Ox Alpha: The Free 1M-Token Stealth AI Model Nobody Can Identify — and How to Try It This Week

Introduction: A Frontier Model Appeared, and Nobody Will Say Whose

On the evening of August 20, 2026, a new AI model quietly appeared on OpenRouter under the provider name “stealth.” It’s called Ox Alpha, it costs nothing to use, it swallows a million tokens of context, and absolutely no one will say who built it. OpenRouter’s listing describes it only as “a reasoning model designed for coding, sustained agentic work, and production workloads,” developed by “a third-party provider who has chosen to remain anonymous during this preview.”

Within hours, the mystery went viral. OpenCode — the open-source terminal coding agent, hosting the model in parallel — announced it would be free for a week with “near unlimited usage” and claimed capacity for 100 trillion tokens a day; the post passed 6.8 million views. Stripe CEO Patrick Collison, whose company is acquiring OpenRouter, called the model “very impressive.” By August 23, TechCrunch was asking the question on everyone’s mind: who’s behind Ox Alpha?

If you use AI coding tools, this isn’t just a parlor game: a frontier-class model is free to test for a few more days. Here’s what it does, who might have built it, and how to plug it into your stack.

What Ox Alpha Actually Is

Strip away the mystery and the spec sheet is genuinely serious — the kind of numbers that until recently belonged exclusively to flagship models from the big labs:

Two numbers matter most for real work. The 1M-token context means you can load an entire codebase or a full API surface in one shot — territory we last saw with DeepSeek’s 1M-context release. And the 131K-token completion ceiling means a single turn can emit a large multi-file refactor without truncation — exactly what “sustained agentic work” requires.

The Guessing Game: GLM, MiMo, or MAI?

Since no one has claimed the model, the community has turned to forensics. Three suspects lead the board:

• Z.ai’s GLM family. A developer ran tokenizer fingerprinting across 25 prompts and found Ox Alpha’s raw token counts line up almost exactly with GLM-5.3 — off by a constant 75-token wrapper every time. Others noted it stumbles on the same “dirty token” that has tripped up Qwen and GLM-family models.
• Xiaomi’s MiMo team. Xiaomi has run this playbook before: MiMo-V2-Pro spent time on OpenRouter under the alias “Hunter Alpha” before Xiaomi claimed it. Ox Alpha’s profile — million-token context, native multimodality, an agentic lean — matches what MiMo has been building.
• Microsoft’s MAI. Early coverage leaned toward GLM, then updated to suggest Ox Alpha could be an unreleased version of Microsoft’s in-house MAI models.

Nobody has confirmed anything, and analyst Andrew Curran noted that confidence has only decreased as more evidence arrived. That’s the point of a stealth drop: the lab gets real-world traffic against production workloads — not curated benchmarks — before putting a name and a price tag on the thing.

The Privacy Catch: Two Hosts, Two Stories

Here’s the detail that should gate what you send through Ox Alpha. The two hosts describe its data handling differently:

• OpenRouter says prompts and completions are retained by the anonymous provider, though not used for training.
• OpenCode promises zero data retention and no training.

Those statements can’t both be fully true for the same underlying route. The practical rule for the free week: treat Ox Alpha like any anonymous endpoint — great for open-source repos and experiments; think twice before piping in proprietary code, customer data, or secrets. The model is only half the risk equation; the pipe matters just as much.

How to Try Ox Alpha Before the Free Window Closes

The preview runs roughly a week from August 20, so the clock is ticking. Three ways in:

• OpenRouter playground or API. Select stealth/ox-alpha and chat with it directly, or call it from any OpenAI-compatible client — the fastest way to run your own tokenizer tests.
• OpenCode terminal agent. The free tier is built in, usage is “near unlimited,” and it doesn’t count against OpenCode Go subscriptions during the window.
• Your existing coding tools. Anything that accepts a custom OpenRouter endpoint — setups built around Claude Code, Cursor, Windsurf, or GitHub Copilot rivals like Gemini CLI — can be pointed at the model for an afternoon of benchmarking.

A fair test beats a fun one: pick two or three real tickets from your backlog, run them through Ox Alpha and your current default, and compare correctness, not vibes. Models like DeepSeek already proved cheap frontier capacity is the defining buyer’s market of 2026.

Why Stealth Drops Matter for the AI Tools You Use

Beyond the whodunit, Ox Alpha marks a real shift in the model market. On OpenRouter, the share of tokens processed by US-built models has reportedly fallen from around 70% a year ago to roughly 30% today, as labs like DeepSeek, Tencent, Xiaomi, and Z.ai eat the difference. Free stealth drops are the sharpest weapon in that fight: developers get frontier output at zero cost, and the lab gets a week of unfiltered stress-testing before committing to a name or a price.

For tool buyers, the lesson is the one Nvidia’s harness research reinforced: the scaffolding around a model matters as much as the model. Keep your stack model-agnostic, keep a cheap default for routine work, and treat weeks like this one as a free upgrade window for your pipeline.

Frequently Asked Questions