Which is better, Gemma 4 31B or GPT-5.2?

Across 50 shared benchmarks tracked on Vector Wire, Gemma 4 31B leads on 10 and GPT-5.2 leads on 39 — GPT-5.2 leads on more. Strengths differ by axis; the table below shows every head-to-head number with provenance.

SEEDRefresh Models —Deals —Regulatory —Sources —

Gemma 4 31B vs GPT-5.2

GooglevsOpenAI50 shared benchmarks10–39 head-to-head

Benchmark	Gemma 4 31B	GPT-5.2
AA Agentic Index	14.4	60.2
AA Intelligence	29.7	43.3
AA-LCR	68.3	79.3
AA-Omniscience	-47.9	0.3
aime_2026	89.2	98.3
arena_elo	1451	1437
arena_text_factuality	1449	1442
arena_vision	1252	1245
Artificial Analysis Coding Index	43.4	48.7
charxiv_rq	67.9	82.1
coding_arena_elo	1364	1418
critpt	1.4	11.6
gdpval	15.5	48.3
GPQA (unspecified)	84.3	92.4
GPQA Diamond	85.7	92.4
HLE	26.5	45.5
HLE (with tools)	26.5	45.5
hmmt_feb_2025	88.7	100
hmmt_feb_2026	77.2	97
hmmt_nov_2025	87.5	99.2
IFBench	75.6	75.4
imo_answer_bench	74.5	86.3
mathvision	85.6	83
MathVista	79.3	82.8
MCP Atlas	57.2	68
MMLU-Pro	85.2	87
mmmlu	88.4	89.6
MMMU-Pro	76.9	79.5
mrcr_v2_8needle_128k_average	66.4	83.8
OCRBench	86.1	80.7
OmniDocBench 1.5	80.1	85.7
OmniScience Accuracy	20	44.3
OmniScience Non-Hallucination	18.1	38.4
scicode	43.4	52.1
simpleqa_verified	9.6	38.9
SimpleVQA	52.9	55.8
SWE-bench Multilingual	51.7	72
SWE-bench Pro	35.7	55.6
SWE-bench Verified	52	80
TauBench V3 - Banking	14.8	11.1
Terminal-Bench 2.0	42.9	54
Terminal-Bench Hard	36.4	62.2
tool_decathlon	21.2	46.3
vectara_answer_rate	100	100
vectara_avg_summary_length	75.8	186.3
vectara_factual_consistency	92.6	89.2
vectara_hallucination_rate ↓	7.4	10.8
τ²-Bench	76.9	85.5
τ²-Bench Telecom (AA run)	65.5	98.7
τ³-Bench	67.5	12

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.