Ox Alpha is free on OpenRouter and nobody will say who built it

A stealth coding model with a 1M-token window landed on OpenRouter August 20 with zero lab name attached. Community forensics point at Zhipu GLM infrastructure, and that raises real routing questions for production code.

SaifullahSaifullah
4 min read
Ox Alpha is free on OpenRouter and nobody will say who built it

On August 20, 2026, a frontier-class coding model showed up on OpenRouter with no logo, no press release, and no lab name. Developers called it Ox Alpha. It reads up to 1,048,576 tokens, accepts text, images, and video, and for about a week it costs $0 per token.

That combination is catnip for agent builders. It is also a compliance trap if you paste proprietary code without reading the fine print.

What showed up on the wire

OpenRouter lists the model as stealth/ox-alpha under its generic Stealth provider label. OpenCode announced the same preview the same day and advertised capacity of 100 trillion tokens per day during the free window.

The system prompt, leaked in community threads, instructs the model to identify only as ox-alpha from an undisclosed organization. That is not an accident. Anonymous previews have become a playbook:

Recent stealth dropLater attribution
Pony Alpha on OpenRouterZhipu GLM-5 preview
MiMo-V2-Pro stealthXiaomi
Ling-2.6-flash stealthAnt Group
LongCat-2.0 stealthMeituan

Ox Alpha is the fifth anonymous drop in roughly six months on OpenRouter, and all four prior reveals traced back to Chinese labs.

Anonymous API request routed through OpenRouter with tokenizer probes matching GLM family fingerprints

The benchmark hype vs the full run

The viral number is an 80% pass rate on a 10-task DeepSWE slice, with community comparisons putting Claude Fable 5 near 65% and GPT-5.6 Sol near 52% on the same small set.

Ben Davis, who ran the initial slice, flagged the sample size himself. A 10-task set moves 10 percentage points per task. Full DeepSWE runs from other testers landed closer to 63%, still strong, but not the headline gap Twitter wanted.

If you pick models from screenshots, read why DeepSWE spreads frontier agents apart before you reroute production harnesses.

DeepSWE benchmark comparison chart with small-sample warning for Ox Alpha scores

Forensics beat press releases

The interesting work this weekend was not leaderboard chasing. It was attribution.

Developer unclecode's open-source modelprint tool sends nine infrastructure probes to an endpoint and matches responses against known models. Ox Alpha aligned with GLM-5.3 on six of nine probes, including all four normalized tokenizer counts. No other lab's best candidate cleared more than two.

Separate testers forced malformed requests and got:

  • Error code 1210 in the same shape Z.ai returns for GLM-5.3
  • Java stack traces naming com.wd.paas.api.domain.v4.chat.ChatCompletionRequest
  • Chinese-language parse errors from what looks like Z.ai hosting, not a US proxy

GLM-5.3 shipped August 14 as text-only. Ox Alpha accepts images and video, so if the GLM theory holds, this is likely an unreleased multimodal variant, not the public SKU.

OpenRouter says it is not the developer, owner, or operator of Ox Alpha and only routes traffic. Fair. That also means you own the routing decision.

Why I care as a builder

I route client agent workloads through explicit model allowlists. Anonymous previews fail three checks I actually run:

  1. Data residency: per-model retention language on OpenRouter can contradict the platform EULA. Treat retained prompts as possible until counsel says otherwise.
  2. Export controls: the US Commerce Department added Zhipu AI to the Entity List in January 2025. Z.ai is the same org under a new brand. Enterprise code to an unnamed endpoint is not a neutral experiment.
  3. Eval vs production: free 1M-context models are perfect for agentic coding economics spreadsheets. They are not a default for repos under customer NDAs.

Stripe agreed to acquire OpenRouter on August 19. Patrick Collison publicly called Ox Alpha impressive. Platform consolidation plus anonymous Chinese-lab previews is the story underneath the mystery model.

What I would do this week

  • Run Ox Alpha in a sandbox repo with fake data only
  • Log tokens per task against your current default (see model routing in 2026)
  • Watch whether Z.ai confirms or denies before the free window ends (~August 27)
  • Keep production on named models with signed DPAs

If Ox Alpha is a flash-tier GLM variant that truly trades blows with frontier coding models at lower token cost, that changes local and routed economics. We still need the reveal, not just the fingerprint.

If you are wiring agentic coding into ops and want a second pair of eyes on model allowlists and harness design, book a free discovery call.

Share this post

Related posts