Which is better, Claude Opus 4.5 or Claude Opus 4.6?

Across 75 shared benchmarks tracked on Vector Wire, Claude Opus 4.5 leads on 19 and Claude Opus 4.6 leads on 56 — Claude Opus 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 Opus 4.5 vs Claude Opus 4.6

AnthropicvsAnthropic75 shared benchmarks19–56 head-to-head

Benchmark	Claude Opus 4.5	Claude Opus 4.6
AA Agentic Index	59.6	67.6
AA Intelligence	41.9	44.9
AA-LCR	76	74.3
AA-Omniscience	14	13.7
AIME 2025	92.8	99.8
aime_2026	95.1	96.7
AIME25 no tools	91	95.6
arc_agi_1	82	93
ARC-AGI-2	37.6	68.8
arena_elo	1473	1497
arena_text_factuality	1458	1487
Artificial Analysis Coding Index	47.8	48.1
browsecomp	37	86.8
browsecomp_with_context_manager	67.8	84
charxiv_rq	68.5	69.1
coding_arena_elo	1495	1545
critpt	4.6	12.6
cybergym	50.6	73.8
deepsearchqa_f1	76.1	91.3
ERQA	46.8	40.8
finance_agent	55.2	76.7
Fortress	13.6	20.5
frontiermath_tier_4	4.2	22.9
gdpval	47.3	55.9
GPQA Diamond	87	91.3
HLE	30.8	53.1
HLE (with tools)	43.4	62.7
hmmt_feb_2026	85.3	96.2
hmmt_nov_2025	93.3	96.3
IFBench	58	62.5
imo_answer_bench	84	75.3
livebench	76	76.3
livebench_agentic_coding	39.7	49
livebench_coding	79.7	78.2
livebench_data_analysis	74.4	69.9
livebench_instruction_following	62.5	63.3
livebench_language	81.3	83.3
livebench_math	90.4	89.3
livebench_reasoning	80.1	88.7
LiveCodeBench	87	88.8
LiveCodeBench v6	84.8	88.8
mathvision	77.1	71.2
MCP Atlas	69.8	76.8
metr_hcast	73.4	77
metr_re_bench	40	97.4
metr_swaa	99.5	98.5
MMLU-Pro	90	89.1
mmmlu	90.8	91.1
MMMU-Pro	74	77.3
multichallenge	59	56
MultiNRC	48.6	57.1
nl2repo	43.2	49.8
OmniDocBench 1.5	87.7	86.6
OmniScience Accuracy	46.6	47
OmniScience Non-Hallucination	39	37.2
OSWorld-Verified	66.3	72.7
RealWorldQA	77	73.9
scicode	50	52
SEAL VISTA	46.4	46.1
simplebench	62	67.6
simpleqa_verified	41.8	46.5
swe_bench_bash	76.8	75.6
SWE-bench Multilingual	77.5	77.8
SWE-bench Pro	57.1	57.3
SWE-bench Verified	81.5	80.8
Tau2 retail	88.9	91.9
Terminal-Bench 2.0	59.3	65.4
Terminal-Bench Hard	47	65.4
tool_decathlon	43.5	47.2
toolathlon	43.5	56.8
vectara_answer_rate	98.7	99.8
vectara_avg_summary_length	114.5	137.6
vectara_factual_consistency	89.1	87.8
vectara_hallucination_rate ↓	10.9	12.2
τ²-Bench Telecom (AA run)	89.5	99.3

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.