Finley

Engineer · AI usage, harness practices and where spend could have gone further.

Demonstration report. This engineer and their usage are synthetic.

This page is for the engineer and their manager to find cheaper defaults and harness fixes. It is never used to rank people.

Period

Summary

FY2027 to date

AI spend

$2,093

126 sessions · $17 each

Savings opportunity

$383

18% of spend, at list price

Practices on target

3 of 7

Focus: rtk coverage, Knowledge MCP use

Cache hit rate

82%

Team 82% · org 82%

FY2027 to date · estimated at list price

$383 of $2,093 spend

Largest saving first

Improve prompt caching$232
Cache hit rate 82% against an 85% target.
Keep stable context at the front of prompts and avoid switching models mid-session, so 85% of input reads from cache.
Keep context lean$63
9% of input tokens were past 100K of context.
Compact or /clear between tasks instead of carrying more than 100K tokens of context.
Right-size models$59
40% of routine spend ran on Opus: code research $67 of $186, documentation $36 of $77, ops $41 of $95.
Run docs, decks, ops and read-only research on Sonnet instead of Opus.
Route commands through rtk$28
3% of 1,044 Bash calls ran through rtk; 2.4M tokens of command output went into context.
rtk compacts test, build, git and search output before it reaches the model (about 70% smaller).

Compared with Video Rendering and the organization

Practice	Finley	Team	Org	Target	Recommendation
rtk coverage	3%Off target	2%	56%	≥ 80% of Bash calls	Install the rtk hook so test, build, git and search commands return compact output.
Knowledge MCP use 4 findings saved back	30%Needs attention	28%	30%	≥ 40% of sessions	Search the knowledge MCP before re-researching a codebase, and save findings back with remember.
Cache hit rate	82%Needs attention	82%	82%	≥ 85% of input	Keep CLAUDE.md and tool lists stable, avoid mid-session model switches, and resume sessions instead of restarting.
Context past 100K	9%On target	17%	10%	≤ 10% of input	On target
Opus on routine work	40%Needs attention	43%	44%	≤ 25% of routine spend	Default routine sessions to Sonnet and keep Opus for harness, design and hard debugging work.
Sessions using subagents	53%On target	48%	50%	≥ 30% of sessions	On target
Tool error rate	3.5%On target	3.0%	3.0%	≤ 4% of tool calls	On target

Is the engineer moving toward the targets?

Month	rtk	Knowledge	Cache	Context	Model fit	Subagents	Savings opportunity
Feb 2026	–	–	58%Off target	10%On target	–	19%Needs attention	$20
Mar 2026	–	–	65%Off target	10%Needs attention	62%Off target	28%Needs attention	$25
Apr 2026	–	–	68%Off target	7%On target	62%Off target	28%Needs attention	$23
May 2026	4%Off target	–	69%Off target	8%On target	55%Off target	39%On target	$36
Jun 2026	1%Off target	8%Off target	71%Needs attention	11%Needs attention	53%Off target	48%On target	$153
Jul 2026	7%Off target	22%Needs attention	81%Needs attention	8%On target	40%Needs attention	47%On target	$32
Aug 2026	4%Off target	37%Needs attention	83%Needs attention	7%On target	38%Needs attention	56%On target	$45
Sep 2026	3%Off target	38%Needs attention	87%On target	10%On target	30%Needs attention	67%On target	$37
Oct 2026	0%Off target	36%Needs attention	87%On target	8%On target	27%Needs attention	53%On target	$14

FY2027 to date

By model tier, by month

Routine types (documentation, presentation, ops, code research) rarely need Opus

Cache reads are billed at a fraction of uncached input; a high share is what good caching looks like

Model	Requests	Uncached input	Cache reads	Cache writes	Output	Spend
Opus 5.5	2,825	11.7M	147.4M	10.2M	3.7M	$202
Opus 5	4,436	23.5M	221M	15.6M	6.9M	$499
Opus 4.8	3,015	35.6M	136.5M	11M	5M	$439
Opus 4.6	1,590	24.2M	62.2M	5.5M	2.6M	$251
Opus 4.5	60	1.3M	2.1M	219K	97K	$11
Sonnet 5.5	2,358	10.5M	132.1M	9.1M	3.9M	$96
Sonnet 5	5,223	29.7M	275.6M	19.5M	7.9M	$364
Sonnet 4.6	1,666	22.8M	70.1M	5.9M	2.6M	$151
Sonnet 4.5	31	783K	1.3M	130K	49K	$4
Haiku 5.5	3,054	14.1M	179.7M	12.4M	5.6M	$8
Haiku 4.5	2,389	15.8M	112.8M	8.2M	3.5M	$55
GPT Realtime (meetings)	133	1.3M	6.8M	518K	212K	$13
Total	26,780	191.4M	1347.7M	98.2M	42.1M	$2,093

Spend and tokens come from LiteLLM gateway logs and imported agent session history, priced at API list rates.
Savings are estimates: routine Opus work repriced at the Sonnet available at the time; uncached input moved to cache reads up to an 85% hit rate; rtk removing 70% of unfiltered command output, which stays in context for about half the session's remaining turns; and 50% of input past 100K context avoided by compacting. Levers overlap, so the total is an upper bound.
rtk and the knowledge MCP are scored from their rollout (rtk May 2026, knowledge MCP Jun 2026).
Generated Oct 20, 2026, 7:46 AM.