Skill Cost

cost

**Token-usage & USD cost report for AI agent sessions — pi, Claude Code, Codex, opencode, Hermes Agent**

[](./LICENSE) [](https://gaiaskilltree.com) [](https://github.com/BerriAI/litellm/blob/main/model_prices_and_context_window.json) [](https://www.python.org) [](#requirements) [](./.github/workflows/refresh-prices.yml)

**Prices sourced from [BerriAI/litellm](https://github.com/BerriAI/litellm) (MIT).** **[Powered by Gaia](https://gaiaskilltree.com) — an open registry of AI agent skills.**

**How much did that agent session actually cost?**

Your harness persists token counts in JSONL logs or a local session database. It often does *not* record dollars — pi records `cost: 0`, self-hosted proxies rarely price anything, and the "cost" number many CLIs surface is stale by a version. `cost` reads the persisted token counts and prices them against [LiteLLM's canonical catalog](https://github.com/BerriAI/litellm/blob/main/model_prices_and_context_window.json), producing a reproducible public-rate estimate. Subscription-included or negotiated billing may differ from this estimate.

No pip. No npm. No config. **One Python file, one JSON price sheet, prices auto-refresh every 7 days.**

---

Why this exists

I kept asking my agent, at the end of long orchestration sessions, "how much did that cost?" It would say things like `~$2` with high confidence. I believed it for a while.

Then I checked the raw token counts. One session it had estimated at $2 was closer to $8 — 4× off