Skip to main content
AI Learning Resources

Quick Overview

This episode of The Code Report, hosted by Fireship, covers the background and official reveal of the anonymous AI model Ox Alpha. The video examines how the model rose to the top of developer leaderboards, details its technical architecture, and puts its coding and vision capabilities to the test.

Key Points

  • 1.The anonymous AI model Ox Alpha gained massive popularity on OpenRouter, serving 42 trillion tokens in its first six days.
  • 2.Chinese AI company Zhipu AI revealed that Ox Alpha was an unannounced test run of its GLM-5.3-Flash model.
  • 3.GLM-5.3-Flash is a 320 billion parameter multimodal mixture-of-experts model running on approximately 100,000 Chinese-made chips.
  • 4.The model is priced at 15 cents per million input tokens and 50 cents per million output tokens, making it up to 40 times cheaper than Claude during promotional pricing.
  • 5.GLM-5.3-Flash demonstrates strong agentic coding and multimodal vision capabilities, though it exhibits slower response speeds and occasional doom loops.

Summary

A mysterious frontier-grade artificial intelligence model named Ox Alpha appeared unexpectedly on OpenRouter, quickly dominating developer attention. Within its first six days online, the model served 42 trillion tokens, accounting for nearly a third of OpenRouter weekly volume and doubling the traffic of its closest competitor, DeepSeek. Developers flocked to the model due to its free availability, one-million token context window, multimodal vision support, and coding capabilities, largely disregarding privacy disclosures stating that user prompts would be retained by an unidentified third party.

Online speculation initially attributed Ox Alpha to various developers including Xiaomi, Zhipu AI, and Google, while some suspected it might be a distilled checkpoint of Claude. Clues quickly emerged linking the model to Chinese research labs, as Ox Alpha followed four earlier anonymous releases on OpenRouter that shared similar naming conventions. Developers examining stack traces, error codes, and tokenizer fingerprints found direct matches to Zhipu AI GLM framework. On August 26, 2026, Zhipu AI officially confirmed that Ox Alpha was a stealth deployment of GLM-5.3-Flash.

GLM-5.3-Flash is a natively multimodal mixture-of-experts model containing 320 billion total parameters and 188 billion active parameters. Zhipu AI released the model weights under an MIT license on Hugging Face and disclosed that its entire inference workload was hosted on 100,000 Chinese-made chips. Following the stealth preview period, commercial API pricing was introduced at 15 cents per million input tokens and 50 cents per million output tokens. With promotional discounts reducing rates further, the model offered API access at roughly one-fortieth the price of proprietary competitors like Claude.

Early viral benchmarks claimed an 80 percent success rate on the DeepSWE benchmark, though verified third-party evaluations settled closer to 58.4 percent. Real-world testing revealed that while the model exhibits verbose reasoning and slower generation speeds, its practical capabilities remain competitive. In a live modernization test of a legacy 2016 AngularJS application named HorseTube, GLM-5.3-Flash systematically rebuilt the user interface using vanilla JavaScript and raw CSS. It correctly resolved visual layout bugs using its vision capabilities and used terminal tooling with FFmpeg to extract video frames, analyze content, and generate relevant metadata.

The Mysterious Rise of Ox Alpha

An anonymous frontier-grade model named Ox Alpha appeared on OpenRouter, rapidly dominating usage leaderboards by processing 42 trillion tokens within six days. Developers heavily used the free tier with a one-million token context window despite telemetry policies stating prompts would be retained by the anonymous host.

Unmasking GLM-5.3-Flash

Digital forensics, stack trace analysis, and tokenization patterns pointed to Chinese origins before Zhipu AI officially claimed the model as GLM-5.3-Flash. The architecture features 320 billion total parameters in a mixture-of-experts configuration, open weights released under an MIT license, and inference powered entirely by 100,000 domestic Chinese chips.

Pricing Structure and Benchmark Realities

Following the stealth trial, pricing was set at 15 cents per million input tokens and 50 cents per million output tokens, dropping even lower during temporary discounts. While early viral claims exaggerated its DeepSWE benchmark score to over 80 percent, verified evaluation placed its actual resolve rate closer to 58.4 percent alongside slower token generation speeds.

Refactoring and Multimodal Practical Testing

In practical coding evaluations, GLM-5.3-Flash successfully modernized a ten-year-old AngularJS codebase into pure HTML, CSS, and vanilla JavaScript while diagnosing CSS overflow bugs using image inputs. It also processed uploaded video content by executing FFmpeg commands to sample frames and generate contextual metadata and comments.

The Bottom Line

The emergence and reveal of Ox Alpha as Zhipu AI GLM-5.3-Flash establishes that Chinese open-weight models can achieve high adoption and competitive coding performance on domestic inference hardware. While independent testing disproves exaggerated initial benchmark claims and highlights issues with output speed, the model provides an exceptionally low cost alternative for agentic workflows. It leaves open the question of how Western frontier labs will respond as high-capability models become increasingly commoditized at fraction-of-a-cent pricing.

FAQ

What is Ox Alpha and what is GLM-5.3-Flash from Zhipu AI?

Ox Alpha was the stealth codename used on OpenRouter for GLM-5.3-Flash, a 320 billion parameter multimodal mixture-of-experts model created by Chinese AI company Zhipu AI.

What are the context window size and API pricing of GLM-5.3-Flash?

GLM-5.3-Flash features a context window of 1,048,576 tokens. Its standard pricing is 15 cents per million input tokens and 50 cents per million output tokens, with cached inputs priced at 3 cents per million tokens.

What hardware was used to serve Ox Alpha during its stealth preview on OpenRouter?

Zhipu AI claimed that the entire stealth run of Ox Alpha was served using approximately 100,000 domestically produced Chinese chips to process online requests.

What was the actual DeepSWE benchmark score achieved by GLM-5.3-Flash compared to viral claims?

While viral screenshots claimed an 80 percent score, the actual verified resolution rate on the DeepSWE benchmark was approximately 58.4 percent.

How did developers deduce the identity of Ox Alpha before the official announcement?

Developers identified the model through Java stack traces matching internal Z.ai APIs, error code similarities, and tokenizer fingerprint tests matching the GLM series.

Worth watching for

Software developers, AI engineers, and technology professionals tracking frontier model benchmarks, API pricing, and open-weight model releases.

  • artificial-intelligence
  • large-language-models
  • glm
  • openrouter
  • machine-learning
  • coding-agents
Watch on YouTubeMore videos from Fireship