Inference
EN

GLM-5 vs Qwen3.5-27B

Change models
GLM-5Z.AI
Qwen3.5-27BAlibaba
IndexAll categories, out of 100.
51.6 ±3.6 best
51.3 ±4.3
GapIndex points below the top.
Top
−0.3 · no clear gap
CoverageIndex weight that has results.
90%
90%
SpeedOutput tokens per second.
45/s
69/s best
EloLMArena rating and rank.
1446 best
1409
ReleasedMonth of release.
2026-02
2026-03
Input / 1MUS dollars per 1M input tokens.
$1.00
$0.26 best
Output / 1MUS dollars per 1M output tokens.
$3.20
$2.60 best
1K chats2K tokens in, 500 out, each.
$3.60
$1.82 best
1K documents20K tokens in, 1K out, each.
$23.20
$7.80 best
1K agent tasks200K tokens in, 20K out, each.
$264.00
$104.00 best
ContextMaximum tokens in one request.
205K
262K best
AgenticMulti-step tasks with tools.
51.6 best
50.3
CodingCode writing and repair.
53.5 best
52.7
ReasoningLogic problems and puzzles.
48.3
50.5 best
MultimodalTasks with images and text.
—
53.9
KnowledgeFacts and expert knowledge.
51.8 best
49.0
MultilingualTasks in many languages.
52.5 best
51.6
InstructionTasks with strict prompt rules.
53.1
57.1 best
MathMath problems.
52.5
—

Benchmarks

18 shared
Only GLM-5: 45Only Qwen3.5-27B: 11
GLM-5
Qwen3.5-27B
Artificial Analysis Humanity's Last ExamKnowledge
29.3% best84 / 222
23.9%101 / 222
Artificial Analysis Long Context ReasoningReasoning
75.7%92 / 217
77.7% best74 / 217
Critical Physics TasksReasoning
2.0% best105 / 214
0.9%124 / 214
Artificial Analysis GPQA DiamondKnowledge
82.0%92 / 199
85.8% best71 / 199
Artificial Analysis Omniscience AccuracyKnowledge
26.3% best96 / 164
20.7%127 / 164
τ²-Bench Tool-Agent-User EvaluationAgentic
98.2% best7 / 149
93.9%29 / 149
Artificial Analysis IFBenchInstruction
72.3%37 / 146
75.6% best21 / 146
Graduate-Level Google-Proof Q&AKnowledge
86.0% best48 / 141
85.5%52 / 141
Software Engineering Benchmark VerifiedCoding
77.8% best30 / 114
72.4%60 / 114
Massive Multitask Language Understanding ProfessionalKnowledge
85.7%16 / 75
86.1% best15 / 75
Instruction-Following EvalInstruction
92.6%16 / 70
95.0% best4 / 70
BrowseCompAgentic
62.0% best50 / 63
61.0%51 / 63
Gert Labs Composite Game BenchmarkAgentic
51.0% best20 / 52
39.4%39 / 52
Terminal-Bench 2.0 · BenchLMAgentic
56.2% best24 / 41
41.6%36 / 41
LongBench v2Reasoning
60.8% best13 / 25
60.6%14 / 25
SuperGPQA: Scaling LLM Evaluation Across 285 Graduate DisciplinesKnowledge
66.8% best13 / 22
65.6%15 / 22
MMLU-ProXMultilingual
83.1% best6 / 17
82.2%9 / 17
SWE-RebenchCoding
62.8%
58.9%