Every figure, date and claim the finished picture puts on screen, chased to a source. Checked 22 August 2026. The script was rewritten after the first pass; this file covers what the finished narration says and what the finished picture puts on screen.
| Fact | Value | Source |
|---|---|---|
| Model id | stealth/ox-alpha |
https://openrouter.ai/stealth/ox-alpha |
| Provider name shown | Stealth | https://openrouter.ai/stealth/ox-alpha |
| Appeared | 20 August 2026 | https://openrouter.ai/stealth/ox-alpha |
| Context window | 1,048,576 tokens | https://openrouter.ai/stealth/ox-alpha |
| Max output | 131,072 tokens | https://openrouter.ai/stealth/ox-alpha |
| Input modalities | Text, image, video | https://openrouter.ai/stealth/ox-alpha |
| Tool calling | Yes, tools and tool_choice |
https://openrouter.ai/stealth/ox-alpha |
| Structured output | Yes, response_format, without JSON schema enforcement |
https://openrouter.ai/stealth/ox-alpha |
| Reasoning | Listed as a reasoning model | https://openrouter.ai/stealth/ox-alpha |
| Price, prompt tokens | $0 | https://openrouter.ai/stealth/ox-alpha |
| Price, completion tokens | $0 | https://openrouter.ai/stealth/ox-alpha |
| Preview length | Free for roughly one week from launch | https://officechai.com/ai/stealth-model-ox-alpha-available-for-free-for-a-week-on-openrouter-and-opencode/ |
The listing positions it for coding, sustained agentic work and production workloads. Source: https://openrouter.ai/stealth/ox-alpha
OpenRouter’s model page states: “Prompts and completions for this model are retained by the provider and are not used for training; all other use is governed by the Stealth Model Terms.”
Source: https://openrouter.ai/stealth/ox-alpha
Retained by the provider and not used for training are two separate statements. The provider is not named during the preview, so there is no published entity behind the retention commitment. Other frontends that route to the same model may state different retention terms.
One run, by independent researcher Ben Davis, on DeepSWE, a long horizon software engineering benchmark. Ten deterministic tasks.
| Model | Score on that run |
|---|---|
| Ox Alpha | 80% pass@1, 8 of 10 |
| Claude Fable 5 | 65% |
| GLM-5.3 | 62% |
| Grok 4.6 | 62% |
| GPT-5.6-sol | 52% |
Source: https://finance.biggo.com/news/9dc856ba-634d-467a-bea2-6ba70233113c
The same report carries the caveat directly: sample sizes and attempt counts vary by model, Ox Alpha’s sample is small, and the figures are preliminary. Ten tasks is not a controlled comparison and should not be read as a ranking.
Not officially confirmed. No lab has claimed the model and OpenRouter has not named the provider. What exists is circumstantial technical evidence, gathered by Ben Davis:
Sources: https://finance.biggo.com/news/9dc856ba-634d-467a-bea2-6ba70233113c https://kingy.ai/blog/ox-alpha-glm-5-3-flash-evidence/
Self identification is explicitly not part of this evidence. A model asked what it is will answer from context rather than from fact.
| Fact | Value | Source |
|---|---|---|
| International rebrand to Z.ai | July 2025, around GLM-4.5 | https://www.turingpost.com/p/zhipu |
| GLM lineage on screen | GLM-4.5, GLM-5, GLM-5.2, GLM-5.3 | https://presenc.ai/research/zhipu-glm-model-lineage-2026 |
| GLM-5 max output | 131,072 tokens | https://glm-5.org/ |
| GLM-5.2 | 13 June 2026, context extended to 1M tokens | https://presenc.ai/research/zhipu-glm-model-lineage-2026 |
GLM-5’s stated 131,072 token max output is the same figure Ox Alpha advertises, which is part of why the family is the leading guess.
stealth/ox-alpha. Every attribution in circulation
is fingerprinting, not an announcement.