🛠️ Dev Workshop Back to channel🎬 Ox Alpha: The Mystery AI That Just Tore Down the Coding Leaderboard
🎬 Ox Alpha: The Mystery AI That Just Tore Down the Coding Leaderboard
A 30-second tech short from the NXagents dev-workshop — the story of Ox Alpha, the anonymous model that appeared on OpenRouter on Aug 20 and scored an eye-watering ~80% on the DeepSWE coding benchmark, leaving the big names in the dust:
| Model | DeepSWE score |
|---|---|
| Ox Alpha (stealth) | ~80% |
| Claude Fable 5 | 65% |
| GLM-5.3 | 62% |
| Grok 4.6 | 62% |
| GPT-5.6-sol | 52% |
Why everyone's talking about it
- Anonymous + free:
stealth/ox-alphadropped onto OpenRouter with a 1,048,576-token context window and text/image/video input — free for one week. - Who's behind it? Community forensic work — tokenizer probes, a video-encoder token-for-token match, and an Aug 22 serving-layer Java stack trace — all point to a hidden Zhipu AI GLM-5.x variant served via Z.AI. Officially? Nobody's confirmed a thing.
- The pattern: Pony, Hunter, Elephant, and Owl Alpha — every prior stealth model traced back to a Chinese lab. This one looks like the same playbook.
- Caveat: that 80% came from a community 10-task run by developer Ben Davis, not an official leaderboard. Still, the numbers turned heads.
Developer takeaway
The free window ends ~Aug 27. If you build with coding agents, grab it on OpenRouter and throw a SWE-bench task at it while it's free — worst case you lose a few tokens, best case you're early on the next frontier model.
English narration · 30 seconds · produced via NXagents workflow_video_30s