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

Across 43 shared benchmarks tracked on Vector Wire, Claude Sonnet 4.5 leads on 4 and Claude Sonnet 4.6 leads on 39 — Claude Sonnet 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 Sonnet 4.5 vs Claude Sonnet 4.6

AnthropicvsAnthropic43 shared benchmarks4–39 head-to-head

Benchmark	Claude Sonnet 4.5	Claude Sonnet 4.6
AA Agentic Index	50.6	61.6
AA Intelligence	37.4	48.4
AA-LCR	68.3	74
AA-Omniscience	-0.1	12.2
arc_agi_1	63.7	86.5
ARC-AGI-2	13.6	60.4
arena_elo	1456	1472
arena_text_factuality	1448	1460
Artificial Analysis Coding Index	52.1	63
browsecomp	24.1	76.2
charxiv_rq	67.2	72.4
coding_arena_elo	1392	1523
critpt	1.1	3.1
frontiermath_tier_4	4.2	8.3
gdpval	40.4	54.8
GPQA (unspecified)	83.4	89.9
GPQA Diamond	83.4	89.9
HLE	19.8	49
HLE (with tools)	32	46.8
IFBench	57.3	56.6
MCP Atlas	59.5	69.5
mmmlu	89.1	89.3
MMMU-Pro	68.7	75.6
OmniScience Accuracy	32.9	40.9
OmniScience Non-Hallucination	50.9	51.6
OSWorld-Verified	61.4	78.5
scicode	45	47
simpleqa_verified	23.6	29
SWE-bench Pro	43.6	58.1
SWE-bench Verified	82	79.6
TauBench V3 - Banking	24.5	34.4
Terminal-Bench 2.0	50	59.1
Terminal-Bench 2.1	55.8	71.2
Terminal-Bench Hard	35.6	59.1
toolathlon	41	49.4
Toolathlon Avg turns	32	16.5
Toolathlon Pass@∞	28.7	38
vectara_answer_rate	95.6	99.9
vectara_avg_summary_length	127.8	114.7
vectara_factual_consistency	88	89.4
vectara_hallucination_rate ↓	12	10.6
τ²-Bench Telecom (AA run)	78.1	97.9
τ³-Bench	19	30.5

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.