AI CompendiumA public chart of the language
Home / Evaluation / Stealth Model
Evaluation

Stealth Model

also called Anonymous preview, Cloaked model, Alpha listing

A model released under a codename, with no lab attached, to be judged on behaviour alone.

Definition

A model made available through a gateway or API under a playful codename, with the developing lab deliberately unnamed. Users get the model and its spec — usually free, for a limited preview — but no vendor, model card or benchmark sheet. The point is to collect real usage and unbiased first impressions before the lab puts its name on it.

The lifecycle repeats: codename launch, free usage at scale, community fingerprinting, then reveal under the real name and a move to a paid, named ID. The codename listing is temporary by design — anything that hardcodes it should expect to break or start costing money.

Fingerprinting is real evidence but limited evidence. Matching a tokenizer, error payloads or knowledge cutoffs establishes shared serving infrastructure, not model identity: one lab can serve two different checkpoints from the same stack, and the crowd’s first guess has been wrong before. Read the gateway’s stealth terms too — prompts may be retained, and may be used for training.

The track record says to plan for a reveal: of thirteen tracked stealth listings, seven were eventually claimed by a lab, and the free window usually closed at reveal. Ox Alpha lasted about six days before Z.ai named it GLM-5.3-Flash. Space Bunny, listed September 23, 2026, was still unclaimed when OpenRouter’s preview ended October 5, with OpenCode extending free access.

How people use it
Deciding whether to build on a free anonymous model.
"Treat the codename as a trial, not a dependency. Keep the model ID in config, keep a small eval set, and know which named model you would switch to."
Worked example

The trial that became a dependency

A team builds a feature on a free stealth listing because it tops their quick tests.

At reveal the codename listing stops and the model moves to a paid, named ID. The team scrambles to re-benchmark and re-price. The fix is the one the gateway guides give: keep the model ID in config, keep a small eval set, and know your fallback before you ship.

Avoid

Treating a fingerprint as proof of identity, or treating a stealth listing as a permanent free tier. Both have bitten people.

Sources
Related terms