Ox Alpha Unmasked as Z.ai GLM-5.3-Flash, a Cheap Open Coding Model

The stealth model has a name

For about a week in late August, the most talked-about coding model among developers had no known maker. Ox Alpha first appeared on OpenRouter on August 20. The listing billed it as a frontier model for efficient coding, sustained agentic work and production use, with a 1-million-token context window and support for text, image and video input5. On August 26, Beijing-based Z.ai, also known as Zhipu, said it built the model. According to the company, Ox Alpha was an anonymous preview of its new GLM-5.3-Flash, run on OpenRouter and OpenCode to collect user feedback before launch12. Bloomberg was first to report the confirmation11. Z.ai's shares then rose as much as 12% the next day13.

Few people were surprised by the answer, and that tells you something. The reveal matters less for who built the model than for what the episode shows about how coding assistants now get tested, adopted and priced. A model with no brand and no documentation became a production tool for thousands of developers within days. Now that it has a name, it is one of the cheapest serious options for agentic coding.

What the coverage agrees on

On the basic facts, the reporting lines up. GLM-5.3-Flash is a multimodal model that handles text, images and video. It has a context window of up to 1 million tokens, and its weights are on Hugging Face for developers to download and modify12. Z.ai's launch materials describe a mixture-of-experts design with 320 billion total parameters and 18 billion active, released under the MIT license. API pricing is $0.15 per million input tokens, $0.50 per million output tokens and $0.03 per million cached tokens3. Bloomberg reported the same input and output prices and put the model in the low-cost, high-efficiency group with DeepSeek, a group that is pulling users away from premium offerings such as Anthropic's13.

The outlets also agree that the free preview drew heavy use. OpenCode, an open-source coding agent, offered Ox Alpha free for a week with "near unlimited usage" and said its provider could handle 100 trillion tokens per day12. Business Insider noted that this is roughly 100 times the tokens Visa says it uses in a month6. One newsletter reported that Ox Alpha took OpenRouter's top spot with usage about double that of second-place DeepSeek, and that the platform called it its biggest debut ever18. Another tally put OpenRouter volume at about 26 trillion tokens over the first four days, around 2.6 times the previous single-model launch record3.

Why coding-assistant users should pay attention

The usage breakdown is the most telling detail for anyone who works with AI coding tools. On OpenRouter's apps chart, the biggest sources of Ox Alpha traffic were agent harnesses: Claude Code at about 9.3 billion tokens and Hermes Agent at about 9.0 billion, followed by Oh-My-Pi, DeepSeek Harness and Z Code3. In other words, developers took harnesses built for other vendors' models and pointed them at an anonymous, free alternative. My reading is that the coding-assistant layer has become a commodity much faster than the model layer. If a harness can swap models with one config change, the model vendor's lock-in is thin, and price and performance decide who wins.

The model's spec sheet fits these workloads. Reviewers point to the 131,072-token output cap, function calling and structured JSON output as the features that make it suited to agentic work and not just chat10. The 1M-token window is large enough for a big repository or a multi-hour agent session17. The same features matter for AI code review. A reviewer that can take in a whole codebase, screenshots of a broken UI and test logs at once is a different tool from one that sees a single diff. Still, one procurement guide warns that a large context limit does not prove the model can reliably retrieve information across an entire repository. It advises teams to test recall and instruction-following directly15.

Hands-on testing pointed to particular strengths. One review found the model especially good at front-end UI generation, with clean typography and consistent spacing across long pages1. Other early reports praised its code quality but said it "spins a lot" and generates slowly2. OpenRouter's dashboard showed about 50 tokens per second and a 4.45% tool-call error rate over the first three days8. For agentic pipelines and automated review bots, that error rate is the number to watch. When an agent fails roughly one in twenty tool calls, it needs retry logic and checks in the harness. The model alone won't cover it.

Where the benchmarks diverge

The coverage disagrees most on performance. Developer Ben Davis reported an 80% score on the DeepSWE coding-agent benchmark, ahead of Fable at 65% and GPT-5.6 Sol at 52%3. That number drove much of the early hype. One writeup later stressed that it came from a single run by one person and that the model wasn't yet listed on Artificial Analysis8. Another warned that the most-shared result came from a very small user-run test and should be treated as an early signal, not proof10.

Z.ai's own figures are lower. With the release, the company reported a DeepSWE v1.1 score of 63.4, an AutomationBench score of 48.8 and an Artificial Analysis Intelligence Index of 57. All of these are vendor-reported numbers17. Community benchmarks also differ. One early review cited a King Bench score of 70 out of 80, good for second place1. A later analysis put the model at 87.5 on Kingbench, just below GLM-5.3's 91.2517. These may be different runs or versions of the test, and the coverage doesn't settle the question.

Overall, the evidence suggests Ox Alpha is a strong mid-tier frontier coding model, not one that beats the best. The low price is what sets it apart. One analysis compared it to GPT-5.6 Sol's mid tier17. Another source reported that Artificial Analysis scored it 57 at a discounted $0.045 per task, about ten times cheaper than models with similar rankings18. For teams running automated review on every pull request, cost per task matters more than a few points on a leaderboard.

The detective work, and a false lead

Community sleuths identified the model well before Z.ai confirmed it. A fingerprinting tool called modelprint matched Ox Alpha to GLM-5.3 on six of nine infrastructure probes. Davis found token counts identical to GLM-5.3 across 25 prompts, apart from a constant offset17. The video-token usage matched GLM-5V-Turbo, and the model refused audio input the same way Zhipu's models do8.

Not every guess was right. Wccftech first pointed to Zhipu, then changed its view and suggested an unreleased Microsoft MAI model9. A theory that Ox Alpha was a new Gemini gained traction, helped along by cryptic posts from Google DeepMind researchers, until the tokenizer evidence and Z.ai's release ended it17. Z.ai has done this before. It tested GLM-5 anonymously as "Pony Alpha"14.

The chip claim and the privacy fine print

Z.ai says all the traffic during the anonymous test ran on Chinese AI chips. It says its engineers worked around memory and bandwidth limits by optimizing software and how workloads were distributed14. One analysis notes that the figure of roughly 100,000 domestic chips running at a cost comparable to Nvidia hardware is a company claim. It has been widely repeated but not audited17. If the claim holds up, it suggests US export controls have slowed China's AI inference less than intended.

For engineering teams, the privacy terms matter more right now. OpenRouter said the preview provider kept prompts and completions but did not train on them. OpenCode advertised zero data retention for its own Ox Alpha traffic3. During the anonymous phase, security analysts told organizations to avoid sending sensitive code or secrets, to limit the permissions given to agents and to validate generated output before deploying it9. That advice still applies to any team that sent proprietary code through the preview. Those historical experiments need their data handling documented15.

What changes now

The free period is over. The old preview IDs are deprecated, and the named model is now available through paid providers on OpenRouter, which lists a context window of 1,310,720 tokens15. Self-hosting is another option: SGLang, vLLM and TokenSpeed all serve the model3. That means a team can trade per-token fees for a one-time hardware cost, including for code review it runs itself.

The broader point is that coding assistants have become the main testing ground for open-weight Chinese models. Z.ai shipped GLM-5.3 earlier in August and said it rivals Anthropic's Fable 5 on some benchmarks11. Each release like this adds pressure on expensive frontier labs19. Ox Alpha shows that developers will adopt a capable model before they know who built it. For OpenAI and Anthropic, that weakens one of their main advantages.