Gemma 4 26B A4B vs Qwen3.5 27B
Wins 0 of 8 areas
—
Wins 8 of 8 areas
Coding · Agents · Reasoning · Facts · Images and charts · Math · Long documents · Following instructions
Qwen3.5 27B is the stronger all-rounder.Gemma 4 26B A4B is cheaper.
Scores updated · 65 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 software05Qwen3.5 27B5 of 7 tests · 2 ties
- AgentsCarrying out multi-step tasks on its own04Qwen3.5 27B4 of 4 tests
- Images and chartsUnderstanding pictures, charts and video14Qwen3.5 27B4 of 5 tests
- MathCompetition and research-level math03Qwen3.5 27B3 of 3 tests
- ReasoningHard problems that need careful thinking02Qwen3.5 27B2 of 3 tests · 1 tie
- FactsGetting facts right instead of making them up12Qwen3.5 27B2 of 3 tests
- Long documentsFinding answers in very long texts01Qwen3.5 27B1 of 1 test
- Following instructionsDoing exactly what it is asked01Qwen3.5 27B1 of 1 test
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 80% 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 27B pulls ahead
- Fixes real GitHub issues in Python projectsSWE-bench Verified+55points ahead
- Fixes real GitHub issues in many programming languagesSWE-bench Multilingual+52points ahead
- Long, multi-file coding tasks in real codebasesSWE-bench Pro+37.4points ahead
Where Gemma 4 26B A4B pulls ahead
No clear win on a test scored out of 100.
Every test, side by side
All 65 tests both models report. The winning score is in its model's colour; marks a score checked independently.
CodingQwen3.5 27B
- SWE-bench VerifiedQwen3.5 27B by 5517.472.4+55
- SWE-bench MultilingualQwen3.5 27B by 5217.369.3+52
- SWE-bench ProQwen3.5 27B by 37.413.851.2+37.4
- Terminal-Bench HardQwen3.5 27B by 1913.632.6+19
- LiveCodeBench v6Qwen3.5 27B by 3.677.180.7+3.6
- SciCodetie4039.5tie
- LMArena · WebDevtie13591357tie
AgentsQwen3.5 27B
- BrowseCompQwen3.5 27B by 34.726.361+34.7
- GDPValQwen3.5 27B by 29.63.433+29.6
- MCP AtlasQwen3.5 27B by 18.45068.4+18.4
- AA IT-Bench SREQwen3.5 27B by 11.923.635.5+11.9
ReasoningQwen3.5 27B
- GPQA DiamondQwen3.5 27B by 6.679.285.8+6.6
- Humanity's Last ExamQwen3.5 27B by 4.619.323.9+4.6
- CritPttie00.9tie
FactsQwen3.5 27B
- Vectara HHEM hallucination ratelower is betterGemma 4 26B A4B by 6.95.212.1+6.9
- AA-Omniscience · Non-hallucinationQwen3.5 27B by 4.913.618.5+4.9
- AA-Omniscience · AccuracyQwen3.5 27B by 1.619.120.7+1.6
Images and chartsQwen3.5 27B
- CharXiv (RQ)Qwen3.5 27B by 10.56979.5+10.5
- MathVistaQwen3.5 27B by 8.479.487.8+8.4
- MMMU-ProQwen3.5 27B by 5.869.275+5.8
- MMMUQwen3.5 27B by 3.978.482.3+3.9
- LMArena · VisionGemma 4 26B A4B by 20 rating points12601240+20 rating
MathQwen3.5 27B
- IMOAnswerBenchQwen3.5 27B by 5.674.379.9+5.6
- AIME 2026Qwen3.5 27B by 3.488.391.7+3.4
- HMMT Feb. 2026Qwen3.5 27B by 2.17981.1+2.1
Following instructionsQwen3.5 27B
- IFBenchQwen3.5 27B by 3.272.475.6+3.2
Other results38 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 27B by 50.343.693.9+50.3
- AA Agentic IndexQwen3.5 27B by 43.61154.6+43.6
- vectara_avg_summary_lengthQwen3.5 27B by 27.367.194.4+27.3
- WideSearchQwen3.5 27B by 22.838.361.1+22.8
- MCPMarkQwen3.5 27B by 22.114.236.3+22.1
- Tool-DecathlonQwen3.5 27B by 19.51231.5+19.5
- Claw Eval (pass@3)Qwen3.5 27B by 18.22846.2+18.2
- NL2RepoQwen3.5 27B by 15.711.627.3+15.7
- SkillsBench Avg5Qwen3.5 27B by 14.912.327.2+14.9
- OmniDocBench 1.5Qwen3.5 27B by 14.574.488.9+14.5
- QwenClawBenchQwen3.5 27B by 13.538.752.2+13.5
- RealWorldQAQwen3.5 27B by 11.572.283.7+11.5
- QwenWebBenchGemma 4 26B A4B by 110 rating points11781068+110 rating
- τ³-BenchQwen3.5 27B by 9.45968.4+9.4
- C-EvalQwen3.5 27B by 882.590.5+8
- Terminal-Bench 2.0Qwen3.5 27B by 7.434.241.6+7.4
- vectara_factual_consistencyGemma 4 26B A4B by 6.994.887.9+6.9
- AA-OmniscienceQwen3.5 27B by 6.8-50.8-44+6.8
- CC-OCRQwen3.5 27B by 6.574.581+6.5
- t2-benchGemma 4 26B A4B by 6.585.579+6.5
- DeepPlanningQwen3.5 27B by 6.416.222.6+6.4
- AA IntelligenceQwen3.5 27B by 6.216.722.9+6.2
- Claw-Eval AvgQwen3.5 27B by 5.558.864.3+5.5
- VITA-BenchQwen3.5 27B by 536.941.9+5
- AI2DQwen3.5 27B by 4.688.392.9+4.6
- Artificial Analysis Coding IndexGemma 4 26B A4B by 4.439.334.9+4.4
- SuperGPQAQwen3.5 27B by 4.261.465.6+4.2
- HallusionBenchQwen3.5 27B by 3.966.170+3.9
- SimpleVQAQwen3.5 27B by 3.852.256+3.8
- IFEvalQwen3.5 27B by 3.691.495+3.6
- MATH-VisionQwen3.5 27B by 3.682.486+3.6
- MMBench EN-DEV-v1.1Qwen3.5 27B by 3.68992.6+3.6
- MMLU-ProQwen3.5 27B by 3.582.686.1+3.5
- HMMT Nov. 2025Qwen3.5 27B by 2.387.589.8+2.3
- MMLU-Reduxtie92.793.2tie
- MMMLUtie86.385.9tie
- HMMT Feb. 2025tie91.792tie
- vectara_answer_ratetie99.899.8tie
Questions people ask
Which is better, Gemma 4 26B A4B or Qwen3.5 27B?
Qwen3.5 27B wins all eight areas we test: coding, agents, reasoning, facts, images and charts, math, long documents and following instructions. Gemma 4 26B A4B wins none, but costs 80% less.
Which is better for coding?
Qwen3.5 27B. It wins 5 of the 7 coding tests both models report; Gemma 4 26B A4B wins none, and 2 are ties.
Which is cheaper?
Gemma 4 26B A4B costs $0.13 per million input tokens and $0.40 per million output tokens; Qwen3.5 27B costs $0.30 and $2.40. That makes Gemma 4 26B A4B about 80% cheaper for the same work.
How do you compare the two?
We use the 65 benchmark tests both models have published scores on. The verdict counts the 27 tests in the eight capability areas, and a gap under one point (ten on rating-style scales) counts as a tie. The other 38 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.