Compare input, cache reuse and estimated cost per model call within observed model workloads. Explore request costs separately.
Measured from 1617 assistant runs across 329 conversations (2026-02-12–2026-10-11). Recent runs count more: their weight halves every 30 days.
Model prices updated: 2026-10-23
3 matched attempts: 3 record model-call counts, 3 record cache usage, and 3 support the cost calculation.
Work throughput: 2.09 recorded tool calls per model call. More tool calls may represent more work; this is not an inefficiency score.
Compare the same model and workload at comparable context sizes and harness versions. Missing cache counters are excluded from cache/cost metrics. Cost uses current model API prices and excludes helper, child, service and cache-write surcharges. System/history/tool-definition attribution and completion quality are not recorded, so these figures cannot establish a harness efficiency trend.
Workload estimates below use 3 recorded attempts from Claude Sonnet 5.5. Unsupported workloads remain unmeasured.
Average primary-model cost per assistant attempt with Claude Sonnet 5.5; all internal model calls are already included
| Action | Cost | Prompt | Output | Cache | Samples |
|---|---|---|---|---|---|
Multi-page Research ~7 tool calls | 49.3¢ | 713K | 4K | 71.5% | 1~est |
Read Emails ~8 tool calls | 47.1¢ | 711K | 3K | 72.3% | 1~est |
Multi-tool Action ~8 tool calls | 22.9¢ | 117K | 4K | 20.7% | 1~est |
Usage comes from one local profile (329 conversations, 4075 messages), weighted toward recent runs. Matched model/workload groups remain separate. More tool calls and higher total usage can represent more capable execution; totals alone do not measure harness efficiency. Examples cover whole assistant attempts, not fixed recipe steps. Unknown task intent is not inferred from tools. Estimates cover primary-model API tokens only, excluding helpers, child agents, tool fees and cache-write surcharges. Unreported cache gets no discount in request estimates and is excluded from efficiency metrics. Subscription usage is repriced at API rates. You pay providers directly; Caiioo adds no token markup.