Gemma 4 31B vs Qwen3.5 397B A17B
Wins 1 of 8 areas
Agents
Wins 7 of 8 areas
Coding · Reasoning · Facts · Images and charts · Math · Long documents · Following instructions
Qwen3.5 397B A17B is the stronger all-rounder.Gemma 4 31B is cheaper and better at agents.
Scores updated · 51 tests both models report · How we compare
Where each one wins
Tests won in each of the eight areas we test. Each piece is one test, so a longer bar means more evidence; grey means the two scored within a point of each other.
- CodingWriting and fixing software07Qwen3.5 397B A17B7 of 8 tests · 1 tie
- MathCompetition and research-level math03Qwen3.5 397B A17B3 of 3 tests
- ReasoningHard problems that need careful thinking02Qwen3.5 397B A17B2 of 3 tests · 1 tie
- FactsGetting facts right instead of making them up12Qwen3.5 397B A17B2 of 3 tests
- Images and chartsUnderstanding pictures, charts and video12Qwen3.5 397B A17B2 of 3 tests
- Long documentsFinding answers in very long texts01Qwen3.5 397B A17B1 of 1 test
- Following instructionsDoing exactly what it is asked01Qwen3.5 397B A17B1 of 1 test
- AgentsCarrying out multi-step tasks on its own21Gemma 4 31B2 of 4 tests · 1 tie
Long documents and following instructions rest on a single test each.
What it costs
Prices per million tokens, roughly 750,000 words. The bars show the cost of a million tokens read plus a million written.
Gemma 4 31B costs 87% less for the same work.
The biggest differences
The three tests each model wins by the widest margin. Scores are out of 100.
Where Qwen3.5 397B A17B pulls ahead
- Fixes real GitHub issues in Python projectsSWE-bench Verified+24.4points ahead
- Fixes real GitHub issues in many programming languagesSWE-bench Multilingual+17.6points ahead
- Short factual questions, answered correctlySimpleQA Verified+15.6points ahead
Where Gemma 4 31B pulls ahead
- Avoids making up answers it doesn't knowAA-Omniscience · Non-hallucination+3.9points ahead
- Finds the root cause of IT system incidentsAA IT-Bench SRE+3.2points ahead
- Customer-service tasks in a simulated bankτ-Bench V3 · Banking+1.4points ahead
Every test, side by side
All 51 tests both models report. The winning score is in its model's colour; marks a score checked independently.
CodingQwen3.5 397B A17B
- SWE-bench VerifiedQwen3.5 397B A17B by 24.45276.4+24.4
- SWE-bench MultilingualQwen3.5 397B A17B by 17.651.769.3+17.6
- SWE-bench ProQwen3.5 397B A17B by 15.235.750.9+15.2
- Terminal-Bench 2.1Qwen3.5 397B A17B by 7.943.451.3+7.9
- Terminal-Bench HardQwen3.5 397B A17B by 4.536.440.9+4.5
- LiveCodeBench v6Qwen3.5 397B A17B by 3.68083.6+3.6
- LMArena · WebDevQwen3.5 397B A17B by 35 rating points13651400+35 rating
- SciCodetie45.544.8tie
AgentsGemma 4 31B
- GDPValQwen3.5 397B A17B by 8.76.114.8+8.7
- AA IT-Bench SREGemma 4 31B by 3.237.334.1+3.2
- τ-Bench V3 · BankingGemma 4 31B by 1.414.813.4+1.4
- Terminal-Bench 4.0tie00tie
ReasoningQwen3.5 397B A17B
- Humanity's Last ExamQwen3.5 397B A17B by 5.423.629+5.4
- GPQA DiamondQwen3.5 397B A17B by 3.685.789.3+3.6
- CritPttie1.41.7tie
FactsQwen3.5 397B A17B
- SimpleQA VerifiedQwen3.5 397B A17B by 15.610.426+15.6
- AA-Omniscience · AccuracyQwen3.5 397B A17B by 10.82030.8+10.8
- AA-Omniscience · Non-hallucinationGemma 4 31B by 3.91511.1+3.9
Images and chartsQwen3.5 397B A17B
- MMMUQwen3.5 397B A17B by 4.680.485+4.6
- MMMU-ProQwen3.5 397B A17B by 3.973.477.3+3.9
- LMArena · VisionGemma 4 31B by 13 rating points12771264+13 rating
MathQwen3.5 397B A17B
- HMMT Feb. 2026Qwen3.5 397B A17B by 10.777.287.9+10.7
- IMOAnswerBenchQwen3.5 397B A17B by 6.474.580.9+6.4
- AIME 2026Qwen3.5 397B A17B by 2.189.291.3+2.1
Long documentsQwen3.5 397B A17B
- AA-LCRQwen3.5 397B A17B by 7.669.777.3+7.6
Following instructionsQwen3.5 397B A17B
- IFBenchQwen3.5 397B A17B by 3.275.678.8+3.2
Other results25 tests, not counted
Tests outside the eight areas. They are not counted above: several are summary scores built from other tests, or the same test under another name.
- WideSearchQwen3.5 397B A17B by 38.835.274+38.8
- τ²-Bench Telecom (AA run)Qwen3.5 397B A17B by 35.759.995.6+35.7
- MCPMarkQwen3.5 397B A17B by 2818.146.1+28
- Claw-Eval AvgQwen3.5 397B A17B by 22.248.570.7+22.2
- HLE (with tools)Qwen3.5 397B A17B by 21.826.548.3+21.8
- AA-OmniscienceQwen3.5 397B A17B by 17.2-47.9-30.8+17.2
- NL2RepoQwen3.5 397B A17B by 16.715.532.2+16.7
- C-EvalQwen3.5 397B A17B by 10.482.693+10.4
- DeepPlanningQwen3.5 397B A17B by 10.32434.3+10.3
- QwenClawBenchQwen3.5 397B A17B by 10.141.751.8+10.1
- Terminal-Bench 2.0Qwen3.5 397B A17B by 9.642.952.5+9.6
- VITA-BenchQwen3.5 397B A17B by 6.74349.7+6.7
- SkillsBench Avg5Qwen3.5 397B A17B by 6.423.630+6.4
- HMMT Feb. 2025Qwen3.5 397B A17B by 6.188.794.8+6.1
- HMMT Nov. 2025Qwen3.5 397B A17B by 5.287.592.7+5.2
- Artificial Analysis Coding IndexQwen3.5 397B A17B by 4.843.448.2+4.8
- SuperGPQAQwen3.5 397B A17B by 4.765.770.4+4.7
- AA Agentic IndexQwen3.5 397B A17B by 3.96.710.6+3.9
- AA IntelligenceQwen3.5 397B A17B by 3.714.718.4+3.7
- GPQA (unspecified)Qwen3.5 397B A17B by 2.884.387.1+2.8
- MMLU-ProQwen3.5 397B A17B by 2.685.287.8+2.6
- MMLU-ReduxQwen3.5 397B A17B by 1.293.794.9+1.2
- QwenWebBenchGemma 4 31B by 11 rating points11971186+11 rating
- t2-benchtie86.486.7tie
- MMMLUtie88.488.5tie
Questions people ask
Which is better, Gemma 4 31B or Qwen3.5 397B A17B?
Qwen3.5 397B A17B wins seven of the eight areas we test: coding, reasoning, facts, images and charts, math, long documents and following instructions. Gemma 4 31B wins agents, and costs 87% less.
Which is better for coding?
Qwen3.5 397B A17B. It wins 7 of the 8 coding tests both models report; Gemma 4 31B wins none, and 1 is a tie.
Which is cheaper?
Gemma 4 31B costs $0.17 per million input tokens and $0.40 per million output tokens; Qwen3.5 397B A17B costs $0.60 and $3.60. That makes Gemma 4 31B about 87% cheaper for the same work.
How do you compare the two?
We use the 51 benchmark tests both models have published scores on. The verdict counts the 26 tests in the eight capability areas, and a gap under one point (ten on rating-style scales) counts as a tie. The other 25 are listed but not counted, because several are summary scores or repeat a test. Each score is the one shown on the model's own page.