Which is better, Claude Opus 4.7 or GPT-5.5?

Across 62 shared benchmarks tracked on Vector Wire, Claude Opus 4.7 leads on 22 and GPT-5.5 leads on 40 — GPT-5.5 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 4.7 vs GPT-5.5

AnthropicvsOpenAI62 shared benchmarks22–40 head-to-head

Benchmark	Claude Opus 4.7	GPT-5.5
AA Agentic Index	64.6	47.4
AA Intelligence	55	56.3
AA-LCR	75.3	79
AA-Omniscience	27.3	20.5
aime_2026	95.8	100
arc_agi_1	92	95
arc_agi_3	0.2	0.4
ARC-AGI-2	75.8	85
arena_elo	1494	1482
arena_text_factuality	1474	1482
arena_vision	1304	1283
Artificial Analysis Coding Index	73.6	74.9
browsecomp	79.8	84.4
coding_arena_elo	1557	1457
critpt	12	27.1
cybergym	73.8	81.8
Finance Agent v2	51.5	51.8
finance_agent	71.5	60
frontiermath_tier_4	22.9	35.4
gdpval	58.6	49.5
GDPval-AA (Elo)	1753	1769
GPQA Diamond	94.2	93.6
GraphWalks — Parents task, 256K-token context subset	93.6	90.1
GraphWalks BFS 256K	76.9	73.7
HiL-Bench	41.7	39.7
HLE	54.7	52.2
HLE (with tools)	54.7	52.2
hmmt_feb_2026	93.9	98.5
IFBench	58.6	75.9
itbenchSre	46.7	45.8
Legal Agent Benchmark	7.1	2.1
livebench	76.9	80.7
livebench_agentic_coding	50.7	54
livebench_coding	82.1	82.2
livebench_data_analysis	78.3	81.6
livebench_instruction_following	66.7	70.7
livebench_language	77.9	87.4
livebench_math	92.8	95.9
livebench_reasoning	87.2	89.7
MCP Atlas	79.1	82.8
MMMU-Pro	78.8	83.2
officeqa_pro	80.6	60.9
OmniScience Accuracy	48.9	58
OmniScience Non-Hallucination	57.7	12
OSWorld-Verified	82.8	78.7
scicode	54.5	56.1
simplebench	61.7	69
simpleqa_verified	50.6	63.1
SWE-bench Pro	64.3	58.6
SWE-bench Verified	87.6	80.6
TauBench V3 - Banking	34.6	39
Terminal-Bench 2.0	69.4	82.7
Terminal-Bench 2.1	83.1	84.3
Terminal-Bench Hard	54.5	60.6
toolathlon	59.3	55.6
usamo_2026	69.3	98.2
vectara_answer_rate	98	100
vectara_avg_summary_length	149.1	129.6
vectara_factual_consistency	88	90.7
vectara_hallucination_rate ↓	12	9.3
τ²-Bench Telecom (AA run)	88.6	93.9
τ³-Bench	28.9	31.3

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.