English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

Mysterious Model Ox Alpha Appears on OpenRouter, DeepSWE Benchmark Beats Claude — Fingerprints Point to Zhipu (GLM)

Forum topic · QianXun · 2026-08-23

Summary

OpenRouter quietly listed an anonymous model called Ox Alpha on August 20, offering free access for one week. Developer Ben Davis benchmarked it on DeepSWE, a suite that evaluates a model's ability to read real codebases, locate vulnerabilities, and produce working patches. Across ten tasks, Ox Alpha scored about 80%, compared with Claude Fable 5 at 65%, GLM-5.3 at 62%, Grok 4.6 at 62%, and GPT-5.6-sol at 52%. Listed specs include a 1,048,576-token context window, multimodal text/image/video input, a 131,072-token output limit, and free preview access. Coding-agent platform OpenCode also integrated it into a free tier claiming one quadrillion tokens per day. Researchers ran tokenizer fingerprinting across 25 prompts and found Ox Alpha's token counts almost identical to GLM-5.3, differing only by a constant 75-token wrapper. Video-encoder behavior and API response patterns also match Zhipu's architecture. The article, summarized from IT Home (2026/8/22), notes these clues collectively point to Zhipu as the likely origin.

Background

On August 20, OpenRouter — one of the largest AI model aggregation platforms — quietly listed an anonymous model called Ox Alpha, made free for one week. Its origin was undisclosed at launch.

DeepSWE Benchmark Results

Developer Ben Davis ran Ox Alpha against DeepSWE, a benchmark that measures an AI model's ability to read real-world codebases, locate vulnerabilities, and generate working patches.

Across ten tasks, Ox Alpha reached roughly 80%. Davis noted the result was hard to interpret given the model's unknown provenance.

| Model | DeepSWE average (10 tasks) | |---|---| | Ox Alpha (anonymous) | 80% | | Claude Fable 5 (max) | 65% | | GLM-5.3 (max) | 62% | | Grok 4.6 (xhigh) | 62% | | GPT-5.6-sol (max) | 52% |

Listed Specifications

  • Context window: 1,048,576 tokens
  • Input modalities: text, image, video
  • Max output: 131,072 tokens
  • Pricing: free during preview
  • The terminal coding-agent platform OpenCode also added Ox Alpha to its free tier, claiming daily capacity of 100 quadrillion tokens (1e17).

    Fingerprint Analysis — Pointing to Zhipu

    Investigators ran tokenizer fingerprinting across 25 prompts:

  • Ox Alpha's token counts were almost identical to GLM-5.3, differing only by a constant 75-token wrapper layer.
  • Video-encoder behavior aligned with Zhipu's architecture.
  • API response characteristics were consistent with Zhipu's stack.
  • These combined signals point toward Zhipu as the likely creator, though no party has officially confirmed it.

    Source

    Adapted from IT Home (2026/8/22): *《神秘 Ox Alpha AI 模型限免上架:DeepSWE 跑分碾压 Claude Fable 5,线索指向智谱》*.

    Key Points

  • Anonymous model Ox Alpha launched free on OpenRouter; topped DeepSWE with ~80%.
  • Specs: 1M-token context, multimodal input, 131K-token output, free preview.
  • Tokenizer and API fingerprints closely match GLM-5.3, suggesting Zhipu as origin.

Tags

#ox-alpha#openrouter#deepswe#benchmark#zhipu#glm#claude#llm

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178633859