Track LLM token usage and cost accurately in Langfuse, including custom model prices for unknown models and manually supplied usage

Documented steps

1. By default Langfuse infers cost from the observation's `model` field plus usage counts, using built-in tokenizers and price tables for common providers (OpenAI, Anthropic, Google). No action is needed when your model name matches a known pattern.
2. For an unknown, custom, or self-hosted model, define it in Project Settings > Models in the UI, or via the API: `POST /api/public/models` with a regex match pattern such as `(?i)^(my-custom-model)$`, a unit (for example TOKENS), and a prices map like `{"input": 0.03, "output": 0.06}`. Manage definitions with GET/POST/DELETE /api/public/models using the same Basic auth scheme as the Public API.
3. Optionally add tokenizer config for count estimation (OpenAI-style models use tokenizerModel, tokensPerName, tokensPerMessage) and pricing tiers when rates vary by context length, e.g. a higher price above 200K input tokens.
4. To bypass inference and supply exact usage yourself, set `usage_details` as non-overlapping buckets where every token is counted in exactly one key, for example input, output, cache_read_input_tokens, reasoning_tokens.
5. Optionally set `cost_details` with USD values per bucket to override the computed cost entirely.
6. Python: `langfuse.update_current_generation(usage_details={"input": 10, "output": 5}, cost_details={"input": 0.01, "output": 0.02})`.
7. JS/TS: `generation.update({ usageDetails: { input: 10, output: 5 }, costDetails: { input: 0.01, output: 0.02 } })`.
8. Open the generation detail view in the UI and confirm the Usage & Cost panel reflects your values rather than showing $0 or unknown.
9. Verified against current official docs on 2026-07-29: https://langfuse.com/docs/observability/features/token-and-cost-tracking