Which is better, Claude Sonnet 4.5 or Gemini 3 Pro?

Across 77 shared benchmarks tracked on Vector Wire, Claude Sonnet 4.5 leads on 16 and Gemini 3 Pro leads on 60 — Gemini 3 Pro leads on more. Strengths differ by axis; the table below shows every head-to-head number with provenance.

SEEDRefresh Models —Deals —Regulatory —Sources —

Claude Sonnet 4.5 vs Gemini 3 Pro

AnthropicvsGoogle77 shared benchmarks16–60 head-to-head

Benchmark	Claude Sonnet 4.5	Gemini 3 Pro
AA Agentic Index	50.6	52
AA Intelligence	37.4	40.6
AA-LCR	68.3	73
AA-Omniscience	-0.1	15.3
AIME 2025	88	100
AIME25 no tools	88	96
arc_agi_1	63.7	75
ARC-AGI-2	13.6	54
ARC-AGI-2 (Verified)	13.6	31.1
arena_elo	1456	1485
arena_text_factuality	1448	1481
Arena-Hard (Creative Writing)	76.7	93.6
Arena-Hard (Hard Prompt)	63.3	72.6
Artificial Analysis Coding Index	52.1	46.5
browsecomp	24.1	59.2
BrowseComp (context management)	26.1	67.6
browsecomp_zh	42.4	66.8
charxiv_rq	67.2	81.4
coding_arena_elo	1392	1438
critpt	1.1	9.1
Fortress	17	41.7
frontiermath_tier_4	4.2	18.8
gdpval	40.4	34.2
GPQA Diamond	83.4	91.9
HLE	19.8	45.8
HLE (with tools)	32	45.8
hmmt_feb_2025	79.2	97.5
hmmt_nov_2025	81.7	93.3
IFBench	57.3	70.4
imo_answer_bench	65.9	83.3
LiveCodeBench	71	92
LiveCodeBench v6	64	90.7
longbench_v2	61.8	68.2
matharena_visual_math_overall	75.8	84.2
MathVista	79.8	89.8
MCP Atlas	59.5	70.3
MMLU-Pro	88.2	90.1
mmmlu	89.1	91.8
MMMU-Pro	68.7	81
mrcr	55.4	89.7
Multi-SWE-Bench	44.3	42.7
multichallenge	55.3	65.7
MultiNRC	35.8	59
OctoCodingbench	22.8	22.9
OJ-Bench (cpp)	30.4	68.5
OmniDocBench 1.5	85.8	88.5
OmniScience Accuracy	32.9	55.8
OmniScience Non-Hallucination	50.9	10
scicode	45	56.1
SEAL VISTA	48.8	51.5
Seal-0	53.4	45.5
simplebench	54.3	76.4
simpleqa_verified	23.6	72.9
SimpleVQA	57.6	69.7
swe_bench_bash	71.4	74.2
SWE-bench Multilingual	68	68.7
SWE-bench Pro	43.6	43.3
SWE-bench Verified	82	78
SWE-Perf	3	6.5
SWT-bench	69.5	79.7
Terminal-Bench 2.0	50	54.2
Terminal-Bench Hard	35.6	56.9
toolathlon	41	36.4
vectara_answer_rate	95.6	99.4
vectara_avg_summary_length	127.8	101.9
vectara_factual_consistency	88	86.4
vectara_hallucination_rate ↓	12	13.6
VIBE (Average)	85.2	82.4
VIBE-Android	87.5	78.7
VIBE-Backend	90.8	78.7
VIBE-iOS	81.2	75.8
VIBE-Simulation	79.1	89.2
VIBE-Web	87.3	89.5
VisualToolBench	6.2	26.9
τ²-Bench	98	98
τ²-Bench (Retail)	86.2	85.3
τ²-Bench Telecom (AA run)	78.1	98

Best tracked score per model per benchmark (default configuration; source-attributed). ↓ marks lower-is-better metrics. Open either model for its full surface, provenance and pricing.