Which is better, Claude Sonnet 4.5 or GPT-5.4?

Across 47 shared benchmarks tracked on Vector Wire, Claude Sonnet 4.5 leads on 3 and GPT-5.4 leads on 44 — GPT-5.4 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-5.4

AnthropicvsOpenAI47 shared benchmarks3–44 head-to-head

Benchmark	Claude Sonnet 4.5	GPT-5.4
AA Agentic Index	50.6	58.2
AA Intelligence	37.4	53.1
AA-LCR	68.3	77.7
AA-Omniscience	-0.1	5.8
arc_agi_1	63.7	93.7
ARC-AGI-2	13.6	74
arena_elo	1456	1477
arena_text_factuality	1448	1476
Artificial Analysis Coding Index	52.1	71.1
browsecomp	24.1	82.7
charxiv_rq	67.2	82.8
coding_arena_elo	1392	1457
critpt	1.1	23.4
frontiermath_tier_4	4.2	27.1
gdpval	40.4	50.1
GPQA Diamond	83.4	93
HLE	19.8	43.7
HLE (with tools)	32	52.1
hmmt_nov_2025	81.7	95.8
IFBench	57.3	73.9
imo_answer_bench	65.9	91.4
matharena_visual_math_overall	75.8	92.5
MCP Atlas	59.5	70.6
MMLU-Pro	88.2	87.5
MMMU-Pro	68.7	81.2
MultiNRC	35.8	58.3
OmniDocBench 1.5	85.8	89.1
OmniScience Accuracy	32.9	50.9
OmniScience Non-Hallucination	50.9	17.4
OSWorld-Verified	61.4	75
scicode	45	56.6
SEAL VISTA	48.8	50.9
simpleqa_verified	23.6	45.3
SWE-bench Multilingual	68	71.7
SWE-bench Pro	43.6	59.1
TauBench V3 - Banking	24.5	39.6
Terminal-Bench 2.0	50	75.1
Terminal-Bench 2.1	55.8	78.3
Terminal-Bench Hard	35.6	57.6
toolathlon	41	54.6
vectara_answer_rate	95.6	99.9
vectara_avg_summary_length	127.8	81.7
vectara_factual_consistency	88	93
vectara_hallucination_rate ↓	12	7
VisualToolBench	6.2	29.2
τ²-Bench Telecom (AA run)	78.1	87.1
τ³-Bench	19	72.9

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.