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

Across 63 shared benchmarks tracked on Vector Wire, Claude Opus 4.6 leads on 20 and GPT-5.5 leads on 43 — 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 Opus 4.6 vs GPT-5.5

AnthropicvsOpenAI63 shared benchmarks20–43 head-to-head

Benchmark	Claude Opus 4.6	GPT-5.5
AA Agentic Index	67.6	47.4
AA Intelligence	44.9	56.3
AA-LCR	74.3	79
AA-Omniscience	13.7	20.5
aime_2026	96.7	100
apexAgents	33	37.7
arc_agi_1	93	95
arc_agi_3	0.5	0.4
ARC-AGI-2	68.8	85
arena_elo	1497	1482
arena_text_factuality	1487	1482
arena_vision	1300	1283
Artificial Analysis Coding Index	48.1	74.9
browsecomp	86.8	84.4
BrowseComp (single-agent)	83.7	84.4
coding_arena_elo	1545	1457
critpt	12.6	27.1
cybergym	73.8	81.8
finance_agent	76.7	60
Fortress	20.5	16.3
frontiermath_tier_4	22.9	35.4
gdpval	55.9	49.5
GDPval-AA (Elo)	1619	1769
GPQA Diamond	91.3	93.6
HiL-Bench	38.3	39.7
HLE	53.1	52.2
HLE (with tools)	62.7	52.2
hmmt_feb_2026	96.2	98.5
hmmt_nov_2025	96.3	96.5
IFBench	62.5	75.9
Legal Agent Benchmark	4.2	2.1
livebench	76.3	80.7
livebench_agentic_coding	49	54
livebench_coding	78.2	82.2
livebench_data_analysis	69.9	81.6
livebench_instruction_following	63.3	70.7
livebench_language	83.3	87.4
livebench_math	89.3	95.9
livebench_reasoning	88.7	89.7
matharena_visual_math_overall	72.3	94.9
MCP Atlas	76.8	82.8
MMMU-Pro	77.3	83.2
nl2repo	49.8	50.7
officeqa_pro	57.1	60.9
OmniScience Accuracy	47	58
OmniScience Non-Hallucination	37.2	12
OSWorld-Verified	72.7	78.7
scicode	52	56.1
simplebench	67.6	69
simpleqa_verified	46.5	63.1
SWE-bench Pro	57.3	58.6
SWE-bench Verified	80.8	80.6
Terminal-Bench 2.0	65.4	82.7
Terminal-Bench Hard	65.4	60.6
tool_decathlon	47.2	55.6
toolathlon	56.8	55.6
usamo_2026	66.2	98.2
vectara_answer_rate	99.8	100
vectara_avg_summary_length	137.6	129.6
vectara_factual_consistency	87.8	90.7
vectara_hallucination_rate ↓	12.2	9.3
τ²-Bench Telecom (AA run)	99.3	93.9
τ³-Bench	72.4	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.