Console

Models

What your key can call, what each call costs you, and which model “auto” picks. 26 models across 9 providers, priced at the same rate the receipt uses.

/dashboard/models is a rate card that knows which account is reading it. It sits in Operate, directly under API Keys, because those two are the same sentence in two halves: here is your credential, and here is what it can reach. It reports nothing about your traffic — that is Usage — so it is not a Monitor surface.

Every figure is derived on the server from the registry the router and the meter both read. There is no separate catalogue behind this page and nothing here is fetched from a provider, so it cannot disagree with Models & providers or with the price on a response.

The rates are what you pay, not what we pay

The registry stores the provider's list price. Every rate on this surface is that figure times 1.25 — the published Lobstack rate, and the same arithmetic that produces x-lobstack-cost-usd on a managed request. What the provider charged us is a separate ledger column and appears nowhere in the Console. The list prices were last checked against the providers' own pages on 2026-09-08, and the page says so under its title.

A rate is never rounded into a shorter number. A per-million figure under a dollar is drawn to three decimals rather than collapsed to <$0.01, because a rate card that cannot tell $0.075 from $0.94 is not a rate card. Where the registry holds no usable price the cell is an em dash, never $0.00 — the same rule the ledger applies to cost_usd: null.

What auto serves

Five cards across the top, one per tier, each naming the complexity score it accepts and the model an "auto" request lands on there. This is the routing ladder rendered against your plan, not a second copy of it.

Tier	Score	Lead model on “auto”	Candidates in the pool
Flagship	≤ 100	`claude-opus-5`	7
Premium	≤ 90	`kimi-k3`	11
Standard	≤ 70	`claude-sonnet-5`	12
Small	≤ 40	`claude-haiku-4-5`	9
Nano	≤ 20	`gemini-3.1-flash-lite`	6

A preference, not a promiseThe router skips a provider whose key is not configured on the deployment, and that is not knowable from the registry alone. So a lead model is what the router prefers at that tier, and the model that actually served you is the one on x-lobstack-model — only that one. How the tier is chosen is on Token Intelligence.

Two labels appear on a tier card and only when they apply, because a row of five cards each saying “within your plan” is five cards saying nothing. Above your plan marks a tier your ceiling excludes. Preference points at a previous-generation key marks a tier whose first choice is a back-compat key rather than a current one — that is a catalogue fault worth seeing rather than hiding, and it is the exact failure that once left auto serving no current-generation model for six weeks.

What your plan reaches

Under the tier cards, one strip states how many of the 26 current keys this account may call. Free is capped at standard; every paid plan reaches flagship. A plan that already reaches everything says so and sells nothing.

Plan	Highest tier reachable
Free	standard
BYOK	flagship
Developer	flagship
Studio	flagship
Scale	flagship

A named model above the ceiling is refused, and auto never routes there. Plans that sell top-ups carry the same buy control the allowance panel on Usage does, because “this model is out of reach” and “buy more allowance” are the two things a reader wants at that moment.

No strip at all means the reach is unknownThe strip is absent in two cases: the allowance has not answered yet, and the account is a legacy subscriber whose ceiling lives in the routing layer rather than in the plan catalogue. Both mean the page does not know this customer's reach, and rendering every model as callable would be telling a capped account it can call a flagship. Absence is the honest state; there is no default.

The catalogue

Below the strip, the current generation grouped by tier, highest first, filtered by a provider control in the title row. Each card carries the model's label and key, its context window, its provider and — when the two differ — the provider-native id that actually goes on the wire. Underneath: the input and output rate per million tokens, and a button that copies a real request with that key already named.

What Copy curl puts on the clipboardbash

curl https://www.lobstack.ai/api/gateway/v1/chat/completions \
  -H "Authorization: Bearer $LOBSTACK_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "claude-sonnet-5",
    "messages": [{ "role": "user", "content": "Say hello." }]
  }'

The pills

Five, and every one is a fact the registry actually records with a consequence a caller can act on.

Pill	What it means
`auto picks first`	The router's first choice at this tier when it picks for you.
`in auto pool`	In the candidate list for this tier, but not the first choice.
`no temperature / top_p`	The provider rejects sampling parameters on this model with a hard 400. Send the request without them; the Gateway drops one if you send it and reports x-lobstack-dropped-params.
`BYOK only`	Not servable on a Lobstack-hosted key.
`price unconfirmed`	The list price behind this rate could not be confirmed against the provider's own pricing page on the date above.

What this surface deliberately does not show

The registry records provider, tier, context window and price. It does not record modality, tool calling, structured output or reasoning, so nothing here renders a capability badge — a badge that is wrong is worse than one that is absent, because somebody ships against it.

Absent	Why
Capability badges	Vision, tools, reasoning and structured output are not fields in the registry. Three honest facts beat seven invented ones.
Per-model rate limits	There are none in this system. The Gateway meters dollars or requests per account, never RPM or TPM per model, so a figure here would be fiction.
A “New” badge	The registry carries no release date. The only distinction it makes is current generation from back-compat, and every card here is already current.
A playground	There is no playground in this product. The copy-curl button is the honest substitute: a real call rather than a dead link.
Previous-generation keys	Back-compat is not a menu. Older keys still resolve and still price so a stored configuration keeps working; they are listed on Models & providers.

The full registry is a docs page, not a Console surfaceThis surface offers the 26 keys a caller should choose from today. The registry also holds 19 previous-generation keys, kept so an older configuration still resolves and still meters rather than being served for free, and a table of non-canonical spellings that resolve onto them. Both are on Models & providers, together with every price and context window — and as JSON from GET /api/gateway/v1/models, which returns canonical keys only.