Which is better, Claude Opus 4.6 or GPT-5.4?

Across 92 shared benchmarks tracked on Vector Wire, Claude Opus 4.6 leads on 40 and GPT-5.4 leads on 52 — 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 Opus 4.6 vs GPT-5.4

AnthropicvsOpenAI92 shared benchmarks40–52 head-to-head

Benchmark	Claude Opus 4.6	GPT-5.4
AA Agentic Index	67.6	58.2
AA Intelligence	44.9	53.1
AA-LCR	74.3	77.7
AA-Omniscience	13.7	5.8
AgentWorldBench - Android	61.7	60
AgentWorldBench - MCP	69.9	70.1
AgentWorldBench - OS	70.2	68.6
AgentWorldBench - Overall	57.8	58.3
AgentWorldBench - Search	29.3	37.3
AgentWorldBench - SWE	64.5	66.3
AgentWorldBench - Terminal	57.5	53.7
AgentWorldBench - Web	51.4	51.8
aime_2026	96.7	99.2
Apex (Pass@1)	34.5	54.1
Apex Shortlist (Pass@1)	85.9	78.1
apexAgents	33	33.3
arc_agi_1	93	93.7
arc_agi_3	0.5	0.2
ARC-AGI-2	68.8	74
arena_elo	1497	1477
arena_text_factuality	1487	1476
arena_vision	1300	1277
Artificial Analysis Coding Index	48.1	71.1
baby_vision	14.8	49.7
baby_vision_with_python	38.4	80.2
browsecomp	86.8	82.7
BrowseComp (Agent Swarm)	83.7	82.7
browsecomp_with_context_manager	84	82.7
charxiv_rq	69.1	82.8
charxiv_rq_with_python	84.7	90
Chinese SimpleQA (C-SimpleQA)	76.4	76.8
coding_arena_elo	1545	1457
critpt	12.6	23.4
cybergym	73.8	66.3
deepsearchqa_accuracy	80.6	63.7
deepsearchqa_f1	91.3	78.6
finance_agent	76.7	57.2
frontiermath_tier_4	22.9	27.1
gdpval	55.9	50.1
GDPval-AA (Elo)	1619	1674
GPQA Diamond	91.3	93
HiL-Bench	38.3	9.7
HLE	53.1	43.7
HLE (with tools)	62.7	52.1
hmmt_feb_2026	96.2	97.7
hmmt_nov_2025	96.3	95.8
IFBench	62.5	73.9
imo_answer_bench	75.3	91.4
Legal Agent Benchmark	4.2	0.4
livebench	76.3	80.3
livebench_agentic_coding	49	53.8
livebench_coding	78.2	77.5
livebench_data_analysis	69.9	79.3
livebench_instruction_following	63.3	70.2
livebench_language	83.3	82.6
livebench_math	89.3	94.2
livebench_reasoning	88.7	88.1
matharena_visual_math_overall	72.3	92.5
mathvision	71.2	92
mathvision_with_python	84.6	96.1
MCP Atlas	76.8	70.6
MCPAtlas Public (Pass@1)	73.8	67.2
mcpmark	56.7	62.5
MMLU-Pro	89.1	87.5
mmmu_pro_with_python	77.3	82.1
MMMU-Pro	77.3	81.2
MultiNRC	57.1	58.3
nl2repo	49.8	41.3
officeqa	73.5	68.1
officeqa_pro	57.1	51.1
OmniDocBench 1.5	86.6	89.1
OmniScience Accuracy	47	50.9
OmniScience Non-Hallucination	37.2	17.4
OSWorld-Verified	72.7	75
scicode	52	56.6
SEAL VISTA	46.1	50.9
simpleqa_verified	46.5	45.3
SWE-bench Multilingual	77.8	71.7
SWE-bench Pro	57.3	59.1
Terminal-Bench 2.0	65.4	75.1
Terminal-Bench Hard	65.4	57.6
tool_decathlon	47.2	54.6
toolathlon	56.8	54.6
usamo_2026	66.2	95.2
V* (w/ python)	86.4	98.4
vectara_answer_rate	99.8	99.9
vectara_avg_summary_length	137.6	81.7
vectara_factual_consistency	87.8	93
vectara_hallucination_rate ↓	12.2	7
VisualToolBench	27.5	29.2
τ²-Bench Telecom (AA run)	99.3	87.1
τ³-Bench	72.4	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.