OpenClaw Main Corrects Responses Token Usage and Anthropic Pricing Overrides

OpenClaw has merged two cost-accounting corrections on main. Clients of /v1/responses could receive incomplete or false token usage, while Anthropic runs could ignore operator-configured model pricing and fall back to catalog values.

Neither fix changes a provider invoice. They change OpenClaw’s representation of usage and estimated cost—the data operators use for budgets, chargeback, alerts and model-routing decisions. Historical reports produced by affected paths may therefore need reconciliation.

Responses usage was losing detail

PR #117533 reports several problems in the OpenAI-compatible Responses surface:

• cached input could be omitted from input_tokens;
• cache-write and reasoning detail could be absent;
• an undercounted total could be trusted; and
• a response whose only non-zero source was lastCallUsage could report zero usage.

The merged Gateway path preserves usage details and derives totals from the available authoritative components rather than accepting a contradictory zero or undercount.

Interpretation: a schema-valid usage object can still be financially false. Validation must include accounting identities and fallback precedence, not merely the presence of numeric fields.

Anthropic configuration now remains authoritative

PR #117275 fixes Anthropic cost estimation when an operator supplies model pricing in configuration. The configured rates could be ignored in favour of built-in catalog pricing, making private agreements, proxy rates or deliberately overridden test pricing invisible to OpenClaw’s estimate.

Operator configuration becoming authoritative is the correct contract, but it also increases responsibility. A typo in a custom rate can now produce a confidently wrong estimate. Keep the unit explicit, review cache read/write rates separately where supported and version pricing changes alongside the configuration that introduced them.

What operators should do

Prioritise review if you:

• expose /v1/responses to internal clients;
• allocate cost by cached input, cache writes or reasoning tokens;
• use OpenClaw totals to select or throttle models;
• configure custom Anthropic pricing; or
• reconcile OpenClaw telemetry with provider invoices.

For a safe acceptance run, use synthetic prompts and a non-production budget:

1. capture the exact model, provider route and containing revision;
2. exercise an uncached request and a cache-hit request;
3. request reasoning where the provider and model support it;
4. assert that detail fields sum consistently with reported totals;
5. set an obviously synthetic Anthropic rate in disposable configuration;
6. confirm the estimate uses that rate, then remove the override; and
7. compare OpenClaw telemetry with the provider’s own usage record without expecting rounding or billing-window semantics to be identical.

Do not put invoice exports, prompt content or credentials into public debugging artifacts. Token counts and opaque request IDs are usually enough to establish the accounting discrepancy.

Historical reconciliation

Do not silently rewrite prior dashboards. Mark the deployment boundary, calculate which routes used /v1/responses or custom Anthropic pricing and annotate the affected interval. Provider billing remains the external financial authority; OpenClaw’s corrected telemetry is an operational estimate and routing signal.

If an alert or quota was driven by an old zero or undercount, inspect whether work exceeded its intended budget. If catalog pricing overrode a lower or higher contract rate, recalculate from retained token counts rather than scaling an already rounded cost total.

Evidence and limits

The official merged PR and commit records in the repository packet establish the reported usage omissions, fallback error and pricing precedence correction. Direct GitHub page extraction was unavailable and independent search was blocked by provider credits. OpenClaw Academy did not run provider calls or compare a live invoice.

Bottom line

Usage telemetry must preserve every billable or operationally relevant component, and configured pricing must beat a generic catalog when the operator explicitly supplies it.

Adopt a containing build through staging, reconcile the affected reporting interval and keep provider invoices as the billing authority. The fix makes OpenClaw’s numbers more truthful; it does not turn estimates into invoices.