Which is better, Claude Opus 4.7 or GPT-5.4?

Across 60 shared benchmarks tracked on Vector Wire, Claude Opus 4.7 leads on 31 and GPT-5.4 leads on 29 — Claude Opus 4.7 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 4.7 vs GPT-5.4

AnthropicvsOpenAI60 shared benchmarks31–29 head-to-head

Benchmark	Claude Opus 4.7	GPT-5.4
AA Agentic Index	64.6	58.2
AA Intelligence	55	53.1
AA-LCR	75.3	77.7
AA-Omniscience	27.3	5.8
aime_2026	95.8	99.2
arc_agi_1	92	93.7
arc_agi_3	0.2	0.2
ARC-AGI-2	75.8	74
arena_elo	1494	1477
arena_text_factuality	1474	1476
arena_vision	1304	1277
Artificial Analysis Coding Index	73.6	71.1
browsecomp	79.8	82.7
charxiv_rq	82.1	82.8
coding_arena_elo	1557	1457
critpt	12	23.4
cybergym	73.8	66.3
finance_agent	71.5	57.2
frontiermath_tier_4	22.9	27.1
gdpval	58.6	50.1
GDPval-AA (Elo)	1753	1674
GPQA Diamond	94.2	93
HiL-Bench	41.7	9.7
HLE	54.7	43.7
HLE (with tools)	54.7	52.1
hmmt_feb_2026	93.9	97.7
IFBench	58.6	73.9
itbenchSre	46.7	34.5
Legal Agent Benchmark	7.1	0.4
livebench	76.9	80.3
livebench_agentic_coding	50.7	53.8
livebench_coding	82.1	77.5
livebench_data_analysis	78.3	79.3
livebench_instruction_following	66.7	70.2
livebench_language	77.9	82.6
livebench_math	92.8	94.2
livebench_reasoning	87.2	88.1
MCP Atlas	79.1	70.6
MMMU-Pro	78.8	81.2
officeqa	86.3	68.1
officeqa_pro	80.6	51.1
OmniScience Accuracy	48.9	50.9
OmniScience Non-Hallucination	57.7	17.4
OSWorld-Verified	82.8	75
scicode	54.5	56.6
simpleqa_verified	50.6	45.3
SWE-bench Multilingual	80.5	71.7
SWE-bench Pro	64.3	59.1
TauBench V3 - Banking	34.6	39.6
Terminal-Bench 2.0	69.4	75.1
Terminal-Bench 2.1	83.1	78.3
Terminal-Bench Hard	54.5	57.6
toolathlon	59.3	54.6
usamo_2026	69.3	95.2
vectara_answer_rate	98	99.9
vectara_avg_summary_length	149.1	81.7
vectara_factual_consistency	88	93
vectara_hallucination_rate ↓	12	7
τ²-Bench Telecom (AA run)	88.6	87.1
τ³-Bench	28.9	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.