Liminalis+ subscriptions
You are buying model time. You should be able to see where it went.
Every model call in Liminalis is metered against real provider cost and written to a durable ledger. Not a monthly allowance of mystery units — an auditable record of what ran, in which role, on which model, against which document.
Four properties that matter more than the price.
Cost at provider price, plus a stated markup
Usage is priced from the actual provider cost of each call — input tokens, output tokens, cached tokens at their reduced rate — with a markup applied on top rather than a rate invented from nowhere. Cached input bills at roughly a tenth, and because the prompt is deliberately structured so the stable half stays cached, that reduction shows up in your ledger rather than in ours.
A durable, auditable ledger
Every charge is a ledger entry, attributable to a role and a document. That's the same mechanism the legislation corpus uses to report what a batch actually cost: spend is read back by document id rather than estimated after the fact. If you can't reconcile a bill, the system is broken.
You choose what runs on what
Chat is one job. Extraction, summarizing, scaffolding, critique and narration support are others, and each is independently assignable to a model. Put the mechanical work on a cheap model and keep the expensive one for replies. Entity extraction — the most expensive background pass — can be switched off entirely per project.
Running out is a soft stop
Exhausting credits refuses new work only. Work already in flight finishes. Documents mid-analysis stay pending and resume from where they were on the next tick after a top-up — no backfill, no catch-up step, nothing lost. An empty balance costs you time, never progress.
A project-level pause stops the autonomous workers without touching chat or anything you explicitly asked for — the line is autonomous versus requested. It exists because an effectively unbounded corpus needs a brake that doesn't also break the thing you're doing right now.
Shapes, not final numbers.
Billing supports both one-off credit packs and recurring subscriptions, through a self-serve portal. The plan structure below reflects what the system supports today.
Pay as you go
Buy credits, use them, top up when they run out. No recurring charge and no expiry treadmill. Suits occasional or bursty work — importing a manuscript, prepping a campaign, running one corpus analysis.
- All lenses
- Full model choice, including the role matrix
- Project sharing and roles
- PLACEHOLDER: pack sizes and prices
Ongoing work
A recurring allowance for continuous projects — a manuscript you're actively drafting, a campaign in weekly play, a corpus you keep adding to. Same metering underneath; the subscription just funds the balance.
- Everything in the credit pack
- Recurring credit allowance
- Self-serve billing portal
- PLACEHOLDER: tiers, allowances and prices
API & MCP access
Scoped API keys for the agent API and the MCP server, metered through the same ledger. Every call reports its own cost, so an agent's spend is attributable rather than aggregate.
- Scoped keys, write off by default
- Webhooks with per-key rate limits
- Per-call cost reporting
- PLACEHOLDER: access and rate terms
The expensive half is always a separate, deliberate action.
- Free
- Browsing, editing, search over what's already indexed, opening the insight charts, previewing an ingest batch, scanning a corpus manifest, and every deterministic analysis — the reverse index, the escalation matrix, the structural half of validation.
- Cheap
- Creating and embedding documents. This is what makes material searchable and citable, and it is deliberately decoupled from analysis so a large corpus can be made useful before it is made expensive.
- The real spend
- Background analysis — entity and fact extraction, summarizing, motif tagging — plus generation: builder stages, chronicles, validation's model pass, and chat itself at the higher thinking levels.
- How it's controlled
- Analysis is released explicitly, per batch, with an estimate first. Thinking levels are visible and selectable. The role matrix puts mechanical work on cheap models. Background processing can be paused per project. And a router-level gate stops a background fan-out from starving — or outspending — the reply you're waiting on.
Questions about cost at your corpus size?
The honest answer usually depends on how much of the corpus needs full analysis versus just needing to be searchable.