Bottom line up front
When a gateway lists a model under an alias with no publicly identified weights behind it, the registry gives you an identifier, a price, and sometimes a set of reasoning-effort settings — and nothing else. You cannot compare it on context length, licence, modalities, or benchmarks, because none of those are established. What you can do is measure the endpoint's behaviour on your own workload, record the conditions carefully, and treat the result as evidence about that endpoint on that date rather than about a model you could name.
Separate what the registry states from what you would like to know
A gateway entry is deliberately small: a model identifier, a display name, sometimes a release date, a status, a list of reasoning-effort settings, and a price per million tokens. For a zero-priced entry that is the entire public record. It tells you the endpoint exists and what it costs today, and it tells you nothing about the model behind it.
Write down what you actually know before you start evaluating, because the temptation is to fill the gaps from a name. A codename is not a specification. Where no developer publishes a model card, the honest position is that the model is unidentified, and any comparison on context, licence, or capability is unavailable rather than pending.
- List the facts the registry gives you, and stop there.
- Mark every other property as unknown rather than inferred.
- Record the exact identifier and the date you read the entry.
Why an unidentified model changes the decision
Identification is not pedantry; it changes three practical things. Licence determines whether you can self-host or redistribute. Context and modality limits determine what your workflow can ask for. Provenance determines whether you can verify a security or data-handling claim with the developer rather than the reseller.
A reseller's terms govern the endpoint, but they do not tell you anything about the weights. If any of those three matters to your use case, an alias is a gap you either accept knowingly or route around by choosing an identified model. For disposable or low-stakes work, accepting it knowingly is a reasonable decision — the point is that it is a decision.
- Ask whether licence, context, or provenance matters for this use.
- If yes, prefer an identified model over an alias.
- If no, record that you accepted an unidentified model knowingly.
A behavioural trial that still produces a decision trail
You can measure an endpoint without knowing what is behind it. Pick one task with a clear output and a review point, run it across a small set of candidates including at least one identified model as a reference, and record the conditions: date, endpoint, model identifier, reasoning-effort setting, prompt, and the failures as prominently as the successes.
The result is evidence about that endpoint on that date, and nothing more. That is still useful — it is how you decide whether the endpoint is good enough for the job in front of you. What it cannot do is support a claim about the model in general, a comparison table entry, or a recommendation to someone else whose workload differs.
- Include an identified model in the trial as a reference point.
- Record the date, identifier, and reasoning-effort setting with every result.
- Keep failures in the record alongside successes.
The same model appears under different prices on different gateways
Free model lists are not one list. Several gateways publish their own zero-cost entries, some of which are the same underlying model priced differently, and some of which are aliases with no public identity. An entry that is free on one gateway may be paid on another, and a free tier can be withdrawn or repriced without a deprecation notice.
The practical consequence is that a free model list is a snapshot, not a catalogue. Record where you read each entry and when, and treat any entry as provisional until you have used it. For anything that must keep running, a paid fallback or a self-hosted path on an identified, openly licensed model is the design that survives a gateway changing its mind.
- Record the gateway and the date for every free model you rely on.
- Check the same model across gateways before assuming one price or one identity.
- Design a paid or self-hosted fallback before the dependency.