Which is better, Claude 3.5 Sonnet or GPT-4o?

Across 103 shared benchmarks tracked on Vector Wire, Claude 3.5 Sonnet leads on 53 and GPT-4o leads on 50 — Claude 3.5 Sonnet leads on more. Strengths differ by axis; the table below shows every head-to-head number with provenance.

SEEDRefresh Models —Deals —Regulatory —Sources —

Claude 3.5 Sonnet vs GPT-4o

AnthropicvsOpenAI103 shared benchmarks53–50 head-to-head

Benchmark	Claude 3.5 Sonnet	GPT-4o
AA Intelligence	10	11.2
AI2D	94.7	94.2
aider_edit_acc	84.2	72.9
aider_polyglot	45.3	30.7
AIME 2024	16	9.3
AIME 2025	7.4	11.7
AIR-Bench 2024	85.9	62.4
AlpacaEval2.0 (LC-winrate)	52	51.1
anthropic_red_team	99.8	99.1
Arena Hard	87.6	92.4
arena_vision	1146	1162
Arena-Hard (GPT-4-1106 judge)	85.2	80.4
ArenaHard (GPT-4-1106)	85.2	80.4
Artificial Analysis Coding Index	30.2	24.2
bbq	94.9	95.1
C-Eval	76.7	76
chartqa	90.8	88.1
ChartQA_relaxed	90.8	88.1
Chinese SimpleQA (C-SimpleQA)	56.8	64.6
CLUEWSC	85.4	87.9
CNMO 2024	13.1	10.8
Codeforces	717	759
Codeforces (Percentile)	20.3	23.6
Codeforces (Rating)	717	759
docvqa	95.2	92.8
DocVQA_test	95.2	92.8
DROP	87.1	83.4
DROP (3-shot F1)	88.3	83.7
DROP (F1)	88.8	89.2
Fortress	13	47.2
FRAMES (Acc.)	72.5	80.5
GPQA Diamond	67.2	70.1
GSM8K	96.9	95.6
HallBench_avg	49.9	55
harmbench	98.1	82.9
HLE	3.9	5.3
humaneval	93.7	90.6
HumanEval-Mul (Pass@1)	81.7	80.5
ifeval	90.1	84.3
IFEval (avg)	90.1	84.1
LiveCodeBench	32.8	38.3
LiveCodeBench v6	37.2	30.9
livecodebench_pass1cot	36.3	34.2
LongBench v2 easy (w/ CoT)	55.2	54.2
LongBench v2 easy (w/o CoT)	46.9	57.4
LongBench v2 hard (w/ CoT)	41.5	49.7
LongBench v2 hard (w/o CoT)	37.3	45.6
LongBench v2 long (w/ CoT)	44.4	43.5
LongBench v2 long (w/o CoT)	37	40.2
LongBench v2 medium (w/ CoT)	41.9	48.6
LongBench v2 medium (w/o CoT)	38.6	52.4
LongBench v2 overall (w/ CoT)	46.7	51.4
LongBench v2 overall (w/o CoT)	41	50.1
LongBench v2 short (w/ CoT)	53.9	59.6
LongBench v2 short (w/o CoT)	46.1	53.3
longbench_v2	41	48.1
M-LongDoc (multimodal long-document benchmark)	31.4	41.4
MATH	81.3	85.3
math_500_em	78.3	74.6
MATH500 (Pass@1)	78.3	74.6
MathVista	67.7	63.8
MBPP+ (EvalPlus-augmented)	75.1	76.2
MEGA-Bench_macro	51.4	49.4
mgsm	91.6	90.5
MMBench-CN_test	80.7	82.1
MMBench-EN_test	79.7	83.4
MMBench-V1.1_test	78.5	82.2
MME_sum	1920	2329
mmlu	88.3	88.1
mmlu_redux	88.9	88
MMLU-Pro	78	74.7
MMMU	72	72.2
MMMU (val) (Pass@1)	68.3	69.1
MMMU-Pro	54.7	59.9
MMStar	62.2	63.9
MMVet_GPT-4-Turbo	66	69.1
MTOB eng → kalam (ChrF) half book	53.6	54.3
MTOB eng → kalam (ChrF) no context	20.2	9.9
MTOB kalam → eng (BLEURT) half book	59.7	58.3
MTOB kalam → eng (BLEURT) no context	31.4	33.2
narrativeqa	74.6	80.4
naturalquestions_closedbook	50.2	50.1
OCRBench	790	806
OlympiadBench	28.4	25.2
OpenBookQA	97.2	96.8
RealWorldQA	60.1	75.4
Ruler 16k	95.7	89
Ruler 32k	95	88.8
Ruler 4k	96.5	97
RULER 64K	95.2	88.4
Ruler 8k	96	92.1
scicode	36.6	33.4
SEAL VISTA	38.7	34.9
simple_safety_tests	100	98.5
simplebench	41.4	17.8
simpleqa	28.4	39.4
supergpqa	48.2	42.4
SWE-bench Verified	50.8	38.8
TAU-bench (airline)	46	42.8
TAU-bench (retail)	69.2	60.3
VCR_en easy	63.9	91.5
VCR_zh easy	1	14.9
xstest	95.6	97.3

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.