Executive summary
18 months into the agentic era, the question has flipped from "is AI making engineers faster?" to "where did the speed go?" This quarter's system data from 1,470 organizations and 21,409 engineers gives an unusually direct answer: in the median organization, output per engineer rose 1.8x in the three quarters from Q3 2025 to Q2 2026, and nearly every metric leaders use to steer that gain is now measuring the wrong thing.
Scroll sideways or zoom to read the full chart.
Four themes define the quarter:
1. Value and volume have decoupled
Volume and value now move on separate curves. Since January 2025 merged PR volume is up 6x and the median merged PR grew from 36 lines to 109, while the value each PR carries rose 59%: three times the code for half again the value. The input side inflated further still, with tokens consumed per unit of shipped output up 14x since January. Platform-wide, AI's share of merged output value crossed the halfway mark in June at 52%, up from 41% in April. Every volume metric (lines, tokens, PRs, adoption) overstates progress; only complexity-weighted output survives as a ruler.
Scroll sideways or zoom to read the full chart.
2. The waste is not where you think
Of every $100 in AI session spend, only $26 ends in a clean ship: a session that completes its work and whose code merges. Another $29 merges without a completion signal, $20 completes and never merges, $17 carries no classified outcome, and $8 ends abandoned or blocked. Fewer than 4 in 10 agent sessions with a classified outcome complete. Of the tool calls that fail, the largest classified bucket at 41% is a file, path, or resource the agent expected and did not find, though an unclassified bucket of the same size sits beside it. The highest-leverage fix we can measure is boring: sessions that invoke a codified skill add 5.4x more lines.
Scroll sideways or zoom to read the full chart.
3. The quality panic is aimed at the wrong layer
The industry expected AI volume to break the quality gates. It hasn't: revert rate fell below 1% while monthly merged PR volume reached 432,000, and the median wait for review is at least 2.5x shorter than its 2025 peak. AI-signed review events grew 16x between Q4 2025 and Q2 2026, though they are still 1.1% of all reviews. The real structural risk is concentration: the top 10% of engineers produce roughly 40% of output at every organization size, and in 100+ engineer organizations 14% of the team holds half the codebase expertise.
4. Spend became governable, but most orgs haven't started
Routing is the AI cost lever that requires no behavior change from engineers. In July 2026, the most recent complete month, the Weave Router saved 58% against list price, and open-weight models generated 61% of those savings on 31% of routed traffic. Almost nobody is doing it yet: fewer than 1% of engineers on the platform routed a request in July, the Weave Router's second month of general availability. Read that as a starting line rather than a weakness. The lever is two months old, and routing is only now starting to look like a normal line item instead of an experiment. The spending gap between the top 1% of engineers by output ($34/day) and the bottom half ($2/day) is not where the waste sits; ungoverned defaults are.
The bottom line for leaders
The gain is real, and it is table stakes; your competitors got it too. What separates organizations now is conversion: measuring value instead of volume, aiming enablement at the bottom half of the distribution, engineering agent context instead of prompt phrasing, and routing every request like the purchasing decision it is. The rest of this report is the benchmark grid, segmented by organization size, so you can locate yourself against peers rather than averages.
Read the full report
Enter your work email to unlock every chart, benchmark and methodology note in the report.