On August 20, 2026, AI model-aggregation platform OpenRouter quietly listed a new anonymous model codenamed Ox Alpha (model ID: stealth/ox-alpha). Operated by an unnamed third-party provider, the model offered free input and output tokens during preview and is positioned as an inference system for long-range software engineering, complex reasoning, and production-grade agent tasks. Early results on the DeepSWE software-engineering benchmark sparked community discussion; although later full evaluations declined, its anonymous identity and high-end specs continued to draw developer and analyst attention.

Details: Specs and Early Performance

Ox Alpha supports roughly a 1,048,576-token context window (1,048,576 tokens) and a maximum output of about 131,000 tokens. Inputs support text, images and video, but not audio; it offers tool calling, structured outputs and adjustable reasoning intensity among other agent features. OpenRouter has explicitly stated that it is only responsible for routing and is not the developer or owner; prompts and completions are retained by the provider but are not used for training.

Early attention focused on the DeepSWE benchmark. Developer Ben Davis's preliminary tests across 10 tasks reported Ox Alpha scoring about 80%, ahead of Claude Fable 5 (65%) and GPT-5.6 Sol (52%). Davis cautioned that the sample size was small with high variance. Subsequent full DeepSWE runs showed the model performing roughly at the mid-range level of GPT-5.6 Sol rather than substantially ahead. The official DeepSWE leaderboard has not yet formally listed the model.

OpenCode and other open-source terminal agents also integrated access, claiming near-unlimited usage and theoretical capacity up to 1e15 tokens per day. Within days of launch, platform data showed token processing at the trillions and rapid growth in users and sessions.

The community probed the model across multiple vectors—tokenizer fingerprints, API error strings, video token budgets and code style—and found Ox Alpha highly similar to Z.ai's GLM-5.3 series, including an almost identical tokenizer behavior. Some signals pointed to a stealth test version of the GLM family, but so far no company has publicly claimed the model.

The free preview window was expected to last about a week (OpenRouter's window through around Aug. 24, OpenCode claiming until roughly Aug. 27), after which access may become paid or be adjusted.

Background: Stealth Releases Become the New Normal

In recent years, some AI labs have adopted a "stealth model" strategy, anonymously testing new models on public platforms to collect real user feedback and assess load capacity while avoiding premature exposure of technical details or sparking competitive responses. As a neutral routing layer, OpenRouter has become an attractive entry point for such tests. Ox Alpha's high-spec free configuration and strong agent orientation align with developers' current demand for long-context coding and tool-calling capabilities.

Industry figures such as Stripe CEO Patrick Collison publicly described it as "very impressive," which further amplified interest. At the same time, the provider's retention of prompt data raised questions about privacy and data flows.

Potential Impact: Testing Strategy and Market Competition

1. Impact on developers During the free period, Ox Alpha provided developers a low-cost opportunity to try cutting-edge coding and agent capabilities, especially for long-context software-engineering tasks. Early impressive results should be interpreted cautiously but have already spurred extensive real-world testing and workflow experiments.

2. Impact on model competition An anonymous launch that quickly gained attention shows that stealth releases can effectivel