RELIABILITY TRACKER · P2 OPEN
Reasoning Tokens Omitted from Usage-Cost Estimates · Issue #68081
A source-linked review of a reported Hermes Agent usage-accounting defect where reasoning tokens are captured during normalization but omitted from estimated per-call and downstream cost calculations.
What the report describes
The report says normalize_usage() captures reasoning tokens, while estimate_usage_cost() prices only input, output, and cache tokens. The issue documents DeepSeek and OpenRouter reproductions in which per-call estimates materially undercounted spending and Langfuse inherited the same estimate. Exact impact depends on provider pricing, workload, and the accuracy of the reported token fields.
Scope note: This page summarizes a linked open issue, not an independent billing determination. Provider invoices and current pricing documentation remain the authoritative source for actual charges.
Verification matrix
| Signal | What to inspect | Safe review response |
|---|---|---|
| Usage capture | Whether normalized usage records include reasoning-token fields and preserve provider units | Retain raw usage and provider identifiers for reconciliation. |
| Pricing path | Whether cost estimation includes every billable token category for the selected provider | Map pricing logic explicitly to the provider usage schema and pricing version. |
| Per-call reconciliation | Compare agent estimate with provider-visible billable usage for controlled calls | Report any difference with the request date, model, price source, and methodology. |
| Downstream telemetry | Whether observability or cost dashboards reuse the same incomplete estimate | Trace derived metrics so an upstream accounting omission is not mistaken for independent confirmation. |
| Budget guardrails | Whether alerts or project budgets rely on undercounted values | Use conservative thresholds or provider reconciliation until the estimate is corrected. |
Reproduction and monitoring protocol
- Choose controlled requests that expose input, output, cache, and reasoning-token fields where supported.
- Record the model, provider price reference, timestamp, and normalized usage payload.
- Calculate a transparent expected estimate from all billable categories and compare it with the displayed estimate.
- Trace whether downstream cost reporting consumes the same estimate.
- Re-test after an upstream fix before relying on project-level budget or attribution figures.
Source evidence
NousResearch/hermes-agent · Issue #68081
Verify current issue state, labels, affected versions, and remediation before making a deployment, security, or budgeting decision.