Models and providers

Namespaced model IDs

Use the exact provider/model form. The first path segment selects an adapter and credential. Everything after the first slash is preserved verbatim and sent to that provider, including additional slashes in hosted model IDs.

Provider prefixes are lowercase and case-sensitive. Empty path segments are invalid. A well-formed ID must also exist in the embedded catalog. If any entry in a models fallback array is unknown, LaneZero rejects the whole request before contacting a provider.

Supported providers

LaneZero's first-party model roster uses the prefixes below. The same prefixes are used in model IDs, named passthrough headers such as x-provider-key-openai, and provider-key management routes.

Host lane

The Host lane is distinct from the first-party model roster: it provides OSS models on rented GPUs through your own hosting-provider account. LaneZero remains BYOK, and billing for model execution stays with the host account you bring.

Host-lane model IDs often contain additional slashes, such as together/meta-llama/Llama-3.3-70B-Instruct-Turbo. Only the first segment is the LaneZero provider prefix.

Live model catalog

The running server exposes its accepted model IDs at GET /v1/models. The result is sorted by ID. Each item includes id, object: "model", and the provider prefix in owned_by.

The live route requires a gateway bearer key, including when opened in a browser. Use the command below when the browser request returns 401.

curl "$RACEWAY_URL/v1/models" \
  -H "Authorization: Bearer $RACEWAY_KEY"

The endpoint lists catalog membership only. It does not expose price, context-window, output-limit, or tool-support metadata.

Pricing and cost estimates

The catalog is seeded from LiteLLM's public model_prices_and_context_window.json pricing dataset, used under its MIT license. Host-lane together_ai/* and fireworks_ai/* records are mapped to LaneZero's together/* and fireworks/* namespaces without inventing missing prices. LaneZero attributes that source in the catalog loader and project notices. Z.ai and Tencent entries are locally curated from their public price tables because the embedded LiteLLM data does not supply those mappings.

When a provider reports token usage and the catalog has both input and output prices, LaneZero calculates:

LaneZero never bills for model use. Advisory spend caps compare the current month's accumulated estimates with a configured cap. They do not stop requests.