Ox Alpha: what a free stealth AI model really costs

Ox Alpha is a stealth model that appeared on OpenRouter with no lab name attached, and the internet immediately started guessing whose it is. TechCrunch covered the guessing in Who's behind the new 'stealth model' Ox Alpha?.

I think the identity question is the least interesting part of this story. If you build software from Sri Lanka on an LKR budget, there is a far more practical question sitting underneath it.

---

🔍 What a stealth model actually is

A stealth model is a model published on a routing platform under a codename, by what OpenRouter labels an anonymous third-party provider. Ox Alpha was listed as "a reasoning model designed for coding, sustained agentic work, and production workload" and shipped free.

That combination is the whole mechanism:

What the lab gets	What you get
Real traffic at real scale, before launch	Frontier-adjacent coding capability at zero cost
Unbiased vibes-testing with no brand halo	No signup friction beyond an OpenRouter key
Live prompts from actual production workloads	A model you cannot cite, pin, or plan around
Free social buzz from the guessing game	No SLA, no changelog, no deprecation notice

Key takeaway: A stealth launch is a paid beta where the currency is your prompts instead of your money. That is not automatically a bad deal, but it is a deal, and you should know you are in it.

---

🕵️ Why the "whose is it?" game is the wrong thing to watch

Per TechCrunch, early speculation pointed at the GLM models from Chinese company Z.ai, and other people argued for an unreleased version of Microsoft's MAI. No consensus emerged. Patrick Collison, whose company is acquiring OpenRouter, called it "very impressive."

Two things worth noticing there, and neither is the answer to the riddle:

1. The endorsement has a stake in it. TechCrunch itself notes the Stripe/OpenRouter acquisition. That does not make the praise wrong. It does mean it is not a neutral benchmark.
2. Nobody produced numbers. The article carries no benchmark table, no eval scores, no context-window figure. The excitement is running entirely on impressions from X and Reddit.

I have watched enough of these cycles to have a rule: when a model arrives with adjectives and no evals, the adjectives are marketing whether or not anyone intended them to be. Treat "very impressive" as a hypothesis, not a result.

---

💰 The free tier is genuinely useful, and that is not a contradiction

None of the above means skip it. If you are a student at UCSC or SLIIT, or a two-person shop in Colombo billing in rupees, a free coding-capable reasoning model is real money saved. A month of moderate agentic coding against a paid frontier API can cost more than a decent monthly data package, and it bills in USD, which is the part that hurts here.

So use it. Just use it in the right slots:

• Good fits: learning, side projects, throwaway scripts, exploratory refactors, generating test fixtures, first drafts of code you will read line by line anyway.
• Bad fits: anything touching client data, credentials, unreleased business logic, medical or financial records, or code under an NDA.
• Worst fit: a production dependency. A codename with no provider can vanish, get renamed, or quietly change weights between Tuesday and Thursday. There is no deprecation policy to appeal to because there is no named party to appeal to.

If you want to know what walking away from the free tier would actually cost, our LLM API price comparison puts the named providers side by side, and the AI token counter will tell you how many tokens your real workload burns before you commit to anything.

Warning: Assume every prompt sent to an anonymous provider is logged and read. You do not know the jurisdiction, the retention window, or the training-data policy, because you do not know the company. That is the definition of a stealth launch.

---

🛠️ How to evaluate a model that will not tell you its name

You cannot read the model card, so you have to generate your own evidence. This is a good habit regardless, and it takes an afternoon:

1. Build a 20-task private eval. Pull real tickets from your own repo. Bugs you already fixed are ideal, because you know the correct answer and the model has not seen your fix.
2. Keep it off the public internet. Anything on GitHub may be in someone's training set. A private eval is the only one that stays honest.
3. Run each task 3–5 times. Reasoning models are stochastic. One good answer is an anecdote; a pass rate is a measurement.
4. Score pass/fail only. No partial credit, no "it was close." Did the test go green or not.
5. Log latency and token counts, not just correctness. A free model that takes 90 seconds per turn is not free when you are the one waiting.
6. Re-run the same suite weekly. With an unnamed provider, silent weight changes are a live risk. Your eval is the only alarm you have.

Signal	What to trust	What to ignore
Public leaderboard rank	Somewhat	Rank changes of one or two places
Screenshots on X	Rarely	Any single cherry-picked prompt
Your own private eval	Yes	Nothing
"Very impressive" from someone with a stake	No	The whole quote

The point of a stealth model is that the usual trust signals are missing. That is not a reason to panic. It just moves the burden of proof onto you, which, honestly, is where it should have been all along.

---

🚀 What this means for you

If you are in Sri Lanka and building on a learning budget, here is what I would actually do this week.

• Try Ox Alpha on throwaway work. Free capability is free capability, and the practice of evaluating an unknown model is a skill worth having.
• Never point it at client data or credentials. Anonymous provider means unknown jurisdiction, unknown retention.
• Do not make it a dependency. Keep a named, paid fallback behind an environment variable so a disappearing codename is a config change, not an outage.
• Write the private eval anyway. You will reuse it for every model release for the next two years. That is the highest-leverage afternoon on this list.
• Ignore the identity guessing. When the lab wants credit, it will take credit. Your build does not get better either way.

The interesting question was never who made Ox Alpha. It is whether you can tell a good model from a hyped one without being told the answer. Right now, most of us cannot, and that is the gap worth closing.