Zhipu GLM-5 Review 2026: Features, Pricing, Pros & Cons
Verdict first: GLM-5 is the strongest open-weights model of 2026 — a 744B-parameter MIT-licensed powerhouse whose coding rivals Claude Opus 4.5 at a tenth of the price. But it's slower than rivals and its price hikes are controversial. Production teams, take note.
Zhipu GLM-5 — the open-weights AI flagship from Z.ai
What Is Zhipu GLM?
Zhipu AI (rebranded as Z.ai) is one of China's earliest and best-known general-purpose AI companies, and its GLM series is the flagship of China's open-source models. In January 2026, Zhipu listed on the Hong Kong Stock Exchange. One month later, on February 11, 2026, it released the next-generation flagship model GLM-5, open-sourced the following day.
A fun detail: an anonymous model called "Pony Alpha" that had been trending on OpenRouter was confirmed to be GLM-5.
GLM-5 core specs:
- 744B parameters MoE (~40B active), double the previous GLM-4.7
- 200K context window (~200K tokens), up to 128K output
- DeepSeek Sparse Attention (DSA) + Multi-Token Prediction (MTP)
- Trained on 28.5 trillion tokens
- MIT license — free commercial use
- Trained entirely on 100,000 Huawei Ascend 910B chips (zero NVIDIA hardware)
Positioning shift: from "Vibe Coding" to "Agentic Engineering" — focused on complex systems engineering and long-horizon agent tasks.
Key Features
1. Top-Tier Coding (#1 Open Source)
Officially aligned with Claude Opus 4.5 and above Gemini 3 Pro:
- SWE-bench Verified: 77.8% — #1 among open-source models
- Terminal-Bench 2.0: 56.2 — open-source SOTA
- Chatbot Arena Code: 1449 Elo — #1 open source, #6 overall
- Strong at multi-file reasoning, stack-trace debugging, and agentic terminal workflows
2. Long-Horizon Agent Ability
Top open-source performance on BrowseComp (web retrieval), MCP-Atlas (tool calling), and τ-Bench (complex multi-tool planning). The "Agentic Engineering" positioning pays off.
3. Office Document Handling
Can agentically process .docx, .pdf, and .xlsx files and deliver results — including natural-language spreadsheet manipulation via an Excel plugin. It does the work, not just instructions.
4. Z Code Tool
A developer tool that drives multi-agent concurrent coding, debugging, and preview via natural language — even remote control of a desktop AI from your phone.
5. Full Domestic-China Compute Stack
Deep adaptation across seven Chinese chip platforms (Huawei Ascend, Moore Threads, Cambricon, Kunlunxin, Muxi, Sunway, Hygon) — outside US export-control reach.
6. "Preserved Thinking" Mode
Maintains reasoning state across 50+ conversation turns — a differentiated feature that keeps long conversations coherent.
7. Full Model Family
| Model | Released | Focus | API (per 1M tokens) |
|---|---|---|---|
| GLM-5 | Feb 2026 | Flagship base (text-only) | $1.00 in / $3.20 out |
| GLM-5-Turbo | Mar 2026 | Agent workflows | $1.20 in / $4.00 out |
| GLM-5V-Turbo | Apr 2026 | Vision + coding (multimodal) | $1.20 in / $4.00 out |
| GLM-5.1 | Apr 2026 | Long-horizon agentic coding | $1.40 in / $4.40 out |
Pricing: API & Plans (2026)
| Model | Input / Million tokens | Output / Million tokens | Price change |
|---|---|---|---|
| GLM-5 | 4 yuan | 18 yuan | Input +100%, output +125% |
| GLM-5-Code | 6 yuan | 28 yuan | Input +200%, output +250% |
GLM-5 API pricing (per million tokens)
Subscriptions & cards:
- GLM Coding Plan: overall +30% starting, first-purchase discount removed; existing subscribers unchanged. GLM Coding Max: 469 yuan/month (auto-renew, +215% vs the old Pro)
- Overseas: Coding plan +30–60%, API calls +67–100%
- Turbo monthly cards: 39 yuan for 35M tokens (experience), 99 yuan for 100M tokens (advanced)
⚠️ Price-hike controversy: In a year when Chinese models are broadly cutting prices, Zhipu raised them — citing commercialization needs. Third-party tests put real per-1,000-call cost at ~61.2 yuan, up ~16.6% vs the previous generation. Views are split.
What We Liked (Test Results)
✅ Capability Ceiling of Open-Weight Models
Artificial Analysis Intelligence Index: 50 — #1 among all open-weights models (beating Kimi K2.5's 47), #4 overall, trailing only Claude Opus 4.6 (53) and GPT-5.2 (51). GLM-5 is the closest any open model gets to the closed frontier.
✅ Hard Reasoning Credentials
- Humanity's Last Exam: 50.4% — beats GPT-5.2 and Claude Opus 4.5
- GPQA-Diamond 86.0%, MMLU-Pro 85.7%, AIME 2026 95.8%
- Vending-Bench simulation: $4,432 account balance — best of all open-weights models
✅ Strong Real-World Coding
In Kilo Code's head-to-head, GLM-5 scored 90.5/100 (vs MiniMax M2.5's 88.5) with a perfect score on API implementation. Excels at multi-file reasoning, stack-trace debugging, and agentic terminal workflows.
✅ Hallucination Control Greatly Improved
Zhipu claims a 35-point improvement in hallucination rates over GLM-4.5 — meaningful for accuracy-critical production work.
✅ Cheap + Open Source
MIT license with full self-hosting and fine-tuning. At ~$1/M input, it's less than a tenth of Claude Opus 4.6's $15/M.
What We Didn't Like
❌ Slower Than Rivals
In Kilo Code's test, GLM-5 took 44 minutes vs MiniMax M2.5's ~22 minutes. It's not optimized for speed/cost-efficiency in agent loops.
❌ Makes "Extra" Changes
Sometimes goes beyond minimal fixes — one test even broke API compatibility. Teams needing strict change control should watch it.
❌ Text-Only Base Model
The base GLM-5 doesn't support vision; you need the separate GLM-5V-Turbo for multimodal. One model won't do everything.
❌ Third-Party Chinese Score Slipped
In a third-party ReLE test (~15,000 questions), GLM-5 scored 71.0% — slightly down from GLM-4.7's 71.5%, ranking #6 → #9. Agent/tool-calling (-3.9%) and instruction-following (-1.7%) declined, contrasting with official claims.
❌ Price Hike & Distillation Controversies
Release came with "distilled Claude data" claims and a price hike that upset some users.
GLM vs DeepSeek vs Claude (2026 Comparison)
| Dimension | Zhipu GLM-5 | DeepSeek V4 | Claude Opus 4.6 | Kimi K2.5 |
|---|---|---|---|---|
| Open source | ✅ MIT | ✅ | ❌ | ✅ |
| Parameters | 744B (~40B active) | High | — | High |
| Context | 200K | 128K | 200K | High |
| Intelligence Index | 50 (#1 open) | — | 53 | 47 |
| Coding (SWE-bench) | 77.8% (#1 open) | Excellent | Top tier | Strong |
| Input price / M | ~$1.00 | Very low | $15 | Medium |
| Multimodal | Via V variant | Text-focused | ✅ | Text-focused |
| Standout | Domestic-CN compute | Reasoning king | All-rounder | Long context |
How to choose:
- Strongest open-weights capability → Zhipu GLM-5
- Deep reasoning & cost → DeepSeek V4
- Absolute capability ceiling → Claude Opus
- Long context + generous free tier → Kimi
Who Should Use Zhipu GLM?
✅ Good fit
- Open-source heavy users: self-hosting, fine-tuning, full model control
- Production-grade coding: code quality and factual accuracy matter more than speed
- Teams under export-control constraints: trained on domestic Chinese compute
- Office document automation: .docx/.pdf/.xlsx delivered directly
- Budget-conscious power users: near-frontier capability at a tenth of Claude's price
❌ Poor fit
- Speed-sensitive users: slow agent-loop iteration hurts real-time interaction
- Native multimodal needs: base model is text-only; use the V variant
- Strict change control: can make changes beyond the minimal fix
- Price-sensitive users: hiking prices against the market trend
Final Verdict: Is Zhipu GLM-5 Worth It?
Strongest open source. Production first choice. But accept the price hike and slower speed.
- If you're a developer or enterprise, GLM-5's #1 open-weights Intelligence Index + MIT license + top coding makes it the most compelling open-source option of 2026.
- If you're a speed-sensitive agent developer, weigh its slower iteration against rivals.
- If you're an individual user, the web version is free, or grab the 39-yuan experience card.
GLM-5's core strengths: strongest open capability, cheap, domestic-CN compute, top coding, low hallucination. Its weaknesses: slower, text-only base, price-hike controversy, third-party Chinese score dip. As the "pinnacle of open-source AI," GLM-5 proves Chinese models' dominance on the open track.
Scorecard: Overall 8.4/10 · Open-weights Capability 9.5 · Coding 8.8 · Value 8.0 · Speed 6.0 · Multimodal 6.0 · Documentation 7.0
Frequently Asked Questions
1. Is Zhipu GLM-5 free?
Model weights are completely free (MIT) for commercial self-hosting. The web interface is free; the cloud API is pay-per-token with 39/99-yuan monthly cards.
2. How much does GLM-5 cost?
API: GLM-5 at 4 yuan input / 18 yuan output per million tokens; GLM-5-Code at 6 / 28 yuan. GLM Coding Max subscription: ~469 yuan/month.
3. Is GLM better than DeepSeek?
GLM-5 wins on open-weights capability ceiling, coding, and office file handling. DeepSeek V4 wins on speed, reasoning, and cost. Top capability → GLM; fast and cheap → DeepSeek.
4. Is GLM-5 open source? Commercial use?
Yes — MIT license: free commercial use, fine-tuning, and retraining.
5. Why did GLM prices increase?
Against the trend of price cuts in Chinese AI, Zhipu raised prices, citing commercialization and sustainability needs. Existing subscribers are unchanged; the move is controversial.
6. Does GLM-5 support multimodal?
The base model is text-only; vision requires the separate GLM-5V-Turbo variant.
More AI Tool Reviews
If you found this Zhipu GLM-5 review useful, check out our other hands-on AI reviews:
- Kimi K3 Review 2026
- Qwen 3.7 Review 2026
- GPT-5.6 Review 2026
- Doubao AI Review 2026
- Xiaomi MiMo Review 2026
- Tencent Hunyuan Review 2026
- Claude Review 2026
Replace each YOUR-XXX-REVIEW-URL with the real post URL, or delete the line if that review isn't published yet.
Disclaimer: This article is based on public information and hands-on media tests from February–August 2026. Pricing and features may change; always confirm with the official source. This content is provided for information purposes only and does not constitute a purchase recommendation.
没有评论:
发表评论