Deeply AI‑native
79% AI‑attributed11.8× at ~$1,300/wk
Hasn’t hit a ceiling. Output keeps scaling with spend, and at equal spend this cohort ships roughly 2× what partial adopters ship.
Larridin Data
The data behind AI‑powered work.
Recurring insights and benchmarks from the Larridin platform: real‑world AI adoption, costs, tools, work produced, and ROI across enterprises and industries. Based on real production telemetry and invoiced spend.
Finding 1
What engineers actually spend on AI tokens
Weekly billed AI‑coding spend per engineer, among engineers who both shipped code and drew billed spend. Whiskers are 95% bootstrap confidence intervals: 10,000 resamples of observed engineers, no synthetic data.
| Percentile | $ / engineer / week | 95% CI | ≈ Monthly |
|---|---|---|---|
| p25 | $86 | $55 – $117 | $370 |
| p50 (median) | $213 | $178 – $262 | $920 |
| p75 | $490 | $409 – $581 | $2,100 |
| p90 | $911 | $681 – $1,255 | $3,900 |
Invoiced dollars only. Engineers on flat‑fee plans consume tokens that never hit a metered bill, so true consumption sits above these figures. Percentiles were stable in each of the four weeks individually.
Median billed spend
$213
per engineer / week · ≈$920 per month
90th percentile
$911/wk
≈$3,900 per month per engineer
Spend spread
10×+
p90 engineer vs p25 engineer
Finding 2
More money doesn’t mean more output. The gap between the curves is skill.
AI spend alone can’t tell you much about productivity: spend and output only correlate where engineers are AI‑native. Both AI‑native cohorts below come from the same company, with the same tools, the same prices, and the same ≈$170/week starting point, so the gap between their curves is skill, not a pricing artifact.
Partially AI‑native
37% AI‑attributed6.2× at 7.5× spend
Real returns, diminishing: marginal payoff drops by about half once spend passes ~$600/week.
Low‑AI
≤15% AI‑attributed~1.9× max, flat
Extra dollars buy activity, not output. Flat across a 20× spend range ($21 to $421/week).
Spend vs shipped output, by cohort
| Cohort | Spend band median | Output multiple |
|---|---|---|
| Deeply AI-native | $162 /wk | 1.0× |
| Deeply AI-native | $411 /wk | 3.5× |
| Deeply AI-native | $797 /wk | 7.5× |
| Deeply AI-native | $1,305 /wk | 11.8× |
| Partially AI-native | $181 /wk | 1.0× |
| Partially AI-native | $355 /wk | 1.7× |
| Partially AI-native | $625 /wk | 3.6× |
| Partially AI-native | $1,351 /wk | 6.2× |
| Low-AI | $21 /wk | 1.0× |
| Low-AI | $53 /wk | 1.6× |
| Low-AI | $179 /wk | 1.8× |
| Low-AI | $271 /wk | 0.5× |
| Low-AI | $421 /wk | 1.9× |
Each point is a spend band’s median output as a multiple of that cohort’s lightest band (1.0× = lightest spenders in that cohort). Cohorts of 27 to 46 engineers grouped by AI‑attributed share of shipped output; all values carry 80% bootstrap bands. The low‑AI cohort comes from a second company with partial repository coverage; the two AI‑native curves are the robust comparison.
Summary
What companies actually pay for AI coding, and where the returns bend
We asked two questions. How much are engineers actually spending on AI coding agents? And does more spend buy proportionally more output?
-
The median engineer draws $213/week (≈$920/month) in billed AI‑coding spend. At the 90th percentile it reaches $911/week, a 10×+ spread across engineers.
-
Spend converts to output only as far as AI‑nativeness carries it. Deeply AI‑native engineers keep climbing at $1,300/week; partial adopters’ payoff halves past ~$600/week; low‑AI output stays flat from the first dollar.
-
Same company, same tools, same prices: the gap between the curves is skill. At equal spend, deeply AI‑native engineers ship roughly 2× what partial adopters ship.
The strategic takeaway
AI capability is still unevenly distributed.
More budget converts to more output only where fluency already exists, so there is no universal “ideal budget” for AI tools. Don’t copy another company’s cap: track your own team’s ROI curve and set review triggers where it levels off. Today the spread between deeply AI‑native engineers and everyone else is roughly 2× at equal spend, and the data suggest that gap is not done widening.
Methodology & integrity
How this was measured
Sample
Engineers instrumented by Larridin who, in the four complete weeks ending August 2, 2026, both merged code and drew billed AI‑coding spend. Company names and exact cohort sizes withheld for customer confidentiality.
Spend
Provider billing feeds only: invoiced dollars. Telemetry‑priced consumption on flat‑fee plans is excluded, making Finding 1 a conservative floor on true consumption.
Output
Each merged pull request is scored by model‑assessed complexity (five levels, weighted 1–24), discounted for low‑quality “slop” and missing tests, and scaled by code churn. Merged‑PR work only. How engineering output is computed →
Statistics
All uncertainty bands are nonparametric bootstrap intervals (10,000 resamples of observed engineers). No data points were simulated, imputed, or extrapolated.
Read before quoting
These relationships are associational, not causal: high‑output engineers may spend more because they ship more. Four weeks of billing is a snapshot; percentiles were stable week to week, but seasonal effects are not averaged out. Engineers whose work is mostly review, operations, or unmerged agent output appear low‑output by construction. Full output methodology →
All insights
New analyses from the Larridin Benchmark land here as they’re published.