Which is better, Claude Sonnet 4.6 or GPT-5.5?

Across 60 shared benchmarks tracked on Vector Wire, Claude Sonnet 4.6 leads on 9 and GPT-5.5 leads on 51 — GPT-5.5 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-5.5

AnthropicvsOpenAI60 shared benchmarks9–51 head-to-head

Benchmark	Claude Sonnet 4.6	GPT-5.5
AA Agentic Index	61.6	47.4
AA Intelligence	48.4	56.3
AA-LCR	74	79
AA-Omniscience	12.2	20.5
apexAgents	28	37.7
arc_agi_1	86.5	95
ARC-AGI-2	60.4	85
arena_elo	1472	1482
arena_text_factuality	1460	1482
arena_vision	1278	1283
Artificial Analysis Coding Index	63	74.9
browsecomp	76.2	84.4
coding_arena_elo	1523	1457
critpt	3.1	27.1
DeepSWE 1.1	30	67
Finance Agent v2	51	51.8
finance_agent	63.3	60
frontiermath_tier_4	8.3	35.4
GDP (Surge AI)	78.6	16.7
gdpval	54.8	49.5
GDPval-AA (Elo)	1676	1769
GDPval-AA v2	1381	1509
GPQA Diamond	89.9	93.6
healthbench_professional	44.2	51.8
HLE	49	52.2
HLE (with tools)	46.8	52.2
IFBench	56.6	75.9
itbenchSre	39.8	45.8
Legal Agent Benchmark	5.4	2.1
Legal Agent Benchmark (Harvey's Held-Out Set)	5.4	2.1
livebench	75.5	80.7
livebench_agentic_coding	42.6	54
livebench_coding	79.3	82.2
livebench_data_analysis	78	81.6
livebench_instruction_following	63.2	70.7
livebench_language	76.1	87.4
livebench_math	87	95.9
livebench_reasoning	84.8	89.7
MCP Atlas	69.5	82.8
MMMU-Pro	75.6	83.2
officeqa_pro	53.4	60.9
OmniScience Accuracy	40.9	58
OmniScience Non-Hallucination	51.6	12
OSWorld-Verified	78.5	78.7
scicode	47	56.1
simpleqa_verified	29	63.1
SWE-bench Pro	58.1	58.6
SWE-bench Verified	79.6	80.6
TauBench V3 - Banking	34.4	39
Terminal-Bench 2.0	59.1	82.7
Terminal-Bench 2.1	71.2	84.3
Terminal-Bench Hard	59.1	60.6
toolathlon	49.4	55.6
usamo_2026	55	98.2
vectara_answer_rate	99.9	100
vectara_avg_summary_length	114.7	129.6
vectara_factual_consistency	89.4	90.7
vectara_hallucination_rate ↓	10.6	9.3
τ²-Bench Telecom (AA run)	97.9	93.9
τ³-Bench	30.5	31.3

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.