llm-cost-meter report: 2026-08-29T23:59:00Z -> 2026-09-05T23:59:00Z (7d)

Fleet priced $0.63 (API-equivalent) over the last 7 days - top spender CODER at $0.40. 11 priced calls, 0.16M priced tokens. Codex: 2 measured turns, 0.02M tokens (subscription price not inferred). Provider logs: 6 reported calls, 0.06M tokens (price not inferred).

Totals: 31 calls | in 24,226 | out 27,576 | cache 191,720 | $0.6325 API-equivalent

By agent                        calls      input    output      cache   cost USD   avg ms   p95 ms
--------------------------------------------------------------------------------------------------
coder                               6      9,400    17,700     86,500     0.3971    7,000    9,800
planner                             5      3,110     5,010     40,500     0.2067    2,100    2,100
reviewer                            3      2,300     1,950     13,500     0.0287        -        -
researcher                          8      7,496     2,383     49,520     0.0000    4,211   16,100
ops                                 6      1,455       460      1,700     0.0000    1,660    3,300
unknown                             3        465        73          0     0.0000    1,003    1,350

By model                        calls      input    output      cache   cost USD   avg ms   p95 ms
--------------------------------------------------------------------------------------------------
claude-opus-4-8                     4      3,450     9,550     51,500     0.3577        -        -
claude-sonnet-4-6                   5      5,400    11,300     68,800     0.2677        -        -
claude-haiku-4-5                    2      1,200     1,050      6,700     0.0071        -        -
gpt-5.1                             5      4,587     2,195     51,220     0.0000    6,180   16,100
gpt-5.1-codex                       2      4,500     2,700     13,500     0.0000    7,000    9,800
kimi-k2                             1      3,000       400          0     0.0000    5,200    5,200
local/qwen-7b                       7      1,174       171          0     0.0000    1,000    1,350
local/qwen-14b                      5        915       210          0     0.0000    1,132    2,100

Findings (review signals, never actions):
  [attention] Local-router owner attribution: Have the router emit an explicit roster_agent field; historical rows without it stay unattributed.
  [attention] Local-router call failures: Inspect failed calls alongside correlation evidence; failed calls alone are not retry waste.
  [unavailable] Model-policy observation: Set `allowed_models` in the config to evaluate model-policy conformance.
  [attention] Context-window pressure: A review signal, not waste: inspect context composition before changing a window or compaction policy.

Cost is API-equivalent for Claude session logs only; codex, provider and router rows are unpriced (tokens/latency only).