Bottom line up front
OpenCode Zen listed ten models at $0 per million input and output tokens when this article was written. Four of them are publicly identifiable open-weight models with documented specifications; three are vendor models served under opaque aliases; the rest sit in between. Free access is a gateway subsidy with rate limits and no continuity guarantee, so it is a research and prototyping budget rather than something to build a production dependency on without a paid fallback.
What the registry actually records
A model registry entry is a small set of facts: a model identifier, a display name, a release date, a status, a list of reasoning-effort settings the gateway accepts, and a price per million input and output tokens. For the ten models listed here, the price is zero on both sides. That is the whole of what the registry establishes.
A registry entry does not establish context length, modalities, benchmark performance, rate limits, uptime, or which underlying weights are being served. Those come from the model developer's own documentation, and for several of these entries no such documentation exists publicly. Treat the registry as an address book, not a specification.
- Record the exact model identifier, not just the display name.
- Record the date you read the entry; gateway lists change without deprecation notices.
- Separate what the registry states from what the developer documents.
Tell an identified model from a gateway alias
Four of these entries map to models whose developers publish specifications: DeepSeek V4.1 Flash, NVIDIA Nemotron 3.5 Lightning, StepFun's Step 5 Preview, and Xiaomi's MiMo-V2.6-Flash. For those, the useful facts — parameter counts, context windows, licences, and output limits — come from the developer's own model card and documentation, and the gateway is simply one way to reach them.
Three entries are gateway aliases with no publicly identified weights behind them. In those cases the honest review states that the model is not publicly identified and cannot be compared on published specifications. Naming a vendor anyway would be a guess, and evaluating a model you cannot identify is not possible beyond measuring the endpoint's behaviour yourself.
- Check whether the developer publishes a model card before comparing specifications.
- Where no card exists, treat the alias as unidentified rather than inferring a vendor.
- Prefer an entry whose weights you can also run yourself, if self-hosting matters to you.
Read free access as a subsidy with conditions
A zero price on a gateway means the operator is absorbing the inference cost. That is usually a deliberate subsidy to attract usage, and it carries the usual consequences: rate limits, narrower concurrency, no service-level commitment, and the possibility of retirement or repricing. A free endpoint is a research budget, not an entitlement.
The practical consequence is architectural. If a free model is embedded in a workflow, that workflow needs a paid fallback and a way to detect when the free endpoint stops responding. Where a model's weights are openly licensed, there is a second route: run it yourself, which converts a subsidy you do not control into a cost you can forecast.
- Find the rate limit and concurrency limit before designing around a free model.
- Record what happens to your workflow when the endpoint is unavailable.
- Keep a paid fallback, or a self-hosted path, for anything that must keep running.
Choose a candidate the way you would choose any dependency
The gateway is a discovery surface. It tells you what exists and what it costs today; it cannot tell you whether a model fits your workload. That part is a bounded trial: pick one task with a clear output, run the same task across two or three candidates, record the conditions, and keep the failures visible alongside the successes.
Reasoning-effort settings are worth treating as part of the experiment rather than a fixed dial. Several of these entries accept low, medium, high, and sometimes higher settings, which changes cost and latency as well as depth of reasoning. Recording the effort level alongside each result is what makes a later comparison meaningful.