Summary

Some OpenRouter-hosted models use tiered pricing — the per-token rate changes once a request crosses a context-size threshold (for example, a long-context tier that kicks in above ~128k input tokens). For these models, the streaming completions API returns a base-tier cost in the final chunk, but OpenRouter's authoritative post-hoc record (/api/v1/generation?id=<responseId>) reports the tiered-tier total_cost, which can be substantially higher.

The OpenClaw gateway currently writes the streamed message.usage.cost.total directly into the assistant JSONL record. For tiered-pricing models, that streamed value is therefore a lower bound, not the billed amount — confirmed as a 4× undercount on long-context qwen/qwen3.6-plus calls.

Evidence

Six random qwen/qwen3.6-plus messages from a single deployment, comparing the cost written into the JSONL against the total_cost returned by GET https://openrouter.ai/api/v1/generation?id=<responseId>:

Below the tier threshold: parity to the penny. Above it: the streamed value is exactly 1/4 of what OpenRouter actually bills.

Full /generation response for the 410k-token call (trimmed):

{
  "tokens_prompt": 485590,
  "native_tokens_prompt": 507438,
  "native_tokens_reasoning": 0,
  "native_tokens_cached": 0,
  "total_cost": 0.6596889,
  "provider_name": "Alibaba",
  "model": "qwen/qwen3.6-plus-04-02"
}

The corresponding JSONL record's message.usage.cost.total was 0.16492710 — exactly 1/4 of the authoritative total_cost.

Over ~30 days of aggregated traffic on one deployment, the cumulative gap is roughly 50%: gateway-recorded OpenRouter spend totaled $105.05 while OpenRouter actually billed $180.19 (+$75.14). The gap is concentrated in the small number of long-context calls, not spread across all traffic. This is most visible when a primary model (e.g. GPT-5.4) exceeds its context window and the gateway falls back to a long-context OpenRouter model.

Suggested fix

Two options in the gateway's OpenRouter response handler:

1. Post-hoc fetch. After the stream completes, call GET https://openrouter.ai/api/v1/generation?id=<responseId> and overwrite message.usage.cost.total with the returned total_cost. Adds ~50–500ms per call; can be done asynchronously after the response is already flushed to the agent.
2. Record both. Keep the streamed value in message.usage.cost.total and add the authoritative value in a new field (e.g. message.usage.cost.total_billed), leaving the choice of which to trust to downstream consumers.

Option 1 is the cleaner fix for correctness. Option 2 preserves the current schema but requires downstream awareness.

Tiered pricing is not unique to qwen — any future OpenRouter-hosted model with context-size or reasoning-token tiers will hit the same failure mode, so the fix generalizes.