Ox Alpha: The Free, Anonymous Model Challenging Claude Fable and GPT-5.6

Preface

Ox Alpha surfaced in late August as a stealth multimodal model available on routing platforms with no clear author attached. This article summarizes what is known about its performance, availability, and the theories about its origin. It explains why the model captured attention so quickly, how benchmark claims spread, and what independent analysis suggests about its likely maker. The purpose is to lay out the evidence and context clearly so readers can judge the claims and implications without hype.

Lazy bag

Ox Alpha launched publicly on routing services as a free, high-capacity model. Early, small-sample results showed it beating Claude Fable 5 and GPT-5.6 Sol on coding tasks, but larger full-run scores put it roughly level with GPT-5.6. The provider has not identified itself; independent fingerprinting points toward Zhipu AI's GLM-5.3 family, though the model's multimodal video support suggests a likely unreleased upgrade. Availability, token throughput, and the stealth release pattern explain why the lab may be testing usage before a formal announcement.

Main Body

On August 20, Ox Alpha appeared on several routing portals and gateways as a publicly callable model with no company name attached. The listing described it as a reasoning-focused model suited to coding, long-running agentic tasks, and production workloads. Within hours and days, the model drew attention for three reasons: it was free to use for a limited window; it accepted text, image, and video inputs while returning text outputs; and an early virally shared sample suggested unusually strong coding performance.

The viral claim originated from a developer who ran a 10-task sample on DeepSWE, a benchmark that measures how often a coding agent resolves real GitHub issues correctly on its first attempt. That initial small sample credited Ox Alpha with an 80% pass rate, ahead of Claude Fable 5 at 65% and GPT-5.6 Sol at 52%. Because the sample was small, the finding was exciting but statistically weak — it could reflect variance in task selection rather than a definitive superiority.

Subsequently, more complete 113-task runs showed Ox Alpha scoring near 63%, broadly similar to GPT-5.6 Sol and not obviously beyond state-of-the-art. Those numbers tempered the initial excitement: while Ox Alpha is competitive on a demanding real-world code benchmark, the early viral claim overstated the strength of evidence.

Beyond performance, Ox Alpha's release mechanics were notable. The model offered a very large single-context window — reportedly up to roughly one million tokens — and accepted multiple modalities (text, image, video). It also supported tool and function calling, though its JSON output did not enforce schemas, which can be a practical shortcoming for agent developers who expect structured responses from tools.

The model was made available through multiple routes: OpenRouter listed it as stealth/ox-alpha; other gateways like OpenCode, Cline, and Nous Research provided access as well. Providers advertised generous throughput capacities (OpenCode claimed capacity around 100 trillion tokens per day; Nous Research suggested even higher figures), suggesting an intent to support heavy real-world usage during the preview window. That combination of capability and no-cost testing is an attractive way for a lab to collect production feedback.

Attention intensified when notable figures and companies commented. A short endorsement from a well-known tech executive helped fuel conversation and broader testing. Still, no lab publicly claimed authorship. That anonymity launched a flurry of speculation. Candidates proposed by observers included large Chinese and international labs — Microsoft (MAI), Xiaomi (MiMo), Alibaba (Qwen), Google (Gemini), DeepSeek, and Zhipu AI — with advocates for each theory citing various behavioral clues.

Careful comparison narrowed the field. Several proposed matches failed to align on concrete technical points: some candidate models accept audio input (which Ox Alpha rejects), others lack video capability, and some use different tokenizers and video encoders. Independent fingerprinting efforts were more decisive: multiple tests matched Ox Alpha’s tokenizer to GLM-5.3 and showed video token consumption patterns consistent with GLM-5V-Turbo. Ox Alpha also shared distinctive error modes and audio rejection behavior observed in GLM-family models. Those converging signs made Zhipu AI’s GLM line the most plausible source.

One wrinkle remained. GLM-5.3 had been released as text-only shortly before Ox Alpha appeared with multimodal support, which suggests that Ox Alpha may be a multimodal revision or internal variant rather than a simple repackaging. Analysts began calling the hypothesis GLM-5.3 Flash to reflect the notion of a near-relative with added video capabilities. Zhipu AI did not confirm or deny the connection publicly during the preview window.

Ox Alpha also illustrates the growing practice of “stealth” releases: labs publish powerful models through routing platforms without attribution to collect usage data and iterate before a formal public launch. Chinese labs in particular have used this approach repeatedly; prior anonymous releases were later traced back to known product lines within days to months. The pattern helps labs test models at scale and observe real-world interactions while retaining flexibility over final disclosures and packaging.

For developers and organizations, the situation offers both opportunity and caution. Free limited-time access lets many try the model and gather empirical impressions. But uncertainty about provenance, undocumented output schema constraints, and the ephemeral free window all argue for careful evaluation before relying on the model in production systems. Benchmarks like DeepSWE provide a useful, objective baseline for comparison, but small-sample viral results should not replace thorough testing across representative tasks.

In short, Ox Alpha is a competitive, high-capacity multimodal model offered anonymously for public preview. Early viral samples suggested it might outperform leading rivals on some coding tasks, but fuller runs put it roughly level with top models. Independent technical fingerprints most strongly implicate the GLM-5.3 family as the underlying architecture, possibly in an unreleased multimodal variant. The stealth release fits a familiar pattern for labs wanting real-world feedback before a formal reveal. Observers should watch for follow-up disclosures or additional technical analyses to confirm the model’s origin and capabilities.

Key Insights Table