An anonymous AI model codenamed Ox Alpha appeared on OpenRouter on August 20, 2026. It offers free access during a preview window that ends around August 27. The model is accessible via the route stealth/ox-alpha.

It features a 1,048,576-token context window and a 131,072-token maximum output. It supports text, image, and video inputs natively. Tool calling and structured JSON output are built in. The architecture is reasoning-first with a visible thinking stream.

Independent researcher Ben Davis tested Ox Alpha against 10 tasks from the DeepSWE software engineering benchmark. He reported an 80% pass rate — 8 of 10 tasks solved on the first attempt. That result outperformed several frontier models tested on the same subset. Claude Fable 5 scored 65%. GLM-5.3 and Grok 4.6 each scored 62%. GPT-5.6 Sol scored 52%. Davis emphasized the small sample size. He noted the DeepSWE official leaderboard has not yet included Ox Alpha.

What's new

  • Model label: stealth/ox-alpha (Ox Alpha)
  • Context window: 1,048,576 tokens (1M)
  • Max output: 131,072 tokens
  • Modalities: Text, image, video
  • Features: Tool calling, structured output (JSON via response_format), reasoning-first architecture with visible thinking stream
  • Pricing: Free during stealth preview (approximately through August 27, 2026)
  • Access: OpenRouter route stealth/ox-alpha; chat interface at oxalpha.com (no account required)
  • DeepSWE result: 80% Pass@1 on 10-task subset (8/10 tasks solved)
  • Reference model scores (same 10 tasks): Fable 5 65%, GLM-5.3 62%, GPT-5.6 Sol 52%, Grok 4.6 62%
  • LiveBench listing: Label "ox-alpha-max" shows Overall 69.2, Reasoning 76.6, Coding 75.8, Agentic coding 52.6, Mathematics 77.5, Data analysis 75.8, Language 66.1, Instruction following 60.3 (snapshot LiveBench-2026-06-25; third-party observation, not route-verified)
  • Architectural fingerprint: Estimated ~744B total parameters, ~40B active via MoE; video encoder behavior matches Zhipu AI GLM-5V-Turbo (99% confidence per Ben Davis)

Why it matters

Ox Alpha combines a 1M-token context window with strong coding and agentic performance at zero cost during its preview period. That combination lowers the barrier for developers to test long-horizon software engineering tasks — debugging entire codebases, multi-file refactors, and sustained agent workflows — without chunking or retrieval workarounds. The model's integration into OpenCode, which claims up to 100 trillion tokens per day of call capacity, signals early production adoption for autonomous coding agents.

The DeepSWE subset is small: 10 tasks, single-run Pass@1. The benchmark differs from SWE-bench Verified, where frontier models score 96%+. The LiveBench snapshot for "ox-alpha-max" cannot be tied to the Tokenra/OpenRouter route. The model's origin remains unconfirmed by any lab. Until route-verified, reproducible results are published under matched conditions, the 80% figure should be treated as a directional indicator rather than a definitive ranking.

Our take

Ox Alpha follows a familiar stealth-release pattern: an anonymous frontier-class model appears on OpenRouter, free for a week, and the community races to benchmark it before the lab reveals itself. The technical fingerprinting pointing to Zhipu AI's GLM-5.x series is compelling, especially the video encoder match. Without official confirmation, the 80% DeepSWE result remains an impressive but unverified data point. For developers, the 1M context window and free access offer a rare chance to stress-test a reasoning model on real production workloads at scale.

Sources