Qwen3 Max vs GLM 5
Qwen3 Max (Qwen, released 2025-09-23) and GLM 5 (Z.ai, released 2026-02-11) are both text-only models built for coding and agent tasks, but they land in different places on Design Arena's benchmarks and on price. Qwen3 Max has the bigger context window; GLM 5 wins on nearly every ranked benchmark category and costs less per token. Both are available on OpenKey with one API key and a flat 3% fee over provider list price.
Spec vs spec
| Spec | Qwen3 Max | GLM 5 |
|---|---|---|
| Context window | 262K | 203K |
| Max output | 33K | — |
| Input modalities | text | text |
| Output modalities | text | text |
| Knowledge cutoff | Jun 30, 2025 | — |
| Released | Sep 23, 2025 | Feb 11, 2026 |
| Reasoning | optional | optional |
Pricing
Per 1M tokens. Provider price plus the flat 3% fee — the sum is what you pay.
qwen/qwen3-max
Input · 1M tokens
$0.780 + 3%$0.803
Output · 1M tokens
$3.90 + 3%$4.02
Cache read · 1M tokens
$0.156 + 3%$0.161
Cache write · 1M tokens
$0.975 + 3%$1.00
FEE — FLAT, EVERY MODEL3%
z-ai/glm-5
Input · 1M tokens
$0.600 + 3%$0.618
Output · 1M tokens
$1.92 + 3%$1.98
Cache read · 1M tokens
$0.120 + 3%$0.124
FEE — FLAT, EVERY MODEL3%
One workload, priced on both
10M input + 2M output tokens at each model's price, flat 3% fee included.
qwen/qwen3-max
$16.07
$15.60 provider + 3%
z-ai/glm-5Cheaper
$10.14
$9.84 provider + 3%
Benchmarks
Design Arena categories where both models have results. Higher Elo and lower rank win.
| Qwen3 Max | GLM 5 | |||
|---|---|---|---|---|
| Category | Elo | Rank | Elo | Rank |
| 3D | 1150 | #62 | 1307 | #15 |
| ASCII art | 1175 | #32 | 1192 | #26 |
| Code | 1159 | #66 | 1295 | #16 |
| Data viz | 1149 | #64 | 1269 | #22 |
| Game dev | 1160 | #62 | 1299 | #15 |
| SVG | 1069 | #60 | 1225 | #19 |
| UI components | 1132 | #67 | 1287 | #21 |
| Websites | 1161 | #66 | 1290 | #18 |
Head-to-head preference voting. How we filter and rank
Design Arena rankings: no contest in the categories that matter for building software
GLM 5 outranks Qwen3 Max in every shared Design Arena category. In codecategories, GLM 5 ranks 16th (elo 1295, 55.6% win rate) against Qwen3 Max's 66th (elo 1159, 44% win rate). In gamedev, GLM 5 ranks 15th (elo 1299) versus Qwen3 Max's 62nd (elo 1160). GLM 5 also has agent-specific benchmarks Qwen3 Max doesn't appear in at all: androidnative rank 6 (elo 1244, 62% win rate), godotgamedev rank 3 (elo 1237), mobileapps rank 10 (elo 1222), and fullstack rank 13 (elo 1190). If your workload is agentic coding — writing, testing, or shipping code autonomously — GLM 5 is the model with track record here.
Pricing: GLM 5 is cheaper on every metric
Provider list price: Qwen3 Max runs $0.78/M input and $3.90/M output. GLM 5 runs $0.60/M input and $1.92/M output. On OpenKey (provider price × 1.03), that's Qwen3 Max at $0.8034/M input and $4.017/M output, versus GLM 5 at $0.618/M input and $1.9776/M output. Input tokens cost 1.3x more on Qwen3 Max (the input_price_ratio). For a 10M-input/2M-output workload: Qwen3 Max costs $15.60, GLM 5 costs $9.84 — a real gap if you're running this at volume.
Context and output limits
Qwen3 Max has a 262,144-token context window; GLM 5 has 202,752 — a 1.29x ratio in Qwen3 Max's favor (the context_ratio). If you're stuffing very large documents or codebases into a single prompt, that extra headroom matters. On the output side, Qwen3 Max caps completions at 32,768 tokens; GLM 5 has no listed max_completion_tokens, which matters for very long generated outputs like full-repo scaffolds. Qwen3 Max also has a known knowledge cutoff of 2025-06-30; GLM 5's cutoff isn't published.
Parameters and reasoning support
GLM 5 supports more inference-time controls: frequency_penalty, min_p, top_k, repetition_penalty, and a `reasoning` parameter with reasoning enabled by default. Qwen3 Max's supported parameter list is shorter and doesn't include reasoning controls, top_k, or repetition_penalty. If you're tuning sampling behavior or want reasoning traces on by default, GLM 5 gives you more knobs out of the box.
Which model for which job
| Use case | Pick | Why |
|---|---|---|
| Agentic coding (Android, full-stack, mobile apps) | GLM 5 | Ranks 6th (androidnative), 13th (fullstack), 10th (mobileapps) on Design Arena agent benchmarks — categories Qwen3 Max isn't ranked in |
| General code generation across categories | GLM 5 | Ranks 16th on codecategories (elo 1295) vs Qwen3 Max's 66th (elo 1159) |
| Very large document or codebase ingestion | Qwen3 Max | 262,144-token context vs GLM 5's 202,752 — 1.29x more room |
| High-volume production workloads on a budget | GLM 5 | $9.84 vs $15.60 for a 10M-input/2M-output workload on OpenKey |
| Long single-shot generation (full repo scaffolds, long docs) | GLM 5 | No published max_completion_tokens cap, versus Qwen3 Max's 32,768-token limit |
| Fine-grained sampling control (top_k, repetition_penalty, reasoning toggle) | GLM 5 | Supports top_k, min_p, repetition_penalty, and a reasoning parameter that Qwen3 Max's parameter list lacks |
Questions
- Which model is better for coding, Qwen3 Max or GLM 5?
- GLM 5, by a wide margin on Design Arena. It ranks 16th in codecategories (elo 1295, 55.6% win rate) versus Qwen3 Max's 66th (elo 1159, 44% win rate), and it has agent-specific benchmarks like androidnative (rank 6) and fullstack (rank 13) where Qwen3 Max has no equivalent ranking.
- How much does each model cost on OpenKey for a typical workload?
- For 10M input tokens and 2M output tokens, Qwen3 Max costs $15.60 and GLM 5 costs $9.84 on OpenKey, which applies a flat 3% fee on top of provider list price ($0.78/$3.90 per M for Qwen3 Max, $0.60/$1.92 per M for GLM 5).
- Which model has a bigger context window?
- Qwen3 Max, with 262,144 tokens versus GLM 5's 202,752 — a 1.29x ratio. If you need to fit very large inputs in one call, Qwen3 Max has more room, though GLM 5 has no published max output token cap while Qwen3 Max caps completions at 32,768 tokens.
- Does either model support reasoning mode?
- GLM 5 lists a `reasoning` supported parameter with reasoning enabled by default (default_enabled: true). Qwen3 Max's supported parameter list doesn't include a reasoning control, though neither model makes reasoning mandatory.