Z.ai Confirms Ox Alpha: A Stealth GLM-Series Model with 1M Context, Outperforming GPT-5.6 on Coding Tasks
Z.ai (智谱AI) officially confirmed on August 26, 2026 via Bloomberg that its stealth model "Ox Alpha" belongs to the GLM series. The model had previously appeared anonymously on LiveBench and other benchmarks, prompting community speculation about its origin. Following confirmation, Z.ai announced plans to release the model weights publicly.
Model Specifications
Ox Alpha is positioned as a frontier reasoning model with the following core specs:
- Context window: 1,048,576 tokens (~1M tokens), large enough to hold entire codebases, 500-page specifications, or hours of transcripts in a single prompt
- Max output: 131K tokens
- Input modalities: Text, image, video
- Reasoning approach: Reasoning-first — plans before answering, streams thinking process in real time
- Tool calling: Supports structured output and tool use
- Pricing: Free (oxalpha.com offers a free chat interface with no account required)
Benchmark Performance
Ox Alpha's official site presents results from an independent 10-task real-world coding benchmark. Ox Alpha solved 8 of 10 tasks, leading all frontier models in the comparison:
- Ox Alpha: 8/10
- Claude Fable 5 (max effort)
- GLM-5.3 (max effort)
- GPT-5.6-sol (max effort)
- Grok 4.6 (xhigh effort)
The tasks include anko-typed-variable-bindings, arktype-json-schema-refs, fastapi-deprecation-headers, and others drawn from real engineering scenarios. Ox Alpha solved a task where all reference models scored 1/4 or worse.
However, LiveBench has not yet included Ox Alpha. The current LiveBench leaderboard (2026-06-25 release) top five are: Claude Fable 5 Max (83.0), GPT-5.6 Sol Max (81.0), GPT-5.5 Thinking xHigh (80.2), Claude 5 Opus Thinking Max (80.1), Kimi K3 (79.2). Ox Alpha's LiveBench score remains to be published.
Community Discussion
Hacker News discussion (430 points) focused on several angles:
One user suggested Ox Alpha may be a "small model punching above its weight" — making mistakes on toy benchmarks but able to self-correct all of them, achieving the same outcomes as GPT-5.6 Sol through more tokens, turns, and tool calls.
Others noted the discrepancy between LiveBench scores (reportedly below GPT-5.4 Nano) and the official benchmark (outperforming Fable 5), likely attributable to different test sets — the official 10-task coding benchmark is more targeted, while LiveBench covers 23 tasks across 7 categories.
Open-Source Implications
Bloomberg reports that Z.ai has committed to releasing Ox Alpha's weights. This follows the broader trend of Chinese AI companies open-sourcing flagship models — DeepSeek, Qwen, Moonshot (Kimi), and others have all released open weights recently. If Ox Alpha weights are released, it would add a reasoning model with a 1M-token context window to the open-source ecosystem.
How to Try
Ox Alpha is available at oxalpha.com — free, no account needed, no subscription. The interface is minimal: type a question or prompt, and the model reasons through it in real time. Mobile browser access is supported.
Z.ai's API platform (open.bigmodel.cn) does not yet list Ox Alpha, suggesting it may still be in an independent testing phase with API access not yet available.