Where every number comes from.

The setup, the baseline, what was counted, how it was priced.

Live demo and pilots

Cost, tokens and speed

Cost

$0.031vs $0.234

per answer: 8.3× lower, 88% smaller bill.

Speed

8.2 svs 18.6 s

question to final answer: 2.2× faster.

Tokens

14×

fewer tokens, every input and output step counted.

When

13 Sep to 19 Oct 2026

Public demo: open-source codebase plus company data, 30,000+ entries.

Two sides

gpt-5-mini

Same model, same questions: Mithrandir vs a mirror agent reading the repo.

Summarised

Medians

Of per-question ratios. Claude Opus list prices, $5 / $25 per M tokens.

Measured, not estimated

An eighth of the cost. Every answer sourced.

8.3 times: lower cost$0.031 vs $0.234 per answer.
Without Mithrandir$0.234
With Mithrandir$0.031
14 times: fewer tokensMeasured across 30,000+ entries.
2.2 times: faster results8.2s vs 18.6s per answer.
100 percent: of answers sourcedEvery answer cites its source.
1.5 times: better answersGraded blind against the same model alone.
zero: leaksAcross more than 30,000 entries.

At a glance

Claim on the homepage	Where it comes from
8.3× lower cost · 88% smaller bill	Live-demo measurement
14× fewer tokens	Live-demo measurement
2.2× faster results	Live-demo measurement
100% of answers sourced	Public demo and pilots, more than 30,000 entries
1.5× better answers	Pilot, blind-judged
0 leaks of context	Pilots, demo and internal audit, more than 30,000 entries
Savings calculator	Estimate. Source: Anthropic
Work that takes people hours, weeks or months, done in 10 to 30 seconds.	What the product does
Finds the relevant story across thousands of documents in about 10 seconds, with every connection.	What the product does
Up to 396 to 792 hours a year saved per person. Up to $5,000 to $11,000 more revenue a year per person.	Estimate. Sources: McKinsey; Coveo; Damodaran; Am Law; Fortune; Sage; Axios

Demo and pilots

Every answer sourced

100%

of answers cite a source.

30,000+

entries on the demo and in pilots.

Pilot

Better answers

1.5×

higher score than the same model alone.

Blind

Blind-judged: labels removed, order shuffled, scored 0 to 5.

Pilots, demo, audit

Zero leaks of context

0 leaks

across 30,000+ entries: pilots, public demo, August 2026 audit.

2,500 / 2,500 held

Brute-force permission attacks across 2,500 entries.
Every one held (security).

How is it estimated

AI cost saving

Heavy-usage AI bill per sector (Anthropic usage figures, Claude Opus list prices)

scaled from 15 people to your team size

× (1 − 0.88) measured saving

Revenue from recovered hours

1.8 to 3.6 h/day × 220 days

× (team size / 15)^0.15, capped at 6 h a day

× team size × sector revenue per employee / 2,080 h

× 25% realized

Search time: McKinsey; Coveo

Sector	Revenue per employee	More revenue a year per person
Legal (per lawyer)	$895,000 (Am Law)	Up to $43,000 to $85,000
Fintech	$750,000 (Axios)	Up to $36,000 to $71,000
Finance	$378,230 (Damodaran)	Up to $18,000 to $36,000
Software	$148,811 (Damodaran)	Up to $7,000 to $14,000
Hardware	$249,155 (Damodaran)	Up to $12,000 to $24,000
Logistics	$69,059 (Damodaran)	Up to $3,000 to $7,000
Enterprise	$687,094 (Fortune)	Up to $33,000 to $65,000
Consulting	$158,000 (Sage)	Up to $8,000 to $15,000
Research lab	$362,653 (Damodaran)	Up to $17,000 to $35,000
Others (US overall)	$111,028 (Damodaran)	Up to $5,000 to $11,000

Measure it on your own work.

One workflow, the same comparison, your raw numbers.

Book a call