Use-case preset
Log analysis cost calculator
Anomaly detection and clustering over server logs; best-effort.
Log analysis jobs paste a 10k–15k token window of structured log lines (JSON, syslog, or mixed) into the prompt and ask the model to identify anomalies, cluster error patterns, or summarise failure modes. Input tokens dominate at 90%; output is a short structured report, typically 200–500 tokens. The 16k window fits a meaningful slice of a production incident without chunking.
Best-effort latency reflects that this runs as a scheduled job or post-incident analysis, not in the hot path. Cache hit rate is low (0–15%): log content changes every run, and only the system prompt and schema description are reusable. Cost lever: pre-filter logs with a regex or embedding-based deduplication step before sending to the LLM — reducing a 10k-token window to 6k tokens cuts cost 40% with minimal quality loss. Models with strong context utilisation at long lengths matter more than raw capability score here.