Investigate API costs with usage records, model filters and CSV export

Match the comparison scope

Open usage records (Chinese interface) and select a date range. Confirm the reporting timezone, account and billing type. Balance charges, subscription usage and per-request tasks are not interchangeable. Compare a monthly estimate with the same period rather than today’s total.

Narrow the change

Filter by API key, model, group, request type and billing mode. Determine whether volume increased or individual requests became larger. If applications share a key, consider separate named keys for future attribution.

Examine context and caching

Conversation history, expert instructions and skill text can increase input tokens. Cache reads and writes may have distinct rates. Long-context tiers, group multipliers, additional modalities and time-based rules can also explain differences from a simple estimate.

Test a specific hypothesis

Suppose an illustrative workflow originally sent 1,500 input tokens, then started attaching a 3,000-token document to every request. Total input increases even if request count does not. Verify that pattern in records before changing the workflow. This is a fictional diagnostic example, not a finding about your account.

Export for reconciliation

Use CSV export where available, retaining the date range and filters. Group rows by model or key and inspect unusually expensive requests. Visibility of error-request records depends on site settings. Remove sensitive account or business data before sharing exports.

Escalate with useful identifiers

If a difference remains unexplained, provide request identifiers, timestamps, model names, groups and a concise description. Do not send full keys or confidential prompts. Stopped generation or interrupted connections may still consume resources; check actual records and billing rules.