Which is better, GPT-5.4 or GPT-5.6 Sol?

Across 49 shared benchmarks tracked on Vector Wire, GPT-5.4 leads on 5 and GPT-5.6 Sol leads on 43 — GPT-5.6 Sol leads on more. Strengths differ by axis; the table below shows every head-to-head number with provenance.

SEEDRefresh Models —Deals —Regulatory —Sources —

GPT-5.4 vs GPT-5.6 Sol

OpenAIvsOpenAI49 shared benchmarks5–43 head-to-head

Benchmark	GPT-5.4	GPT-5.6 Sol
AA Agentic Index	58.2	57.8
AA Intelligence	53.1	61
AA-LCR	77.7	77.7
AA-Omniscience	5.8	22
aime_2026	99.2	99.9
arc_agi_1	93.7	97.5
arc_agi_3	0.2	7.8
ARC-AGI-2	74	92.5
arena_elo	1477	1482
arena_text_factuality	1476	1474
Artificial Analysis Coding Index	71.1	78.3
baby_vision_with_python	80.2	88.9
browsecomp	82.7	90.4
charxiv_rq	82.8	87.8
coding_arena_elo	1457	1619
critpt	23.4	32.3
cybergym	66.3	83.6
DeepSWE 1.1	52	73
DNA sequence design for transcription factor binding	12.8	13.7
gdpval	50.1	61.1
GPQA Diamond	93	94.6
Hard-negative protein binding prediction	3.5	7.6
HiL-Bench	9.7	32.3
HLE	43.7	49.5
HLE (with tools)	52.1	58
IFBench	73.9	72.7
itbenchSre	34.5	56.2
livebench_agentic_coding	53.8	56.2
livebench_coding	77.5	83.9
livebench_data_analysis	79.3	79.8
livebench_instruction_following	70.2	71.8
livebench_language	82.6	87.7
livebench_math	94.2	96.2
livebench_reasoning	88.1	91.7
mathvision	92	95.8
MCP Atlas	70.6	83.6
MMMU-Pro	81.2	84.6
officeqa_pro	51.1	63.2
OmniScience Accuracy	50.9	59.4
OmniScience Non-Hallucination	17.4	10.6
OSWorld-Verified	75	83
scicode	56.6	56.9
simpleqa_verified	45.3	71.6
SWE-bench Pro	59.1	64.6
TauBench V3 - Banking	39.6	44.3
Terminal-Bench 2.1	78.3	89.5
Terminal-Bench Hard	57.6	65.9
toolathlon	54.6	58
τ²-Bench Telecom (AA run)	87.1	85.1

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.