Zhipu confirmed on August 26 that Ox Alpha, a model that had been appearing anonymously on OpenRouter, is a new iteration in its own GLM series, and said it would release the model weights that same evening. Developers had been guessing at the model's identity ever since it went live, with no official response — until usage volume pushed it to the top of the platform's leaderboard.
Ox Alpha appeared on OpenRouter and OpenCode on August 20, listed only under a provider tag of "stealth" with the model ID stealth/ox-alpha. It offers a 1.04 million token context window, accepts text, image and video input, and caps output at around 128,000 tokens — all free during the preview period. There was no launch event, no technical report, not even an official blurb.
Tokenization fingerprints gave it away first
The community didn't wait for an official statement. Around August 22, several independent tests pointed the finger at Zhipu. Testers fed the same batches of prompts to both Ox Alpha and GLM-5.3 and compared the tokenization output: across 25 to 30 different prompts, token counts on both sides matched almost exactly, off by a constant 75-token wrapper overhead. The encoding overhead for video input also lined up with Zhipu's GLM-5V-Turbo. A more direct clue came from a Java stack trace triggered by a malformed request, which contained package names from Zhipu's internal API along with error code 1214.
At the time, all of this evidence stayed at "strongly suspected." One researcher put the confidence level at 0.98, but Zhipu, OpenRouter and the anonymous provider all stayed silent. This confirmation effectively closes out six days of speculation.
A free preview pushed usage to the top of the chart
What put Ox Alpha in the spotlight was usage volume. Four days after launch, OpenRouter noted that its daily processing volume was approaching 6 trillion tokens; it now sits at No. 1 for usage on the platform, with public reports putting its volume at more than twice DeepSeek's. The free preview is expected to run for roughly another week, with official pricing yet to be announced.
Turning a free preview into a distribution tactic is something Chinese model makers have gotten increasingly good at over the past couple of years. Getting developers using the model inside real workflows first, and talking about pricing later, moves faster than a launch event full of benchmark numbers. That approach comes with a real cost, though: at a rough daily processing volume of 6 trillion tokens, even at an estimated cost of a few tenths of a yuan per million tokens, the daily inference bill lands in the millions of yuan — a customer-acquisition budget that adds up to a substantial sum over a week. What it buys is a leaderboard spot, real usage data, and a batch of developers who have already wired the API into their projects.
The benchmark numbers need a discount
The most widely circulated early claim was that Ox Alpha beat Claude Fable 5 and GPT-5.6 Sol on DeepSWE. That conclusion came from a sample of just 10 tasks. Two later, independent full runs across all 113 tasks put the score at around 63% — roughly on par with GPT-5.6 Sol. That's still first-tier performance, but a long way from "crushing it."
Small-sample benchmark results spread on social media far faster than the corrected numbers do, a pattern that has repeated across several recent model launches. For anyone making an actual technology choice, waiting for the full evaluation and the open weights before drawing conclusions is a much cheaper way to go.
After the weights go public
Once Zhipu releases the weights, Ox Alpha won't stay a "closed API you get to use for free for a week." The GLM series has gone the open-source route multiple times before, and once model weights go public, third-party inference providers typically deploy it and the community starts quantizing and fine-tuning it within days.
For Chinese developers, two things are worth watching from here: first, where the official pricing lands, and whether it keeps GLM's usual low-price strategy; second, the licensing terms attached to the open weights — a fully permissive Apache-style license, or one with commercial restrictions. The first determines whether it can stay in daily workflows long-term; the second determines whether it can get into the enterprise.
Sources: Bloomberg, OpenRouter, CocoLoop, IT Home; context length, free-preview duration and the DeepSWE re-run results were each verified against platform pages and public test methodology.