Cost is an estimate from list prices (tune MODEL_PRICING in ClaudeClient.ts). Local Ollama calls are $0. When AI runs on Ollama, spend stays $0 while token volume still shows here.
Loading…
How to read this page before concluding anything
Every user is on Claude Haiku since 2026-10-02: the Node gateway and the Python coach service both read ANTHROPIC_MODEL (claude-haiku-4-5), inside per-plan monthly counts and a spend cap of an eighth of each plan’s price ($25 a month for admins). When Anthropic fails, the local model answers instead, and the user still gets a reply.
So $0.00 against Anthropic means nothing ran, or every call failed. A failed call writes no ledger row, so this page cannot tell the two apart: the Claude key button at the top of every admin page can (it asks Anthropic live), and Who is on Claude above says whether Claude is switched off for everyone. A zero is only good news once you have seen a call succeed.
Comparing across dates: rows written before 2026-09-15 may mix Opus 5 ($5 / $25 per million tokens) with Haiku 4.5 ($1 / $5); check the model column before comparing cost per call.
The monthly counts each plan includes (AI_COACH_MONTHLY_CAPS, AI_CHAT_MONTHLY_CAPS) were sized on what one AI Coach run and one My Second chat message cost on Claude Haiku. They were measured on 24 runs and 5 messages, too few to trust. This measures both again, from the ledger, since everyone moved to Claude, and says when there are enough of each to re-size on. tests/haikuforall.test.ts recomputes the margins from the sizing figures.
Loading…
Cost per call is what makes this extrapolable: what one answer costs is what says what ten thousand would. Ollama runs locally and is $0 at any volume.
| Provider · model | Calls | Tokens in/out | Cost | Per call |
|---|
| Loading… |
| Feature | Calls | Cost |
|---|
| Loading… |
ChessMoves AI