reports/api_usage.md

Observed API usage

model	attempts	completed	attempts_without_reported_usage	reported_prompt_tokens	reported_completion_tokens	reported_reasoning_tokens	estimated_reported_usage_cost_CNY	unknown_or_pending_reservation_CNY	estimated_cost_with_reservations_CNY	mean_completed_input	mean_completed_output
DeepSeek-V4.1-Flash	25	1	24	40	29	23	0.000312	0.5706	0.570912	40	29
GLM-5.3	2412	2400	12	1587657	804255	123028	48.0468	2.48216	50.5289	661.524	335.106

Token columns are sums of reported usage only. Missing usage is unknown, not zero; its attempt count and reserved budget are separate. Reasoning tokens are a subset of completion tokens, not an extra charge. Costs use the conservative planning rates in configs/api_budget.yaml, not a verified provider invoice. Failed/unknown usage retains an upper-bound reservation. The two pre-ledger connectivity tests are listed separately in api_budget.md. Codex threads use the existing ChatGPT account quota, which is separate from this intervention API ledger.