Case Calendar costs money in two independent ways, and both apply whether or not you enable AI case summaries:

  1. LLM API calls — the extractor runs a small/fast model on every relevant docket entry (always, even with summaries off); case summaries add a higher-tier model on top (opt-in).
  2. CourtListener API quota — every sync makes REST calls against CourtListener, whose free tier is rate-limited and whose higher tiers come from a paid Free Law Project membership.

This page covers both — with real measured numbers — and how to measure your own spend.

← Back to docs

LLM cost

Two tracks run on different model tiers. The extractor (and its end-of-sync verify pass) handle high-volume structured classification — date, key, significance — so they default to the cheap small/fast tier and run whether or not summaries are enabled. Summaries are the opt-in synthesis track on a higher tier.

Track Runs when Per-provider model (Anthropic / Gemini / OpenAI)
Extraction + verify Always Claude Haiku 4.5 / Gemini 3.1 Flash Lite / GPT-5.4-nano
Summaries Opt-in (case_summaries.enabled) Claude Sonnet 4.6 / Gemini 2.5 Pro / GPT-5.4

The bold model in each row is the zero-config default for that track. As of 0.13.0 the default is a split: Gemini 3.1 Flash Lite for extraction and Claude Sonnet 4.6 for summaries (the rest are what you get if you pin that provider). Zero-config auto-detection (you set the API keys, no LLM_* env vars) wires up exactly this assignment — the numbers below let you decide whether a different provider fits your caseload. See Architecture → Why the default is a split for why.

Cost scales with your caseload — the number of dockets, how many entries each has, and how long the documents are — so a single universal figure would mislead. Instead, here are real measured numbers from a full from-scratch backfill of the maintainer’s own calendar (28 cases / 34 logical dockets, measured 2026-06-07 on 0.16.0), broken out by provider and track. “Backfill” means processing every historical docket entry plus generating every summary — the one-time cost of onboarding a caseload:

Provider (extraction / summary model) Extraction Verify Summary Backfill total
OpenAI (GPT-5.4-nano / GPT-5.4) $1.17 $0.12 $1.58 $2.87
Gemini (3.1 Flash Lite / 2.5 Pro) $1.87 $0.16 $1.16 $3.19
Default (Gemini 3.1 Flash Lite / Claude Sonnet 4.6) $1.87 $0.16 $2.51 $4.54
OpenAI (GPT-5.4-mini / GPT-5.4) $3.94 $0.44 $1.78 $6.15
Anthropic (Haiku 4.5 / Sonnet 4.6) $6.99 $0.74 $2.51 $10.24

The Default row is what zero-config gives you — it pairs the two tracks’ own defaults, Gemini extraction with Anthropic summaries; the single-provider rows run one provider on both tracks for comparison. Rows are ordered cheapest to priciest backfill.

The Extraction and Verify columns are what you pay with summaries off; the Summary column is the opt-in add-on. Extraction runs roughly $0.04–0.25 per case (it scales with entry count, so a busy docket costs more — the high end is Anthropic’s $6.99 over 28 cases), and summaries roughly $0.03–0.07 per docket. After the backfill, existing summaries are reused unless a docket gets a new primary document or disposition, so ongoing spend is pennies a week; the verify track (one focused call per non-terminal hearing/deadline + the new source-entry-aware context as of 0.11.0) is what runs on every sync, and even that stayed under $1.00 across the whole caseload on the priciest provider.

These figures already include the 0.16.0 change where the extractor reads the PDF of every order (an order’s operative dates often live only in a schedule table the one-line docket text doesn’t echo) and stops fetching transcript bodies (testimony with no forward-looking scheduling). Reading order PDFs is what nudged extraction up versus the prior 2026-05-31 measurement (Gemini $1.82 → $1.87, Anthropic $6.72 → $6.99 over the same 28 cases); it’s a one-time backfill cost — fingerprint dedup means steady-state spend is unchanged.

These are estimates, and they reflect one specific caseload on one date — don’t take them on faith; measure your own with the llm-tokens lines.

CourtListener API limits

The LLM dollars above are only half the picture — CourtListener’s REST API is rate-limited, and higher limits come from a Free Law Project membership (which also funds the project). The tiers, with the per-minute / hour / day request ceilings (free.law/membership, verified 2026-05-26):

Tier Price Requests (min / hour / day)
Free (non-member) — 5 / 50 / 125
Tier 1 $10/mo · $100/yr 10 / 75 / 300
Tier 2 $25/mo · $250/yr 15 / 150 / 600
Tier 3 $50/mo · $500/yr 20 / 250 / 1,000
Tier 4 $100/mo · $1,000/yr 25 / 300 / 1,400

Membership API access is intended for small firms, small government / media outlets, academics, and pre-revenue / pre-funding organizations — not large or funded ones.

Against the backfill above: it made 37 API calls total, so its hourly and daily totals (37 / 37) sit inside even the free tier (50 / 125). The catch is the per-minute burst — the backfill peaked at 14 requests in one minute (cold-docket lookups firing back-to-back), which exceeds the free (5/min) and Tier 1 (10/min) caps but stays inside Tier 2 (15/min). That isn’t data loss: the client honors Retry-After and backs off automatically (just slower). A from-scratch backfill is smoothest on Tier 2 (15/min); a lower tier still completes, just with back-off pauses. Steady-state polling is a handful of requests per sync, far below every tier. The per-run courtlistener-requests log line reports your actual peak min / hour / day so you can size your membership.

Measuring real token usage and estimated cost

Every LLM call — extraction and summaries alike — logs its token counts at INFO, with a rough dollar estimate alongside. The lines are prefixed llm-tokens:

llm-tokens call purpose=summary provider=anthropic model=claude-sonnet-4-6 docket=12345 in=48210 out=312 cached=7858 cache_write=0 cost_est=$0.1282
llm-tokens call purpose=extract provider=gemini model=gemini-3.1-flash-lite docket=12345 in=1820 out=64 cached=0 cache_write=0 cost_est=$0.0007

purpose distinguishes the cheap extractor calls (extract, verify_hearing, verify_deadline, dedupe_hearings) from the higher-tier summary calls, and docket is the CourtListener docket id.

At the end of a sync or summarize run, a per-docket subtotal, a per-model subtotal, and a grand total are logged so you don’t have to add them up yourself:

llm-tokens docket=12345 calls=9 in=63140 out=540 cached=58000 cache_write=2100 cost_est=$0.0241
llm-tokens model=gemini-3.1-flash-lite calls=32 in=58400 out=1700 cached=0 cache_write=0 cost_est=$0.0061
llm-tokens model=claude-sonnet-4-6 calls=5 in=152480 out=310 cached=138400 cache_write=2100 cost_est=$0.0492
llm-tokens sync TOTAL calls=37 dockets=4 models=2 in=210880 out=2010 cached=138400 cache_write=2100 cost_est=$0.0553

The per-model lines are what let you read the cheap extractor track’s spend (here gemini-3.1-flash-lite) apart from the higher-tier summary track’s (claude-sonnet-4-6) — the two run on different models, so the by-model split is the by-track split. The extraction LLM: / summary LLM: lines logged at the start of the run name which model is which.

The token counts are normalized so in always means the same thing across providers (Anthropic reports cache reads/writes separately from its input count; we fold them in), and the cost estimate prices each slice — uncached input, cache reads, cache writes, output — at its own rate, which matters because the system prompt is cached on almost every call.

cost_est is an estimate, not a bill. It comes from a small static price table (case_calendar/costs.py) hand-copied from each provider’s pricing page and dated with PRICES_VERIFIED — the default Anthropic, Gemini, and OpenAI (5.4 / 5.5 family) models are priced at their standard tier. It does not model batch discounts, long-context (>200k) tiers, or data-residency multipliers. When a model isn’t in the table — a legacy model, or anything you set via LLM_MODEL that hasn’t been added — the call logs cost_est=? and the run TOTAL notes how many calls had no price entry, so a partial estimate is obvious rather than silently low. Add a model’s rates to that file to price it. Nothing is persisted — to track spend over time, sum the TOTAL lines from your run logs.

To force a regeneration of every summary after a model upgrade or prompt change (useful when you want to re-measure the summary track’s cost):

uv run case-calendar summarize --force
# or, bundled into a polling sync to share the CourtListener session:
uv run case-calendar sync --force-summaries

Next steps