Lexery - Cost Ledger
Tracks estimated and actual spend for wiki maintenance and automation — separate from product Brain inference costs (Portal user traffic, batch jobs, etc.). Pair with Lexery - Maintenance Runbook for when each tier runs.
Provider context: Lexery - Provider Topology. Design for automation: Lexery - Automation Architecture.
Budget envelope
- Target: ~$2/month ongoing for second-brain maintenance (Tier 1–2 dominant, Tier 3 occasional).
- Spike allowance: one-time build or large restructure months may exceed the envelope; mark rows clearly as non-recurring.
Note: Early “initial wiki” rows below may exceed the monthly target by design. Normalize future months toward the envelope via delta-first Tier 2.
Cost model (maintenance only)
| Tier | Use case | Typical tooling | Est. cost / run |
|---|---|---|---|
| Tier 1 | Git diff, file hashing, metadata, branch lists | None (local CLI) | $0 |
| Tier 2 | Commit digests, PR/issue rollups, short delta summaries | Cheap model via OpenRouter | ~$0.001–0.01 |
| Tier 3 | Wiki restructuring, merging contradictory pages, net-new canonical docs | Claude Opus-class or equivalent | ~$0.10–0.50 / substantial session |
Token discipline
- Pass diffs and summaries, not whole repos, into Tier 2–3.
- Cache file hashes in
_system/state/to skip unchanged large files (see Lexery - Maintenance Runbook).
OpenRouter Model Costs (Product Pipeline)
Хоча ця сторінка фокусується на wiki maintenance, розуміння product pipeline costs важливе для загальної cost awareness. Всі model calls проходять через OpenRouter:
Per-Model Cost Profile
| Model | Pipeline Role | Relative Cost | Типовий Use Case |
|---|---|---|---|
gpt-5.2 | Premium / legal reasoning | $$$ | ORCH decisions, U8, U10 — кожен run з повним reasoning cycle коштує найбільше |
gpt-4o-mini | Classification | $ | U2 classification, intent detection, metadata extraction — швидкі cheap calls |
gpt-5-nano | Routine | ¢ | Delta summaries, memory operations, formatting — мінімальна вартість per call |
Daily Cost Drivers
Основні джерела витрат у product pipeline (від найбільших до найменших):
- U10 Write ($$$) — найдорожчий stage: довгий structured output з legal reasoning, citations, formatting. Кожен write call використовує
gpt-5.2з extended context window - U8 Legal Reasoning ($$) — analysis pass перед writing, також на premium model
- ORCH decisions ($$) — mid-run orchestration calls; зменшені через deterministic canary
- U2 Classification ($) — cheap model, але виконується на кожному запиті
- Delta summaries (¢) — memory і context updates на
gpt-5-nano, мінімальна вартість
ORCH Cost Cutting: Deterministic Canary
Deterministic canary — ключовий механізм зменшення витрат. Для straightforward cases (single act, clear article reference) ORCH приймає рішення без дорогого gpt-5.2 call. Ефект: ~50% зменшення mid-run ORCH costs для типових запитів. Детальніше: Lexery - Retry and Recovery.
Brain Maintenance Target
Для wiki і brain-admin automation:
- Target: ~$2/month через OpenRouter
- Включає: scheduled brain-admin scans (
.github/workflows/lldbi-brain-admin.yml), import proposal generation, catalog health checks - Використовує переважно cheap models (
gpt-4o-mini,gpt-5-nano) для routine tasks - Expensive model calls тільки для substantive wiki restructuring sessions
Running total
| Date | Operation | Tier | Tokens (est.) | Est. cost (USD) | Notes |
|---|---|---|---|---|---|
| 2026-04-09 | Initial wiki build | 3 | ~200k | ~2.00 | One-time scaffold |
| 2026-04-09 | Second brain expansion | 3 | ~300k | ~3.00 | Large page batch |
Reconciliation: When invoice data exists (OpenRouter, Anthropic, Cursor), add a “verified” row or footnote — until then, treat numbers as order-of-magnitude estimates.
Monthly projection
With delta-first processing (weekly Tier 2 + one monthly Tier 3 audit):
- Expected steady state: ~$0.50–1.50/month if Tier 2 stays short and Tier 3 is bounded.
- Risk drivers: full-repo re-ingest, attaching large PDFs to prompts, repeated multi-hour Opus sessions without checkpointing.
Mitigations
- Enforce Lexery - Maintenance Runbook tier boundaries.
- Log every paid pass in this table for retroactive budgeting.
Related product costs (out of scope here)
Brain runtime spend (U2/U8/U10/ORCH, embeddings, rerankers) belongs in infra / finance dashboards — link from Lexery - Deployment and Infra when documented. This ledger is wiki ops only.
Governance
- Who can add rows? Anyone maintaining the vault; sign with initials in Notes if multi-user later.
- Rounding: estimates may be rounded to nearest $0.05 for readability; keep raw notes in monthly Lexery - Log if precision matters.
- Cap enforcement: if two consecutive weeks exceed target, switch to Tier 1 only until the next planned Tier 3.
Open questions (cost side)
See Lexery - Unknowns Queue item monthly AI spend (aggregate) — until that is answered, this ledger cannot be reconciled against company-wide burn; it only tracks documented wiki maintenance sessions.
Links
- Lexery - Maintenance Runbook
- Lexery - Automation Architecture
- Lexery - Provider Topology
- Lexery - Log — append spend notes alongside maintenance entries.
- Lexery - Unknowns Queue — item on aggregate AI spend for the product.
- Lexery - Retry and Recovery — deterministic canary cost optimization.