Rumor Mill
AI Minute Newsroom
2026-08-25
Rumour: developers say the anonymous model beating GPT-5.6 on OpenRouter is Zhipu's — the evidence is a tokenizer and a stack trace
This is an unverified community claim, not a confirmed release. A model called Ox Alpha appeared on OpenRouter around 20 August with a 1,048,576-token context window, 131,072-token maximum output, free access for a preview week, and no company name attached. Developers running forensics on it report that its tokenizer matches Zhipu AI's GLM series on every probe they tried, that a Java stack trace surfaced Zhipu-internal API class names, and that its error dialect matches Z.ai-hosted GLM models. Zhipu has not claimed it and nobody else has either. The claims that it beats GPT-5.6 Sol and Fable 5 at coding come from small user-run tests, not published benchmarks. Note also: with an anonymous provider, assume anything you send it is retained by a party you cannot name.
Why it mattersStealth previews are now a routine launch tactic — Z.ai's own GLM-5, Xiaomi's MiMo-V2-Pro, Ant Group's Ling-2.6-flash and Meituan's LongCat-2.0 all ran anonymously first and were claimed once the preview closed. The lab gets clean, unbiased reactions before the brand is attached; the users pay for it in prompts sent to an unidentified server. If the pattern holds, whoever built Ox Alpha will put their name on it within a couple of weeks.
⚠ Unverified rumor — treat with caution
Read in the app — free, in 9 languages
Related stories
A lab lost the 40,000 chips it was going to train on. So it sold the factory to the company that makes them.
2026-08-25Developers noticed the effort dial reading low. Anthropic says it is testing serving configs, and the number means something different this week.
2026-08-24A search index built for coding agents rather than people: 70 million READMEs, issues and pull requests, and no key needed to try it
2026-08-23The standard that plugs AI into your tools was built for a human clicking 'allow'. Its new roadmap admits the caller is now a machine.
2026-08-23Alibaba trained a model on a rack of 100 real phones — it now finishes 92 out of 100 tasks on an actual handset
2026-08-23