Gemma 4 26B A4B vs Qwen3.5 397B A17B
Wins 0 of 8 areas
—
Wins 7 of 8 areas
Coding · Agents · Reasoning · Images and charts · Math · Long documents · Following instructions
Qwen3.5 397B A17B is the stronger all-rounder.Gemma 4 26B A4B is cheaper.
Scores updated · 54 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 software08Qwen3.5 397B A17B8 of 8 tests
- AgentsCarrying out multi-step tasks on its own04Qwen3.5 397B A17B4 of 4 tests
- ReasoningHard problems that need careful thinking03Qwen3.5 397B A17B3 of 3 tests
- MathCompetition and research-level math03Qwen3.5 397B A17B3 of 3 tests
- Images and chartsUnderstanding pictures, charts and video02Qwen3.5 397B A17B2 of 3 tests · 1 tie
- 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
- FactsGetting facts right instead of making them up11Even1 each
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 26B A4B costs 87% less for the same work.
The biggest differences
The tests each model wins by the widest margin, up to three each. Scores are out of 100.
Where Qwen3.5 397B A17B pulls ahead
- Fixes real GitHub issues in Python projectsSWE-bench Verified+59points ahead
- Fixes real GitHub issues in many programming languagesSWE-bench Multilingual+52points ahead
- Finds hard-to-locate facts by browsing the webBrowseComp+42.7points ahead
Where Gemma 4 26B A4B pulls ahead
- Avoids making up answers it doesn't knowAA-Omniscience · Non-hallucination+2.5points ahead
Every test, side by side
All 54 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 5917.476.4+59
- SWE-bench MultilingualQwen3.5 397B A17B by 5217.369.3+52
- SWE-bench ProQwen3.5 397B A17B by 37.113.850.9+37.1
- Terminal-Bench HardQwen3.5 397B A17B by 27.313.640.9+27.3
- Terminal-Bench 2.1Qwen3.5 397B A17B by 12.33951.3+12.3
- LiveCodeBench v6Qwen3.5 397B A17B by 6.577.183.6+6.5
- SciCodeQwen3.5 397B A17B by 4.84044.8+4.8
- LMArena · WebDevQwen3.5 397B A17B by 41 rating points13591400+41 rating
AgentsQwen3.5 397B A17B
- BrowseCompQwen3.5 397B A17B by 42.726.369+42.7
- GDPValQwen3.5 397B A17B by 11.43.414.8+11.4
- AA IT-Bench SREQwen3.5 397B A17B by 10.523.634.1+10.5
- τ-Bench V3 · BankingQwen3.5 397B A17B by 1.41213.4+1.4
ReasoningQwen3.5 397B A17B
- GPQA DiamondQwen3.5 397B A17B by 10.179.289.3+10.1
- Humanity's Last ExamQwen3.5 397B A17B by 9.719.329+9.7
- CritPtQwen3.5 397B A17B by 1.701.7+1.7
FactsEven
- AA-Omniscience · AccuracyQwen3.5 397B A17B by 11.719.130.8+11.7
- AA-Omniscience · Non-hallucinationGemma 4 26B A4B by 2.513.611.1+2.5
Images and chartsQwen3.5 397B A17B
- MMMU-ProQwen3.5 397B A17B by 8.169.277.3+8.1
- MMMUQwen3.5 397B A17B by 6.678.485+6.6
- LMArena · Visiontie12601264tie
MathQwen3.5 397B A17B
- HMMT Feb. 2026Qwen3.5 397B A17B by 8.97987.9+8.9
- IMOAnswerBenchQwen3.5 397B A17B by 6.674.380.9+6.6
- AIME 2026Qwen3.5 397B A17B by 388.391.3+3
Long documentsQwen3.5 397B A17B
- AA-LCRQwen3.5 397B A17B by 11.665.777.3+11.6
Following instructionsQwen3.5 397B A17B
- IFBenchQwen3.5 397B A17B by 6.472.478.8+6.4
Other results29 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.
- τ²-Bench Telecom (AA run)Qwen3.5 397B A17B by 5243.695.6+52
- WideSearchQwen3.5 397B A17B by 35.738.374+35.7
- MCPMarkQwen3.5 397B A17B by 31.914.246.1+31.9
- HLE (with tools)Qwen3.5 397B A17B by 31.117.248.3+31.1
- NL2RepoQwen3.5 397B A17B by 20.611.632.2+20.6
- AA-OmniscienceQwen3.5 397B A17B by 20.1-50.8-30.8+20.1
- Terminal-Bench 2.0Qwen3.5 397B A17B by 18.334.252.5+18.3
- DeepPlanningQwen3.5 397B A17B by 18.116.234.3+18.1
- SkillsBench Avg5Qwen3.5 397B A17B by 17.712.330+17.7
- GDPval-AA v2Qwen3.5 397B A17B by 155 rating points807962+155 rating
- QwenClawBenchQwen3.5 397B A17B by 13.138.751.8+13.1
- VITA-BenchQwen3.5 397B A17B by 12.836.949.7+12.8
- Claw-Eval AvgQwen3.5 397B A17B by 11.958.870.7+11.9
- PinchBenchQwen3.5 397B A17B by 11.974.786.6+11.9
- C-EvalQwen3.5 397B A17B by 10.582.593+10.5
- SuperGPQAQwen3.5 397B A17B by 961.470.4+9
- Artificial Analysis Coding IndexQwen3.5 397B A17B by 8.939.348.2+8.9
- HMMT Nov. 2025Qwen3.5 397B A17B by 5.287.592.7+5.2
- MMLU-ProQwen3.5 397B A17B by 5.282.687.8+5.2
- GPQA (unspecified)Qwen3.5 397B A17B by 4.882.387.1+4.8
- HMMT Feb. 2025Qwen3.5 397B A17B by 3.191.794.8+3.1
- MMLU-ReduxQwen3.5 397B A17B by 2.292.794.9+2.2
- MMMLUQwen3.5 397B A17B by 2.286.388.5+2.2
- AA IntelligenceQwen3.5 397B A17B by 1.716.718.4+1.7
- IFEvalQwen3.5 397B A17B by 1.291.492.6+1.2
- t2-benchQwen3.5 397B A17B by 1.285.586.7+1.2
- QwenWebBenchtie11781186tie
- τ³-Bench Bankingtie1413.4tie
- AA Agentic Indextie1110.6tie
Questions people ask
Which is better, Gemma 4 26B A4B or Qwen3.5 397B A17B?
Qwen3.5 397B A17B wins seven of the eight areas we test: coding, agents, reasoning, images and charts, math, long documents and following instructions. Gemma 4 26B A4B wins none, but costs 87% less. They are level on facts.
Which is better for coding?
Qwen3.5 397B A17B. It wins 8 of the 8 coding tests both models report; Gemma 4 26B A4B wins none.
Which is cheaper?
Gemma 4 26B A4B costs $0.13 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 26B A4B about 87% cheaper for the same work.
How do you compare the two?
We use the 54 benchmark tests both models have published scores on. The verdict counts the 25 tests in the eight capability areas, and a gap under one point (ten on rating-style scales) counts as a tie. The other 29 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.