Model
Live API
claude-sonnet-5 via the live Anthropic API. Exact billed input/output tokens — not estimates.
Measured evidence
We didn’t believe it either. So we ran it — again and again.
Same question. Same model. Same files. The only change: an AI re-reading whole files to find the answer versus asking Find and Seek. When it can’t find it, it reads everything — and that’s the bill.
41–69%
cheaper per run on the same task
27–45%
faster wall-clock, start to answer
Bounded
the bill can’t spiral — no runaway re-read tail
69–76%
fewer input tokens — why the bill drops
53–63%
fewer tool calls, fewer round-trips
Comparison
Three workloads. Both sides finish.
| Workload | Side | Tool calls | Input tokens | Cost / run | Wall-clock |
|---|---|---|---|---|---|
| Spring Glow exec summary1 team · campaign one-pager | a normal AI assistant | 30 | 150,873 | $0.368 | 81.6s |
| find & seek | 11 −63% | 39,504 −74% | $0.114 −69% | 44.5s −45% | |
| Cross-team reconciliation3 teams · finance/legal spend disagreement | a normal AI assistant | 58 | 444,415 | $1.052 | 194.5s |
| find & seek | 26 −55% | 138,070 −69% | $0.386 −63% | 119.7s −38% | |
| Full program audit6 teams · ops/CFO spend-and-risk sweep | a normal AI assistant | 93 | 1,169,902 | $0.758 | 394.3s |
| find & seek | 44 −53% | 284,855 −76% | $0.445 −41% | 287.0s −27% |
Cost and speed advantages narrow as tasks span more teams — but token reduction stays strong and never reverses on a correctly measured run.
Three reports
Open the full write-up. Read the answers.
Each report is one real workload: what we asked, what both sides wrote, what surprised us, and what the numbers do — and don’t — prove.
Exhibit · 6 teams · Start here
The work itself
Don’t take the numbers — read the audit memo an agent actually produced. Six conflicting budgets reconciled, a self-billing vendor caught 3 runs out of 3, every claim cited.
Report · 1 team · Live
Spring Glow executive summary
One campaign. One page. Both sides finish — Find and Seek ~69% cheaper. Full answers published.
Report · 3 teams · In progress
Cross-team spend reconciliation
Campaign, Finance, and Legal disagree on the total. ~63% cheaper — full answers coming next.
Report · 6 teams · Live
Full program audit
Six departments. Conflicting totals. Risk register. Identity flags. Full answers published.
The corpus is synthetic by design: the same document types a real company holds, with no real customer data, which is why we can publish the raw output in full.
How we tested
What we tested.
Files
Synthetic company files
Multi-team permission groups — like separate folders on a shared drive. Same questions a marketing lead, finance team, or CFO would actually ask.
Runs
3 per task
Both sides allowed to finish normally. We publish ranges, never a single unexplained percentage.
Without Find and Seek: the agent lists folders, searches filenames, and reads whole files.
With Find and Seek: the agent gets the most relevant passages with their sources, not whole-file dumps.
We grade these numbers harder than you will
The ranges hold when Find and Seek answers questions about connected files — not general chat that never touches them. Quote the range across 1 / 3 / 6 permission groups, never one lucky run: the consistent wins are fewer round-trips, a bill that can’t spiral, and the exact passage with its source. Walls are per team today; per-person sign-on ships with the enterprise build.