Sage

Lead 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

$3,057

205 sessions · $15 each

Savings opportunity

$357

12% of spend, at list price

Practices on target

6 of 7

Focus: rtk coverage

Cache hit rate

88%

Team 84% · org 82%

FY2027 to date · estimated at list price

$357 of $3,057 spend

Largest saving first

Improve prompt caching$164
Cache hit rate 88% 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.
Right-size models$134
21% of routine spend ran on Opus: code research $113 of $552, documentation $35 of $170, ops $176 of $845, presentation $5 of $33.
Run docs, decks, ops and read-only research on Sonnet instead of Opus.
Keep context lean$48
4% of input tokens were past 100K of context.
Compact or /clear between tasks instead of carrying more than 100K tokens of context.
Route commands through rtk$10
77% of 2,105 Bash calls ran through rtk; 2.8M 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 Infrastructure and the organization

Practice	Sage	Team	Org	Target	Recommendation
rtk coverage	77%Needs attention	69%	56%	≥ 80% of Bash calls	Install the rtk hook so test, build, git and search commands return compact output.
Knowledge MCP use 15 findings saved back	51%On target	37%	30%	≥ 40% of sessions	On target
Cache hit rate	88%On target	84%	82%	≥ 85% of input	On target
Context past 100K	4%On target	6%	10%	≤ 10% of input	On target
Opus on routine work	21%On target	33%	44%	≤ 25% of routine spend	On target
Sessions using subagents	53%On target	51%	50%	≥ 30% of sessions	On target
Tool error rate	2.9%On target	3.1%	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	–	–	66%Off target	3%On target	30%Needs attention	37%On target	$54
Mar 2026	–	–	68%Off target	6%On target	30%Needs attention	40%On target	$45
Apr 2026	–	–	72%Needs attention	2%On target	30%Needs attention	34%On target	$78
May 2026	30%Off target	–	82%Needs attention	6%On target	26%Needs attention	47%On target	$56
Jun 2026	55%Needs attention	23%Needs attention	86%On target	6%On target	26%Needs attention	36%On target	$39
Jul 2026	88%On target	41%On target	88%On target	4%On target	19%On target	53%On target	$18
Aug 2026	94%On target	63%On target	88%On target	6%On target	17%On target	61%On target	$34
Sep 2026	95%On target	61%On target	93%On target	6%On target	13%On target	62%On target	$17
Oct 2026	85%On target	63%On target	93%On target	3%On target	11%On target	69%On target	$16

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	1,960	1.3M	118.4M	7.6M	2.8M	$123
Opus 5	2,191	5.4M	109.1M	7.3M	3.2M	$207
Opus 4.8	2,100	8.4M	100.5M	6.9M	2.6M	$201
Opus 4.6	2,414	24M	106.3M	8.3M	3.7M	$317
Opus 4.5	107	1.5M	3.6M	328K	142K	$15
Sonnet 5.5	10,861	6.2M	563.3M	36.4M	16.2M	$321
Sonnet 5	13,960	35M	744.1M	49.7M	19.2M	$803
Sonnet 4.6	10,402	81.8M	498.6M	37M	15.9M	$772
Sonnet 4.5	602	9.1M	21.5M	2M	784K	$53
Haiku 5.5	12,379	7.6M	703.2M	45.4M	16.9M	$22
Haiku 4.5	11,345	46.9M	529.7M	36.8M	14.2M	$217
GPT Realtime (meetings)	74	284K	3.8M	262K	116K	$6
Total	68,395	227.6M	3502.1M	238.1M	95.7M	$3,057

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.