10 families · 5 categories · high
base configuration
rank #63
10 families · 5 categories · high
high effort
rank #78
10 families · 5 categories · high
Benchmark evidence
47 results
agentic coding
| τ²-Bench | high effort | 87.1 % | independent | Artificial Analysis |
| τ²-Bench | low effort | 68.1 % | independent | Artificial Analysis |
coding
| LiveCodeBench v6 implementation=artificial-analysis | high effort | 91.7 % | independent | Artificial Analysis |
| LiveCodeBench v6 implementation=artificial-analysis | low effort | 85.7 % | independent | Artificial Analysis |
| SciCode | high effort | 56.1 % | independent | Artificial Analysis |
| SciCode | low effort | 49.9 % | independent | Artificial Analysis |
external indices
| Intelligence Index v4.1 | high effort | 39.6 points | independent | Artificial Analysis |
| Intelligence Index v4.1 | low effort | 33.1 points | independent | Artificial Analysis |
| AA Math Index | high effort | 95.7 points | independent | Artificial Analysis |
| AA Math Index | low effort | 86.7 points | independent | Artificial Analysis |
| ECI | 153.2 points | independent | Epoch AI Benchmarking Hub |
factuality
| MASK contamination=Potential contamination warning: This model was evaluated after the public release of MASK, allowing model builder access to the prompts and solutions. | 42.6 % 38.6–46.6 | independent | Scale Labs | |
| SimpleQA Verified | 72.9 % | independent | Epoch AI Benchmarking Hub 2025-12-08 |
knowledge science
| GPQA Diamond | 92.6 % | independent | Epoch AI Benchmarking Hub 2025-11-19 | |
| GPQA Diamond implementation=vals-ai | high effort | 91.7 % | independent | Vals AI |
| GPQA Diamond implementation=artificial-analysis | high effort | 90.8 % | independent | Artificial Analysis |
| GPQA Diamond implementation=artificial-analysis | low effort | 88.7 % | independent | Artificial Analysis |
| Humanity's Last Exam contamination=Potential contamination warning: This model was evaluated after the public release of HLE, allowing model builder access to the prompts and solutions. · implementation=scale | 37.5 % 35.6–39.4 | independent | Scale Labs | |
| Humanity's Last Exam implementation=artificial-analysis | high effort | 37.2 % | independent | Artificial Analysis |
| Humanity's Last Exam implementation=artificial-analysis | low effort | 27.6 % | independent | Artificial Analysis |
| MMLU-Pro implementation=vals-ai | high effort | 90.1 % | independent | Vals AI |
| MMLU-Pro implementation=artificial-analysis | high effort | 89.8 % | independent | Artificial Analysis |
| MMLU-Pro implementation=artificial-analysis | low effort | 89.5 % | independent | Artificial Analysis |
long context instruction
| AA-LCR | high effort | 70.7 % | independent | Artificial Analysis |
| AA-LCR | low effort | 67.3 % | independent | Artificial Analysis |
| IFBench | high effort | 70.4 % | independent | Artificial Analysis |
| IFBench | low effort | 49.7 % | independent | Artificial Analysis |
| MultiChallenge | 65.7 % 63.5–67.9 | independent | Scale Labs |
multimodal
| MMMU implementation=vals-ai | high effort | 87.5 % | independent | Vals AI |
| VISTA | 51.5 % 50.7–52.3 | independent | Scale Labs |
professional
| CaseLaw | high effort | 53.1 % | independent | Vals AI |
| CorpFin | high effort | 63.7 % | independent | Vals AI |
| LegalBench | high effort | 87.0 % | independent | Vals AI |
| MedQA | high effort | 96.0 % | independent | Vals AI |
| TaxEval | high effort | 72.6 % | independent | Vals AI |
reasoning math
| AIME implementation=vals-ai | high effort | 96.7 % | independent | Vals AI |
| AIME year=2025 · implementation=artificial-analysis | high effort | 95.7 % | independent | Artificial Analysis |
| AIME year=2025 · implementation=artificial-analysis | low effort | 86.7 % | independent | Artificial Analysis |
| ARC-AGI-1older version split=semi_private · model_type=CoT | 75.0 % | independent | ARC Prize Leaderboard | |
| ARC-AGI-2 split=semi_private · model_type=CoT | 31.1 % | independent | ARC Prize Leaderboard | |
| ARC-AGI-2 split=semi_private · model_type=CoT | 0.8 usd_per_task | independent | ARC Prize Leaderboard | |
| ARC-AGI-1older version split=semi_private · model_type=CoT | 0.5 usd_per_task | independent | ARC Prize Leaderboard | |
| EnigmaEval contamination=Potential contamination warning: This model was evaluated after the public release of EnigmaEval, allowing model builder access to the prompts and solutions. | 18.2 % 16.0–20.4 | independent | Scale Labs | |
| FrontierMath | 37.6 % | independent | Epoch AI Benchmarking Hub 2025-11-21 | |
| FrontierMath Tier 4older version | 18.8 % | independent | Epoch AI Benchmarking Hub 2025-11-21 | |
| MATH-500 implementation=vals-ai | high effort | 96.4 % | independent | Vals AI |
| OTIS Mock AIME 2024–2025 | 91.4 % | independent | Epoch AI Benchmarking Hub 2025-11-19 |
Agent + model results
systems, not bare-model scores
| agent + model live-SWE-agent + Gemini 3 Pro Preview | SWE-bench Verified | 77.4 % | unverified | SWE-bench Leaderboard |
| agent + model mini-SWE-agent + Gemini 3 Pro Preview | SWE-bench bash-only | 74.2 % | community | SWE-bench Leaderboard |
| agent + model mini-SWE-agent + Gemini 3 Pro Preview | SWE-bench Verified | 74.2 % | community | SWE-bench Leaderboard |
| agent + model Epoch Inspect harness + Gemini 3 Pro Preview | SWE-bench Verified | 72.9 % | independent | Epoch AI Benchmarking Hub |
| agent + model mini-SWE-agent + Gemini 3 Pro Preview | SWE-bench Verified | 69.6 % | unverified | SWE-bench Leaderboard |
| agent + model mini-SWE-agent + Gemini 3 Pro Preview | SWE-bench bash-only | 69.6 % | unverified | SWE-bench Leaderboard |
| agent + model mini-SWE-agent + Gemini 3 Pro Preview | SWE-bench Verified | 1.0 usd_per_task | unverified | SWE-bench Leaderboard |
| agent + model mini-SWE-agent + Gemini 3 Pro Preview | SWE-bench bash-only | 1.0 usd_per_task | unverified | SWE-bench Leaderboard |
| agent + model mini-SWE-agent + Gemini 3 Pro Preview | SWE-bench Verified | 0.5 usd_per_task | community | SWE-bench Leaderboard |
| agent + model mini-SWE-agent + Gemini 3 Pro Preview | SWE-bench bash-only | 0.5 usd_per_task | community | SWE-bench Leaderboard |
| agent + model terminus-2 + Gemini 3 Pro Preview | Terminal-Bench 2.1 | 74.4 % | community | Terminal-Bench Leaderboard |
| agent + model ante + Gemini 3 Pro Preview | Terminal-Bench 2.0 | 69.4 % | community | Terminal-Bench Leaderboard |
| agent + model gemini-cli + Gemini 3 Pro Preview | Terminal-Bench 2.1 | 66.3 % | community | Terminal-Bench Leaderboard |
| agent + model sage + Gemini 3 Pro Preview | Terminal-Bench 2.0 | 65.2 % | unverified | Terminal-Bench Leaderboard |
| agent + model ruley + Gemini 3 Pro Preview | Terminal-Bench 2.0 | 62.3 % | unverified | Terminal-Bench Leaderboard |
| agent + model ii-agent-simple + Gemini 3 Pro Preview | Terminal-Bench 2.0 | 61.8 % | unverified | Terminal-Bench Leaderboard |
| agent + model Droid + Gemini 3 Pro Preview | Terminal-Bench 2.0 | 61.1 % | unverified | Terminal-Bench Leaderboard |
| agent + model terminus-2 + Gemini 3 Pro Preview | Terminal-Bench 2.0 | 56.9 % | community | Terminal-Bench Leaderboard |
| agent + model letta-code + Gemini 3 Pro Preview | Terminal-Bench 2.0 | 56.0 % | unverified | Terminal-Bench Leaderboard |
| agent + model Artificial Analysis harness + Gemini 3 Pro Preview | Terminal-Bench Hard | 41.7 % | independent | Artificial Analysis |
| agent + model Artificial Analysis harness + Gemini 3 Pro Preview | Terminal-Bench Hard | 34.1 % | independent | Artificial Analysis |
These scores measure the whole agent system (scaffold, tools, budgets) — they are never merged into the bare model’s numbers.