Which is better, Gemma 4 31B or GPT-5.4?

Across 45 shared benchmarks tracked on Vector Wire, Gemma 4 31B leads on 4 and GPT-5.4 leads on 41 — GPT-5.4 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-5.4

GooglevsOpenAI45 shared benchmarks4–41 head-to-head

Benchmark	Gemma 4 31B	GPT-5.4
AA Agentic Index	14.4	58.2
AA Intelligence	29.7	53.1
AA-LCR	68.3	77.7
AA-Omniscience	-47.9	5.8
aime_2026	89.2	99.2
arena_elo	1451	1477
arena_text_factuality	1449	1476
arena_vision	1252	1277
Artificial Analysis Coding Index	43.4	71.1
charxiv_rq	67.9	82.8
coding_arena_elo	1364	1457
critpt	1.4	23.4
gdpval	15.5	50.1
GPQA Diamond	85.7	93
HLE	26.5	43.7
HLE (with tools)	26.5	52.1
hmmt_feb_2026	77.2	97.7
hmmt_nov_2025	87.5	95.8
IFBench	75.6	73.9
imo_answer_bench	74.5	91.4
itbenchSre	37.3	34.5
mathvision	85.6	92
MCP Atlas	57.2	70.6
mcpmark	18.1	62.5
MMLU-Pro	85.2	87.5
MMMU-Pro	76.9	81.2
nl2repo	15.5	41.3
OmniDocBench 1.5	80.1	89.1
OmniScience Accuracy	20	50.9
OmniScience Non-Hallucination	18.1	17.4
scicode	43.4	56.6
simpleqa_verified	9.6	45.3
SWE-bench Multilingual	51.7	71.7
SWE-bench Pro	35.7	59.1
TauBench V3 - Banking	14.8	39.6
Terminal-Bench 2.0	42.9	75.1
Terminal-Bench 2.1	43.4	78.3
Terminal-Bench Hard	36.4	57.6
tool_decathlon	21.2	54.6
vectara_answer_rate	100	99.9
vectara_avg_summary_length	75.8	81.7
vectara_factual_consistency	92.6	93
vectara_hallucination_rate ↓	7.4	7
τ²-Bench Telecom (AA run)	65.5	87.1
τ³-Bench	67.5	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.