Which is better, Claude Opus 4.8 or Opus 5?

Across 54 shared benchmarks tracked on Vector Wire, Claude Opus 4.8 leads on 6 and Opus 5 leads on 48 — Opus 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 4.8 vs Opus 5

AnthropicvsAnthropic54 shared benchmarks6–48 head-to-head

Benchmark	Claude Opus 4.8	Opus 5
AA Agentic Index	49.4	59.2
AA Intelligence	57.3	63.1
AA-LCR	73	78.7
AA-Omniscience	28.8	37.1
arc_agi_1	92.5	97.5
arc_agi_3	1.5	30.2
ARC-AGI-2	72.1	90.4
arena_elo	1482	1487
arena_text_factuality	1464	1481
Artificial Analysis Coding Index	74.3	78
browsecomp	88.5	90.8
coding_arena_elo	1539	1662
critpt	20.9	29.1
DeepSWE 1.1	59	68.8
Finance Agent v2	53.9	58.6
GDP (Surge AI)	84.8	85.5
gdpval	54.2	67.2
GDPval-AA (Elo)	1890	1827
GDPval-AA v2	1593	1861
GPQA Diamond	93.6	93.7
HealthBench	59.3	67.1
HealthBench (raw)	58.8	67.1
HealthBench Professional (raw)	60.3	73.4
healthbench_professional	57.4	59.8
HiL-Bench	35.3	57
HLE	57.9	64.7
HLE (with tools)	57.9	64.7
Legal Agent Benchmark	10	23
livebench_agentic_coding	50.5	65.2
livebench_coding	81.8	81.5
livebench_data_analysis	66	74.5
livebench_instruction_following	72	63.8
livebench_language	79.7	88.7
livebench_math	94.3	95.7
livebench_reasoning	89.2	91.2
MCP Atlas	83.6	85.8
MMMU-Pro	78.9	84.7
officeqa	77.6	78.1
officeqa_pro	66.2	66.9
OmniScience Accuracy	48.8	60.9
OmniScience Non-Hallucination	60.7	40.5
OSWorld 2.0	55.7	70.6
OSWorld-Verified	83.4	4
RiemannBench	34	60
scicode	53.5	55.7
simplebench	64.8	80.6
simpleqa_verified	39.5	56.7
swe_bench_multimodal	38.4	59.4
SWE-bench Multilingual	84.4	89.5
SWE-bench Pro	69.2	79.2
SWE-bench Verified	88.6	96
TauBench V3 - Banking	34.2	44.7
Terminal-Bench 2.1	85	89.1
Toolathlon Verified Pass@3	88	87

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.