Which is better, Gemma 4 26B A4B or Gemma 4 31B?

Across 71 shared benchmarks tracked on Vector Wire, Gemma 4 26B A4B leads on 12 and Gemma 4 31B leads on 57 — Gemma 4 31B 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 26B A4B vs Gemma 4 31B

GooglevsGoogle71 shared benchmarks12–57 head-to-head

Benchmark	Gemma 4 26B A4B	Gemma 4 31B
AA Agentic Index	32.1	14.4
AA Intelligence	26.1	29.7
AA-LCR	61.7	68.3
AA-Omniscience	-50.8	-47.9
AI2D	88.3	89
aime_2026	88.3	89.2
arena_elo	1438	1451
arena_text_factuality	1445	1449
arena_vision	1238	1252
Artificial Analysis Coding Index	39.3	43.4
BIG-Bench Extra Hard	64.8	74.4
C-Eval	82.5	82.6
CC-OCR	74.5	75.7
charxiv_rq	69	67.9
Claw Eval (pass@3)	28	25
Claw-Eval Avg	58.8	48.5
Codeforces ELO	1718	2150
coding_arena_elo	1361	1364
critpt	0	1.4
DeepPlanning	16.2	24
gdpval	25.7	15.5
GPQA (unspecified)	82.3	84.3
GPQA Diamond	82.3	85.7
HallusionBench	66.1	67.4
HLE	19.3	26.5
HLE (with tools)	17.2	26.5
hmmt_feb_2025	91.7	88.7
hmmt_feb_2026	79	77.2
hmmt_nov_2025	87.5	87.5
IFBench	72.4	75.6
imo_answer_bench	74.3	74.5
itbenchSre	23.6	37.3
LiveCodeBench v6	77.1	80
mathvision	82.4	85.6
MathVista	79.4	79.3
MCP Atlas	50	57.2
mcpmark	14.2	18.1
MedXPertQA MM	58.1	61.3
MMBench EN-DEV-v1.1	89	90.9
mmlu_redux	92.7	93.7
MMLU-Pro	85.2	85.2
mmmlu	86.3	88.4
MMMU	78.4	80.4
MMMU-Pro	73.8	76.9
mrcr_v2_8needle_128k_average	44.1	66.4
nl2repo	11.6	15.5
OmniDocBench 1.5	74.4	80.1
OmniDocBench 1.5 (average edit distance, lower is better)	0.1	0.1
OmniScience Accuracy	19.1	20
OmniScience Non-Hallucination	13.6	18.1
QwenClawBench	38.7	41.7
QwenWebBench	1178	1197
RealWorldQA	72.2	72.3
scicode	40.3	43.4
SimpleVQA	52.2	52.9
SkillsBench Avg5	12.3	23.6
supergpqa	61.4	65.7
SWE-bench Multilingual	43.4	51.7
SWE-bench Pro	13.8	35.7
SWE-bench Verified	57.4	52
TauBench V3 - Banking	12	14.8
Terminal-Bench 2.0	34.2	42.9
Terminal-Bench 2.1	39	43.4
Terminal-Bench Hard	25	36.4
tool_decathlon	12	21.2
VITA-Bench	36.9	43
WideSearch	38.3	35.2
ZEROBench_sub	26.3	26
τ²-Bench	68.2	76.9
τ²-Bench Telecom (AA run)	43.6	65.5
τ³-Bench	59	67.5

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.