Tokens & cost#

Session Exporter turns the raw usage recorded by each tool into per-session and per-selection token accounting and cost estimates — with prompt caching handled correctly.

Token accounting#

For every session it aggregates the usage the provider actually recorded:

These come straight from the session logs — they are measured, not estimated (except for Cursor, see below).

Token basis#

Each session is tagged with how its numbers were obtained:

• recorded — Claude Code, Codex, Antigravity, Pi Agent, and Kimi Code expose provider-recorded usage. Fully accurate.
• context-snapshot — Cursor only records the size of the final context window, not cumulative spend or cache activity. A ~ marks these numbers, and cost is not computed for them (see Data & privacy).

Cost estimation#

Every recorded session is priced from its tokens × per-model rates. Caching is billed the way each provider actually bills it:

• Cache reads — the model's configured cache_read rate (falling back to 0.1× input when omitted).
• Anthropic cache writes — 1.25× for the 5-minute cache, 2× for the 1-hour cache. Session Exporter reads the 5-min / 1-hour split out of Claude's logs, so the write cost is exact, not approximated.
• Codex / OpenAI — the recorded input already includes the cached portion, so cost uses (input − cached) × input_rate + cached × cache_read_rate + output × output_rate.
• Kimi Code — inputOther is uncached input, cache reads use the published cached-input rate, and cache creation is a cache miss billed at the regular input rate.

Cost shows up as a Cost column in the table, a tile in the detail drawer, and a running total on the selection bar and the top bar.

The Stats panel#

Click Stats in the toolbar to open a breakdown of the current filter:

• Summary tiles — sessions, tokens, estimated cost, and cache-hit rate.
• By model — a table of every model, its session count, tokens, and cost, with cost bars, sorted most-expensive first.
• By date — the same, grouped by day (newest first).

Because it respects the active filter, questions like "how much did the last 3 days cost?" or "how much have I spent on gpt-5.5?" are one glance: set the date range or source filter, open Stats, read the total.

Rows using an estimated rate (a model whose exact price isn't published) are flagged with ~. Cursor sessions are excluded from cost and shown as —.

Editing prices — pricing.json#

Prices live in an editable pricing.json next to the app. Rates are USD per 1,000,000 tokens:

{
  "models": {
    "claude-opus-4-8": {
      "input": 5, "output": 25, "cache_read": 0.5,
      "cache_write_5m": 6.25, "cache_write_1h": 10
    },
    "gpt-5.5": { "input": 5.0, "output": 30.0, "cache_read": 0.5 },
    "kimi-k3": { "input": 3.0, "output": 15.0, "cache_read": 0.3 }
  },
  "aliases": { "openai": "gpt-5.5" }
}

• cache_read — the discounted cached-input rate.
• cache_write_5m / cache_write_1h — Anthropic / Pi cache-write rates. Kimi cache creation uses input automatically.
• aliases — map a raw model string (e.g. a provider fallback name) to a priced model.
• "estimated": true on a model flags it in the UI with ~.

Model keys are matched case-insensitively. A model missing from pricing.json is shown as — (not priced) rather than guessed — so a dash always means "no price configured", never "free".

Ships with rates filled in: Claude from Anthropic's official pricing; OpenAI, Z.ai (GLM), Moonshot (Kimi) from their public pricing. Change any number to match your own contract, then hit ↻ Reload prices in the Stats panel — it applies instantly, without re-scanning your data.