20260615-海通国际-Frontier_Models_Suspension_Benefits_Chinese_Open_Ecosystem_3页_311kb
报告摘要
Frontier Models Suspension Benefits Chinese Open Ecosystem Summary
Core Content
On June 12, 2026, the US Commerce (BIS) ordered Anthropic to suspend Fable 5 and Mythos 5 for all foreign nationals, citing national security concerns. This move caused Anthropic to fully offline both models for every customer, marking the first time a leading lab has taken a deployed model offline due to direct government intervention. The decision highlights the growing operational risks associated with access to frontier models and the potential policy risks of relying on a single vendor.
On June 13, 2026, Zhipu released GLM-5.2 and Moonshot launched Kimi K2.7 Code, both of which are open-weight models designed for coding and agent-first applications. These releases occurred as Anthropic's models were suspended, indicating a strategic timing to fill the gap left by the US action.
Main Points
- Government Intervention: The US government's decision to suspend Anthropic's models for foreign users signals a new era of regulatory scrutiny and control over AI technologies.
- Open Ecosystem Response: Chinese companies have responded by launching new models with open weights and permissively-licensed options, such as GLM-5.2 and Kimi K2.7 Code.
- Market Positioning: GLM-5.2 is positioned as a permissively-licensed alternative to Claude Code and GPT-5.5, with a 1M context window and a max output of 128K tokens.
- Kimi K2.7 Code: This model is open-weight MoE, with 1T total parameters and 32B activated parameters, offering a 256K context window and improved efficiency in thinking-token usage.
- Performance: Both models show strong performance in coding and agent tasks, with Kimi K2.7 Code beating Opus 4.8 on one benchmark but still trailing GPT-5.5 and Opus 4.8 on most.
- Cost Efficiency: The Chinese models are designed to be more cost-efficient, offering near-frontier capabilities at a fraction of the price of their overseas counterparts.
- Strategic Timing: The release of these models coincided with the suspension of Anthropic's models, reinforcing the Chinese open ecosystem's ability to capture market share.
Key Information
-
GLM-5.2:
- Context window: 1M
- Max output: 128K tokens
- Two thinking-effort presets: High and Max (Max default for coding)
- Pricing: Lite ~$18/month (~400 prompts/week); Max ~8,000 prompts/week
- Compatibility: First-day compatibility with major agentic CLIs
-
Kimi K2.7 Code:
- Open-weight MoE
- 1T total / 32B activated parameters
- 256K context window
- MLA + MoonViT vision encoder
- Thinking-only, cutting thinking-token usage ~30% vs K2.6
- Full benchmark table included
-
Market Trends:
- Chinese companies are leveraging multi-model deployment and rising capabilities to drive inferencing demand.
- The open ecosystem is gaining traction as overseas models face regulatory challenges.
Risks
- Restoration Risk: There is a risk that Fable 5 access could be quickly restored, which might mitigate the impact on the Chinese open ecosystem.
- Pricing Trends: The pricing of frontier models in China has shown significant increases, particularly for Qwen3.7-Max and Doubao Seed 2.0 Pro, indicating a potential shift in cost dynamics.
Figures and Data
Fig.1 Frontier Model API Pricing: Overseas vs China — List Price, USD/MTok, and Change vs Prior Version
| Region | Company | Model | Release | List ccy | Input (list) | Output (list) | Input (US$/MTok) | Output (US$/MTok) | Input chg vs prev ver | Output chg vs prev ver |
|---|---|---|---|---|---|---|---|---|---|---|
| Overseas | Anthropic | Claude Fable 5 | 2026-06-09 | USD | 10.00 | 50.00 | 10.00 | 50.00 | +100% | +100% |
| Overseas | OpenAI | GPT-5.5 | 2026-04-23 | USD | 5.00 | 30.00 | 5.00 | 30.00 | +100% | +100% |
| Overseas | Gemini 3.1 Pro | 2026-02-19 | USD | 2.00 | 12.00 | 2.00 | 12.00 | flat | flat | |
| Overseas | xAI | Grok 4.3 | 2026-04-30 | USD | 1.25 | 2.50 | 1.25 | 2.50 | flat | flat |
| China | Alibaba | Qwen3.7-Max | 2026-05-22 | RMB | 17.00 | 51.00 | 2.50 | 7.50 | +525% | +213% |
| China | Moonshot | Kimi K2.6 | 2026-04-20 | USD | 6.46 | 27.20 | 0.95 | 4.00 | +58% | +33% |
| China | Zhipu | GLM-5.1 | 2026-03-27 | RMB | 6.00 | 24.00 | 0.88 | 3.53 | +50% | +33% |
| China | ByteDance | Doubao Seed 2.0 Pro | 2026-02-14 | RMB | 3.20 | 16.00 | 0.47 | 2.35 | +300% | +100% |
| China | MiniMax | MiniMax M3 | 2026-06-01 | RMB | 2.10 | 8.40 | 0.31 | 1.24 | flat | flat |
| China | DeepSeek | DeepSeek V4-Pro | 2026-04-24 | RMB | 3.00 | 6.00 | 0.44 | 0.88 | +50% | +100% |
Fig.2 GLM & Kimi Spec Sheet: New vs Prior Generation
| Metric | GLM-5.2 (new) | GLM-5.1 (prev) | Kimi K2.7 Code (new) | Kimi K2.6 (prev) |
|---|---|---|---|---|
| Release date | Jun 13, 2026 | Apr 7, 2026 | 12-Jun-26 | Apr 20, 2026 |
| Positioning | Coding/agent-first flagship | Long-horizon agentic coding-first flagship | Coding-specialized agentic | General multimodal agentic flagship |
| Architecture | MoE (details undisclosed; likely GlmMoeDSA per 5.1) | MoE (GlmMoeDSA): MLA + DSA + MTP | MoE + MLA + MoonViT | MoE; 384 experts (8+1); 61 layers; 64 heads; MLA + MoonViT |
| Total / active params | Not disclosed | 744B / 40B (256 routed +1 shared, top-8) | 1T / 32B | 1T / 32B |
| Context window | 1M | 200K | 256K (262,144) | 256K (262,144) |
| Long-horizon | 1M context for multi-hour runs | 8-hour autonomous execution | Long-context coding | Agent Swarm: 300 sub-agents, 4,000 steps |
Analysts
-
王宪策 Nathan
SFC HK执业证书编号：BKS605 -
吴叡霖
SFC HK执业证书编号：BVA376
Disclaimer
This report is for informational purposes only and does not constitute investment advice, recommendations, or any form of guarantee. The information and views expressed in this report are based on data and assumptions as of the date of publication and may be subject to change. Readers are advised to consult the full report on the HTI research website for a comprehensive understanding of the analysis and to make informed investment decisions.
试读结束，高清完整版pdf/doc/ppt,请点下载