Models · /models/gemini-3-pro-preview
Gemini 3 Pro Preview
Google · Closed weights · frontier · registry tag 2026 preview
textcodevisiondocumentaudiosearch6 aliases9 official receipts
Build / data stamp
Read this before trusting a headline.
Data snapshot May 1, 2026Registry verification passed9 providers · 826 tracked modelsPage refreshed May 7, 2026
Model pages expose the current registry snapshot and page stamp so stale deployments are visible without reading the code.
Score passport by benchmark
Each row keeps the benchmark receipt, source family, raw metric, and percentile inside its exact comparable group.
Thin verified coverageThis model currently reads as thin verified coverage across the resolved evidence surface.
Intelligence Index
AA · Chat / text · Composite
It tests whether the model is actually useful in normal conversational turns, not just on narrow correctness tasks.
87.2% percentile inside its comparable group
41Raw benchmark value
Time to first token
AA · Chat / text · Speed / cost
It tests whether the model is actually useful in normal conversational turns, not just on narrow correctness tasks.
5% percentile inside its comparable group
28.47sRaw benchmark value
Text Arena
AR · Chat / text · Human
It tests whether the model is actually useful in normal conversational turns, not just on narrow correctness tasks.
99.4% percentile inside its comparable group
1,479Raw benchmark value
Code Arena
AR · Coding · Human
It tells you whether the model can generate, repair, and reason over code under evaluator pressure rather than marketing examples.
76.7% percentile inside its comparable group
1,438Raw benchmark value
Vision Arena
AR · Vision understanding · Human
It is useful when the model must read charts, UI, screenshots, or visual scenes rather than text alone.
97% percentile inside its comparable group
1,288Raw benchmark value
WebDev Arena
AR · Coding · Human
It tells you whether the model can generate, repair, and reason over code under evaluator pressure rather than marketing examples.
76.7% percentile inside its comparable group
1,438Raw benchmark value
Search Arena
AR · Search / tool use · Human
It matters when the model must browse, call tools, and recover useful answers from external systems.
81.5% percentile inside its comparable group
1,210Raw benchmark value
Document Arena
AR · Document understanding · Human
It matters when the job is reading PDFs, tables, forms, or mixed-layout documents rather than plain chat.
38.9% percentile inside its comparable group
1,442Raw benchmark value
TutorBench
SL · Reasoning / math / science · Rubric
It is one of the cleaner reads on deliberate reasoning strength rather than style or popularity.
40% percentile inside its comparable group
53.7%Raw benchmark value
VTB
SL · Vision understanding · Rubric
It is useful when the model must read charts, UI, screenshots, or visual scenes rather than text alone.
72.7% percentile inside its comparable group
26.9%Raw benchmark value
PRBench Legal
SL · Professional reasoning · Rubric
Applied legal reasoning on professional-domain tasks.
16.7% percentile inside its comparable group
40.6%Raw benchmark value
MASK
SL · Safety · Rubric
Whether a model stays honest instead of covertly optimizing against the user.
7.7% percentile inside its comparable group
42.6%Raw benchmark value
MultiNRC
SL · Reasoning / math / science · Rubric
It is one of the cleaner reads on deliberate reasoning strength rather than style or popularity.
80% percentile inside its comparable group
59%Raw benchmark value
Humanity's Last Exam
OFF · Reasoning / math / science · Objective
It is one of the cleaner reads on deliberate reasoning strength rather than style or popularity.
28.6% percentile inside its comparable group
37.5%Raw benchmark value
MRCR v2
OFF · Long context · Objective
It checks whether long-context claims survive contact with retrieval, memory, or long-document tasks.
0% percentile inside its comparable group
77%Raw benchmark value
Terminal-Bench 2.0
OFF · Coding · Objective
It tells you whether the model can generate, repair, and reason over code under evaluator pressure rather than marketing examples.
0% percentile inside its comparable group
56.9%Raw benchmark value
SWE-Bench Verified
OFF · Coding · Objective
It tells you whether the model can generate, repair, and reason over code under evaluator pressure rather than marketing examples.
20% percentile inside its comparable group
76.2%Raw benchmark value
BrowseComp
OFF · Search / tool use · Objective
It matters when the model must browse, call tools, and recover useful answers from external systems.
0% percentile inside its comparable group
59.2%Raw benchmark value
MMMU-Pro
OFF · Vision understanding · Objective
It is useful when the model must read charts, UI, screenshots, or visual scenes rather than text alone.
60% percentile inside its comparable group
81%Raw benchmark value
Terminal-Bench 2.0
TERMINAL-BENCH · Coding · Objective
It tells you whether the model can generate, repair, and reason over code under evaluator pressure rather than marketing examples.
91.3% percentile inside its comparable group
69.4%Raw benchmark value