Routing and reliability
Models and the catalog
What the catalog lists, how model ids are spelled, and which context figure to budget against.
このページはまだ翻訳されていません。英語版を表示しています。
The catalog is every model you can call by slug. One rule holds it together: what GET /v1/models lists, the public catalog shows, and the gateway serves are the same set. A model is never listed if no request can reach it, and the gateway never serves a model the catalog hides.
Three ways to read the catalog
| Endpoint | Shape | Key |
|---|---|---|
GET /v1/models | OpenAI list of slugs, with data_policy per model | Yes |
GET /api/models | Full catalog: prices, context, capabilities, providers, filters | Optional |
GET /api/v1/models | OpenRouter shape, <vendor>/<slug> ids | Optional |
With your key, the listings add your organization's private models and the :free spellings you are eligible for.
/v1/modelsGET/api/modelsGET/api/models/{slug}curl "https://api.experientiallabs.ai/api/models?sort=preferred&limit=20"Spellings of one model
A model has one slug. Several spellings resolve to it at request time, so tools written for other gateways work unchanged. Usage and billing always record the slug.
| You send | Resolves to | Note |
|---|---|---|
claude-sonnet-5.5 | claude-sonnet-5.5 | The slug (dot form) |
anthropic/claude-sonnet-5.5 | claude-sonnet-5.5 | OpenRouter-style vendor prefix |
claude-sonnet-5-5 | claude-sonnet-5.5 | Anthropic's dashed id |
<slug>:free | the free lane of <slug> | Free only, never charged, see Free models |
<slug>:fast | the priority tier of <slug> | Only on models that list a fast path; priced higher |
claude-sonnet-5-5-20261001 | refused | Dated ids are not spellings: 403 model_not_granted |
client.chat.completions.create(
model="anthropic/claude-sonnet-5.5",
messages=[{"role": "user", "content": "Hi"}],
)
# response.model == "claude-sonnet-5.5"Context window vs max input tokens
context_window is the total: prompt plus the model's maximum output. Providers count the whole maximum output inside the window, whatever max_tokens you ask for. Size your prompt budget, and any agent compaction setting, to max_input_tokens.
| Field | gpt-6-luna |
|---|---|
context_window | 1,050,000 |
max_output_tokens | 128,000 |
max_input_tokens | 922,000 (window minus max output) |
{
"model": {
"slug": "gpt-6-luna",
"context_window": 1050000,
"max_output_tokens": 128000,
"max_input_tokens": 922000
}
}Pinning a coding agent's context limit to context_window lets a session grow past what the provider accepts, and the session can no longer compact.
What a catalog entry tells you
| Field | Meaning |
|---|---|
providers[] | Each rung: provider, prices in nano-USD per million tokens, capabilities |
billing_source | host_managed (credits) or customer_managed (your key) |
pricing_source | Where the price comes from. An estimate is never billed |
retention | zdr_all_rungs, zdr_enforceable or not_zdr, see Data controls |
| uptime and latency | Measured from real traffic once a route has enough volume |
Filter GET /api/models with modality, provider, min_context, supports, retention, free=true and sort. Some makers (OpenAI, Anthropic, Gemini) serve only requests from their supported countries; listing a model does not establish regional eligibility.
{
"provider": "anthropic",
"billing_source": "host_managed",
"input_nano_usd_per_million": 4000000000,
"cached_input_nano_usd_per_million": 200000000,
"output_nano_usd_per_million": 20000000000,
"pricing_source": "provider-docs"
}1,000,000,000 nano-USD is $1, so this rung is $4 input and $20 output per million tokens.