See when token usage is missing
Valeriy MeleshkinWe're retiring automatic tokenization so missing usage is visible instead of hidden behind misleading cost estimates.
You can now identify missing token usage directly in generation headers and follow links to fix your instrumentation.
- Check that a new integration submits provider usage.
- Find out whether a generation's missing cost comes from missing usage or a missing model price.
We're retiring automatic tokenization in Langfuse: estimating token counts from a generation's input and output when no usage was submitted.
With thinking models and cached token pricing, these estimates can produce incorrect costs and mask broken instrumentation. We think it is better to be explicit when instrumentation needs fixing.
Langfuse still calculates costs from submitted usage and model prices, and you can still submit costs directly. If your integration already sends usage or costs, no change is needed.
If you relied on automatic tokenization, update your instrumentation to send provider usage or costs. The change applies to new writes and reprocessing; historical observations are not bulk-rewritten.
See the troubleshooting guide for help with missing usage or cost.