Which is better, Claude Sonnet 4.5 or GPT-5.2?

Across 64 shared benchmarks tracked on Vector Wire, Claude Sonnet 4.5 leads on 15 and GPT-5.2 leads on 49 — GPT-5.2 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 GPT-5.2

AnthropicvsOpenAI64 shared benchmarks15–49 head-to-head

Benchmark	Claude Sonnet 4.5	GPT-5.2
AA Agentic Index	50.6	60.2
AA Intelligence	37.4	43.3
AA-LCR	68.3	79.3
AA-Omniscience	-0.1	0.3
AIME 2025	88	100
AIME25 no tools	88	98
arc_agi_1	63.7	86.2
ARC-AGI-2	13.6	52.9
arena_elo	1456	1437
arena_text_factuality	1448	1442
Artificial Analysis Coding Index	52.1	48.7
browsecomp	24.1	65.8
BrowseComp (context management)	26.1	70
browsecomp_zh	42.4	76.1
charxiv_rq	67.2	82.1
coding_arena_elo	1392	1418
critpt	1.1	11.6
Fortress	17	17.5
frontiermath_tier_4	4.2	18.8
gdpval	40.4	48.3
GPQA (unspecified)	83.4	92.4
GPQA Diamond	83.4	92.4
HLE	19.8	45.5
HLE (with tools)	32	45.5
HMMT 2025	74.6	99.4
hmmt_feb_2025	79.2	100
hmmt_nov_2025	81.7	99.2
IFBench	57.3	75.4
imo_answer_bench	65.9	86.3
LiveCodeBench	71	89
longbench_v2	61.8	54.5
matharena_visual_math_overall	75.8	86.5
MathVista	79.8	82.8
MCP Atlas	59.5	68
MMLU-Pro	88.2	87
mmmlu	89.1	89.6
MMMU-Pro	68.7	79.5
MultiNRC	35.8	42.2
OmniDocBench 1.5	85.8	85.7
OmniScience Accuracy	32.9	44.3
OmniScience Non-Hallucination	50.9	38.4
scicode	45	52.1
SEAL VISTA	48.8	46.6
Seal-0	53.4	45
simplebench	54.3	45.8
simpleqa_verified	23.6	38.9
SimpleVQA	57.6	55.8
swe_bench_bash	71.4	72.8
SWE-bench Multilingual	68	72
SWE-bench Pro	43.6	55.6
SWE-bench Verified	82	80
SWE-Perf	3	3.6
SWT-bench	69.5	80.7
TauBench V3 - Banking	24.5	11.1
Terminal-Bench 2.0	50	54
Terminal-Bench Hard	35.6	62.2
toolathlon	41	46.3
vectara_answer_rate	95.6	100
vectara_avg_summary_length	127.8	186.3
vectara_factual_consistency	88	89.2
vectara_hallucination_rate ↓	12	10.8
τ²-Bench	98	85.5
τ²-Bench Telecom (AA run)	78.1	98.7
τ³-Bench	19	12

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.