llm-spend

Price what your llm history actually cost.

llm already logs every response's token usage to its SQLite database (logs.db) — input tokens, output tokens, provider token details — but it never turns those numbers into dollars. This plugin adds one command that does:

llm install llm-spend-report
llm spend

| Model | Calls | Input tokens | Output tokens | Est. cost |
|---|---:|---:|---:|---:|
| gpt-5 | 2 | 3,000,000 | 500,000 | $7.6250 |
| claude-sonnet-4-5 | 1 | 1,000,000 | 200,000 | $4.6500 |
| local-llama | 1 | 5,000 | 800 | — |
| **Total** | **4** | **4,005,000** | **700,800** | **$12.2750** |

_Total covers priced models only; no price is known for: local-llama (add one with --prices)._
_1 logged call(s) recorded no token usage and are excluded._

Usage

llm spend                          # all logged calls, grouped by model
llm spend --by day                 # grouped by day instead
llm spend --days 30                # only the last 30 days
llm spend --model gpt-5            # one model only
llm spend --json                   # machine-readable report
llm spend --prices my-prices.json  # override/extend the price table
llm spend --database /path/to/logs.db   # a non-default database

The database is opened read-only — the report never writes to, migrates, or locks your live logs.db.

How it works

• Both storage eras. Current llm versions log usage to the turns table; older versions used responses. A call is written to exactly one of them, so both are read and merged.
• Cached tokens are priced at the cached rate when the provider's token details record them (OpenAI- and Anthropic-style detail shapes are recognized) and the price table has a cached-input price. Anything unrecognized is priced at the full input rate — never double-discounted.
• Unknown models are never silently $0. A model with no price in the table is listed with its token counts and a — cost, and named in the notes. Calls that recorded no token usage at all are counted and excluded.
• Model names come from the logged resolved_model when present, so an alias you called is priced as the model that actually ran.

Prices

The bundled table is example data as of 2026-10-30 (GPT-5 family, GPT-4o family, Claude Sonnet/Haiku 4.5, Gemini 2.5 Pro/Flash) — providers change prices often, so verify before relying on totals, or supply your own file (same shape as below; entries override or extend by name):

{"models": {"my-model": {"input_per_m": 0.40, "output_per_m": 1.60,
                         "cached_input_per_m": 0.04}}}

Prices are USD per 1M tokens; cached_input_per_m is optional.

Limitations

• Estimates, not invoices. The log records what the provider reported; free tiers, credits, batch discounts, and per-request fees are not in the log and cannot be priced from it.
• Embeddings usage is not priced in this version.
• Alias-only model ids that never resolved to a priced model stay unpriced until you add them via --prices.

Development

python -m venv .venv && source .venv/bin/activate
pip install -e ".[test]"
python -m pytest        # 11 tests, fully offline

Requires Python ≥ 3.10 and llm itself. Zero additional runtime dependencies.

License

MIT — see LICENSE. Copyright (c) 2026 Gang Ji.

Not affiliated with the llm project; llm is by Simon Willison.