The economics of Devin
A running ledger of the published, auditable cost/return numbers — the thing every evaluation eventually asks for. Only figures that appear in a named customer story or a public report are included; nothing is extrapolated.
Unit economics
| Number | Deployment | Context |
|---|---|---|
| ~$17/hour of engineering output | Compiled Health | ~430 engineer-hours delivered by 2 researchers in 2 months — “≈ +4 engineers” on a 4-person team |
| 20× cost savings | Nubank | Migration program vs. the same work staffed conventionally; 8–12× engineering efficiency alongside |
| 85% cost reduction on legacy-app revamps | The Citation Group | ROI-measured rollout; 271 merged PRs @ 80% in 3 months |
| 50% engineering efficiency gain | The Citation Group | Same study |
Throughput per person
| Number | Deployment | |
|---|---|---|
| ~100 PRs/day from a small team | Melius | |
| 1,800+ PRs in 6 weeks from an 8-person team | Tavrn (HIPAA-restricted fleet) | |
| 50 → 170 weekly backend PRs in four weeks | Común (13-person fintech) | |
| 80 merged PRs/week, 10,000+ engineer-hours saved | Ramp | |
| 92% of merged PRs authored by Devin | EBANX | |
| 30% more PRs per engineer vs. Claude Code alone | Latent Health | |
| ~2× user-story output in three months | GE Aerospace | |
| 2× engineering throughput; 4× faster automation rollout | Rohlik Group | |
| 3× deployment growth in four months; 78% PR merge rate | Endava (2,000+ employees) | |
| 93% of merged code Devin-originated; alert→investigation 35s | Tavrn | |
| 20–40 bugs/alerts first-passed daily in ~5 min; Sentry noise −50% | Crosby | |
| 25% of total code volume — tied top contributor | Hamming | |
| 40× merged-PR throughput on 2.7× headcount (>10×/engineer) | Cognition internal |
The honest-ledger rule
Cognition’s own 18-month performance review supplies the rarest numbers — a vendor publishing its own yardstick: PR merge rate 34%→67%, vulnerability fixes 20× faster than humans (30 min→1.5 min each), ETL migrations 10×, typical coverage 50–60%→80–90%, ~⅓ of its own webapp commits Devin-authored. Vendor-published, but unusually falsifiable.
The most useful economics report is also the least flattering: Leverages (JP) published its rollout with a literal ACU-spend-vs-salary comparison across 14 teams and 177 merged PRs — including the conclusion that returns depend on task mix, not on “agent hours” as a fungible unit (report).
Pattern across the numbers that hold up:
- Wins concentrate where work is voluminous and checkable — migrations, connector maintenance, hygiene loops, test coverage — not where work is ambiguous.
- The reported multipliers come from orgs that built the surrounding machinery (playbooks, review gates, reference benchmarks). No published case reports high ROI from ad-hoc usage.
- Time-to-first-value is consistently short: AHEAD measured full productivity in 2 business days; Courtney shipped to production in ~21 hours.
Sources
Every figure above links to the named customer story at devin.ai/customers or the cited public report. If a number here is later revised by its source, the table follows the source.