Which is better, Claude Sonnet 4.6 or GPT-5.2?

Across 52 shared benchmarks tracked on Vector Wire, Claude Sonnet 4.6 leads on 31 and GPT-5.2 leads on 21 — Claude Sonnet 4.6 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 4.6 vs GPT-5.2

AnthropicvsOpenAI52 shared benchmarks31–21 head-to-head

Benchmark	Claude Sonnet 4.6	GPT-5.2
AA Agentic Index	61.6	60.2
AA Intelligence	48.4	43.3
AA-LCR	74	79.3
AA-Omniscience	12.2	0.3
apexAgents	28	23
arc_agi_1	86.5	86.2
ARC-AGI-2	60.4	52.9
arena_elo	1472	1437
arena_text_factuality	1460	1442
arena_vision	1278	1245
Artificial Analysis Coding Index	63	48.7
browsecomp	76.2	65.8
charxiv_rq	72.4	82.1
coding_arena_elo	1523	1418
critpt	3.1	11.6
frontiermath_tier_4	8.3	18.8
gdpval	54.8	48.3
GDPval-AA (Elo)	1676	1462
GPQA (unspecified)	89.9	92.4
GPQA Diamond	89.9	92.4
HLE	49	45.5
HLE (with tools)	46.8	45.5
IFBench	56.6	75.4
livebench	75.5	74.8
livebench_agentic_coding	42.6	50.3
livebench_coding	79.3	76.1
livebench_data_analysis	78	78.2
livebench_instruction_following	63.2	61.8
livebench_language	76.1	79.8
livebench_math	87	93.2
livebench_reasoning	84.8	83.2
MCP Atlas	69.5	68
mmmlu	89.3	89.6
MMMU-Pro	75.6	79.5
mrcr_v2_8needle_128k_average	84.9	83.8
OmniScience Accuracy	40.9	44.3
OmniScience Non-Hallucination	51.6	38.4
scicode	47	52.1
simpleqa_verified	29	38.9
SWE-bench Pro	58.1	55.6
SWE-bench Verified	79.6	80
Tau2 retail	91.7	82
TauBench V3 - Banking	34.4	11.1
Terminal-Bench 2.0	59.1	54
Terminal-Bench Hard	59.1	62.2
toolathlon	49.4	46.3
vectara_answer_rate	99.9	100
vectara_avg_summary_length	114.7	186.3
vectara_factual_consistency	89.4	89.2
vectara_hallucination_rate ↓	10.6	10.8
τ²-Bench Telecom (AA run)	97.9	98.7
τ³-Bench	30.5	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.