Trust the catalogue, not the claims
Context window, max output, modality, tool support and price are API facts you can verify in seconds. Architecture claims, parameter counts and benchmark tables usually are not.
Stealth line · explainer
Alpha models: the anonymous line on OpenRouter
Every few weeks an unnamed model shows up on OpenRouter with a large context window, no price and no explanation. The ones named “… Alpha” have followed the same script since March 2026: appear quietly, absorb a burst of community testing, then get claimed by a real lab. This page keeps the whole line in one place, with a clear line between what the API says and what the internet thinks.
The line, in one sentence each
This page is the explainer, not the register. The full table — context windows, input modalities, price while anonymous, price after the reveal — lives on the stealth models index, which is the one place we extend every time a codename appears.
- Hunter Alpha — appeared 2026-03-12, 1M context, no. Xiaomi MiMo-V2.5
- OX Alpha — appeared 2026-08-20, 1M context, no. Z.ai GLM 5.3 Flash
- Union Alpha — appeared 2026-09-16, 256K context, yes. Unbiased Pareto , revealed two days after it appeared — the shortest window of the three.
The pattern, in four moves
- 1. Silent listing. The model appears in the public catalogue under a neutral or anonymous namespace, free, with a context window well above the price point's norm.
- 2. Community testing sprint. Someone notices within hours. Discussion splits into two camps: people benchmarking it, and people guessing whose weights it is — usually from its behaviour and its refusals.
- 3. Identity guessing, with weak evidence. Models are unreliable witnesses about themselves, so identity theories are mostly vibes plus one anecdote. Our trackers label these separately for that reason.
- 4. Reveal and reprice. The maker claims it, the codename becomes a product name, and the free window closes — as happened to both earlier models.
The rules we apply to every release
These are editorial principles, not a procedure. The operational checklist — what to re-read on the catalogue, in what order — is on the stealth models index.
Assume the free window closes
Plan for the endpoint to be repriced or withdrawn. Do not put anonymous endpoints in a production path, and never send customer data or credentials to one.
Test it on your workload, not on the hype
A stealth model's value is entirely task-dependent. Run your own prompts against it and a paid model you already trust, and compare cost per completed task rather than a leaderboard number.
FAQ
What are the Alpha models?
Alpha models is the informal name for a line of anonymous AI models that appeared on OpenRouter as free previews before their makers were disclosed: Hunter Alpha (2026-03-12), OX Alpha (2026-08-20) and Union Alpha (2026-09-16). They share the same pattern: no announcement, no model card, zero pricing, a large context window, and a reveal weeks or months later.
What did the Alpha models turn out to be?
All three are now confirmed: Hunter Alpha was Xiaomi MiMo-V2.5, OX Alpha was Z.ai GLM 5.3 Flash, and Union Alpha — revealed on 2026-09-18, two days after it appeared — was Unbiased Pareto, the first model in the line with image input.
Why do labs release stealth models on OpenRouter?
It is a low-cost way to collect real usage data and preference signals under a neutral name: no brand expectations, no launch marketing, no commitment. If the model performs well the maker claims it; if not, nothing was announced.
Should I build on an Alpha model?
For evaluation and experiments, yes — they are free and capable while they last. For production, no: the provider is anonymous, there is no data-retention or privacy policy, pricing can change to paid without notice, and the endpoint can be withdrawn. Both earlier Alpha models were repriced after their reveal.
How do I tell what a stealth model really is?
You largely cannot, and self-reports are the weakest evidence: models frequently confabulate their own identity, and a system prompt can instruct one to claim a codename. Treat API-reported metadata as fact, bench runs you can reproduce as evidence, and everything else — including a model's own answers about who made it — as speculation.
Union Alpha → Unbiased Pareto Current free models Compare models