Ox Alpha is GLM: Evidence, Benchmarks & Access Guide

Ox Alpha is an anonymous AI model released through Open Code and later listed on OpenRouter under the Stealth provider label. The central question is whether this undisclosed model is a next-generation GLM system from Z.ai, potentially related to the GLM multimodal family. Current evidence makes that theory plausible, but it does not establish the model’s identity as fact.

The model attracted attention because it combines a 1,048,576-token context window, multimodal input support, no-data-retention claims, and strong early benchmark results. Open Code also promoted temporary free access during the August 2026 testing period, creating a large amount of community interest and usage.

Video Highlights:

• Ox Alpha reportedly scored 70 out of 80 on a Kingbench evaluation.
• Video-token behavior reportedly matched GLM 5V Turbo across controlled tests.
• Tokenizer counts reportedly matched GLM 5.3 across 25 prompts.
• The model’s writing style and audio-input behavior resemble existing GLM systems.

The available evidence should be read as a model-identification investigation rather than a product announcement. The label “Ox Alpha” may describe a stealth checkpoint, a temporary evaluation alias, or a system that shares technology with GLM without being an unreleased GLM product.

The case for a GLM connection rests on several independent signals. None is conclusive in isolation, but the combination is more informative than a simple comparison of writing style or benchmark scores.

The most notable clue is reported video encoder fingerprinting. Controlled tests compared how different models tokenize video input under changes to frame rate, duration, and resolution. Ox Alpha reportedly produced token counts that matched GLM 5V Turbo across those scenarios. The cited behavior included similar frame sampling, comparable duration scaling near 147 tokens per second, and matching resolution-related changes.

A second clue comes from text tokenization. Ox Alpha reportedly matched GLM 5.3’s tokenizer counts across 25 prompts. A close tokenizer match can suggest a shared or nearly identical vocabulary, although it does not prove that two models were trained by the same organization. Tokenizer reuse can also occur through licensing, infrastructure choices, or compatibility goals.

The investigation also reportedly considered and ruled against several alternative explanations involving DeepSeek, Qwen, Xiaomi, and Western laboratories. Those eliminations are useful because model attribution often depends as much on excluding inconsistent fingerprints as it does on finding a positive match.

However, the theory has important gaps. There is no official statement from Z.ai confirming that Ox Alpha is a GLM model. The public provider label remains Stealth, and OpenRouter’s comparison page does not identify Ox Alpha as a Z.ai release. Until a reveal, technical disclosure, or provider update appears, the responsible conclusion is that Ox Alpha is most likely GLM-related, rather than definitively GLM 5.5.

Ox Alpha’s performance is the other major reason the GLM theory gained attention. In one reported Kingbench test, the model scored 70 out of 80, equivalent to 87.5%. That result placed it second on the cited leaderboard, behind GLM 5.3 at 91.25% and ahead of several other frontier models listed in the same test.

The strongest individual results included perfect scores on a 3JS contact-lens case, a Panda SVG task, a difficult permutation problem, and a Gemma fine-tuning task. The model also performed well on an elevator simulation and a bow-and-arrow task. Lower scores appeared on a folding-table problem and a 3D wrist-clock task, although the latter was described as unusually difficult for many systems.

A separate 10-task DeepSWA subset reportedly produced an 80% result. That score was higher than the comparison results cited for Fable 5, GLM 5.3, Grok 4.6, and GPT 5.6 Soul. Because the sample contained only 10 tasks, the result should be treated as an encouraging signal rather than a stable universal ranking.

The contrast between the two evaluations is especially important. Ox Alpha scored below GLM 5.3 on Kingbench but substantially above it on the cited DeepSWA subset. This may indicate that Ox Alpha is tuned toward agentic coding, tool-oriented work, or multi-step execution rather than optimizing for every form of one-shot generation.

OpenRouter’s public comparison also shows different operational characteristics. Ox Alpha is listed with a p50 latency of 5.74 seconds and throughput of 22 tokens per second, while GLM 5.3 is listed at 3.07 seconds and 36 tokens per second. In practical terms, Ox Alpha may offer strong capability at no listed token cost while responding more slowly in the observed comparison.

A careful evaluation should focus on repeatable tasks rather than a single impressive score. The following process helps compare Ox Alpha with GLM 5.3 or another model while controlling for prompt and workload differences.

For a fair comparison, create a simple scorecard with separate columns for quality, reliability, speed, cost, and integration effort. Do not combine every factor into one number unless the weighting reflects your actual priorities.

Ox Alpha is listed on OpenRouter as an API-accessible model from Stealth. The comparison page shows a 1,048,576-token context window, multimodal input support, tool-use capability, stream cancellation, and no-prompt-training labeling. Some fields remain unknown or unavailable, including quantization details and Artificial Analysis intelligence, coding, and agentic scores.

The model was also promoted as free during the temporary August 2026 access period. The reported free window was expected to end around August 27, 2026, but availability, pricing, and rate limits can change. Check the live model page before planning a long-running application around the temporary terms.

OpenRouter’s comparison page is the most useful public reference for live access details: Ox Alpha vs GLM 5.3 on OpenRouter. The page also makes clear that switching between the two models can use the same API integration with a model-slug change, subject to provider configuration and account access.

The available evidence supports a strong connection between Ox Alpha and the GLM ecosystem, especially because the reported video-token fingerprint and tokenizer behavior align with GLM systems. Its benchmark results also show a capable model that may be particularly effective for coding, structured generation, and agentic workflows.

Still, the identity question remains unresolved. The safest editorial verdict is “probably GLM-related, possibly a next-generation multimodal checkpoint, but not officially confirmed.” That wording preserves the most useful conclusion without turning an investigation into a fact.