navigaraResearch
Live dataReportMethodologyContact
Navigara ResearchNo. 01 · Q1 2026Published 2026‑04‑30

Engineering performance
measured at scale

Commit‑level analysis across six organizations, five quarters.
676 qualifying OSS engineers. One shared axis.

Read the White PaperRead the Methodology

Cloudflare · Vercel · OpenAI · Google · Meta · Microsoft

Disclosure

A short note before the data.

Lines of code and merged PRs measure activity, not value. Neither survives questions about complexity, architecture, or impact. We spent a year building something different, then ran it against codebases you've used: vscode, react, TypeScript, next.js, perfetto, codex, playwright, folly, etc. We have a commercial interest in the conclusion that engineering work is intrinsically resistant to reductive measurement. The paper makes only descriptive and correlational claims. Methodology, sample boundaries, and known limitations are in Section 6 and Appendix A.

The Finding

OSS engineers are delivering more value.

Average performance grew 116% Year over Year.

Across six big tech companies: Cloudflare, Google, Meta, Microsoft, OpenAI, Vercel, engineering throughput value in OSS development grew steadily through 2025. In Q1 2026, it more than doubled.

We checked the obvious explanation first: was it just new contributors inflating the average? No. When we narrow to the 418 engineers who were active every quarter, output still grew 98%.

While we don’t claim it’s only driven by AI, the jump lines up with the rollout of AI coding tools. We understand that engineers might also work more hours on average.

+116%
Open cohort
565 → 676 engineers
CI [+84%, +148%]
+98%
Fixed panel
418 continuously‑active SWEs
CI [+63%, +140%]
+51%
Per‑commit
intensity, not just frequency

The Shape

One quarter in 2026 suggests a trend.

The real question is whether Q2 2026 holds the growth trajectory.

Through 2025, performance per engineer → what we call Engineering Throughput Value (ETV), grew at a steady pace. Single digits each quarter. Q1 2026 changed the slope.

Performance per engineer · % Δ vs. Q1 '25

PERFORMANCE CHANGE VS Q1 '250%+25%+50%+75%+100%+125%+116%Q1 '25Q2 '25Q3 '25Q4 '25Q1 '26
Q1 '25
—
N = 565
Q2 '25
+9%
N = 638
Q3 '25
+18%
N = 683
Q4 '25
+25%
N = 694
Q1 '26
+116%
N = 676

The fixed‑panel cohort moves the same direction. Same shape, different population. The rise survives holding the cohort constant. The corresponding chart is in the white paper (Figure 1b). Whether it survives another quarter is still being determined.

The Composition

The extra performance goes to new value, not maintenance.

So what type of work changed?

Maintenance share fell -10 percentage points (pp). Growth gained +7pp. Fixes moved +3pp. Engineers didn't just ship more → they shifted toward building new things.

ClassificationQ1 '25Q1 '26Δ
Growth(New value)29%36%+7pp
Maintenance56%46%−10pp
Fixes15%18%+3pp

The mix shift rules out one of the three stories from above. This isn't operational drag from incidents and defects → The rise is concentrated in growth‑classified output.

The Spread

The variability in the data shows a wide discrepancy between companies.

Six organizations, six different slopes.

Performance per engineer · ETV · % Δ since first qualifying quarter

The +116% headline averages engineer performance, not company. Every engineer in the sample counts once, regardless of company. Underneath, movement ranges from +51% to +373%. That doesn't mean OpenAI ships 7× more than Meta → it means OpenAI's per‑developer output moved more from its baseline. We're not publishing absolute output figures here.

Organization
Performance Δ
% vs. baseline
OpenAI
cohort entered Q2 '25
+373%
Cloudflare
+157%
Microsoft
+137%
Vercel
+92%
Google
+56%
Meta
+51%

OpenAI's cohort entered Q2 '25 small and grew fast, read its line as a ramp, not a baseline. Cloudflare and Microsoft sit in the middle. Meta and Google moved least, despite being the largest cohorts.

Two patterns worth flagging. The AI‑forward sample (Cloudflare, Vercel, OpenAI) moved more than the incumbent sample (Google, Meta, Microsoft), directionally consistent with the AI‑adoption hypothesis, not proof of it. Cohort size and movement are inversely correlated: smaller orgs moved more.

Per‑quarter trajectory for each company is in the white paper (Section 4, Figures 3a and 3b).

The Mechanism

Bigger changes, not just more of them.

More commits, or bigger commits?

Both, but not equally. Commits per software engineer (SWE) moved +35%. Performance per commit moved +51%. The intensity of each ETV rose more than the frequency of work itself.

Commits per engineer
Performance per commit
0%+15%+30%+45%+60%+35%+51%Q1 '25Q2 '25Q3 '25Q4 '25Q1 '26
Commits per SWE · +35%

More granular commit practices, or genuine throughput gains. The data describes the move without claiming the cause.

Performance per commit · +51%

Larger, more complex changes per commit. The intensity of each unit of work has risen more than 50% over five quarters.

A commit in Q1 2026 carries 51% more weight than a commit in Q1 2025 → larger changes, more complex changes, or both. That's the more interesting half of the headline. Engineers aren't just committing more often. They're shipping bigger units when they do.

Note: the two changes do not compound to the headline. The headline is a mean across SWEs of per‑SWE performance, while the two series here are aggregate commit‑weighted means. Different math, different number (1 + commits/SWE Δ) × (1 + performance/commit Δ) is not expected to equal the +116% reported above.

Methodology

Most of it is open, some of it is proprietary.

How we measured. And what we can't see.

Two‑layer design. A per‑commit scoring engine produces Growth, Maintenance, and Fixes sub‑scores for every merged commit; their sum is the Engineering Throughput Value (ETV). The report layer aggregates ETV into per‑developer, per‑quarter, and cross‑org means.

Organization selection. The six organizations were not sampled randomly. Three — Cloudflare, Vercel, and OpenAI — were selected because each makes frequent public claims about AI productivity gains in its engineering organization. The other three — Google, Meta, and Microsoft — were selected as incumbents of substantially larger scale with strong public reputations for engineering talent density. This is a purposive sample, not a representative one; cross‑org comparisons reflect that.

Sample composition. Public‑repository activity at Google, Meta, and Microsoft is concentrated in developer tooling, SDKs, and open‑source frameworks (TypeScript, react, vscode, playwright, perfetto, hermes, buck2, and similar) rather than the internal product codebases where most engineering at those companies happens. Public‑repo activity at Cloudflare, Vercel, and OpenAI is closer to their core product surface but is still public‑by‑construction. Findings describe activity in the in‑scope public surfaces; generalization to private internal engineering at any of these organizations is not supported by the data.

Who counts. A contributor qualifies as a SWE when the role classifier assigns Software Engineer and they have recorded commit activity in at least 10 weeks of the window. Per‑org results appear only once a cohort reaches the 20‑SWE sample floor.

Authorship attribution. Each commit is attributed to its primary author. Bots are excluded by pattern‑matching on email and display name. Commits authored by a human with AI assistance — Cursor, Copilot, Claude Code, Cody, Devin, agentic coding tools — are attributed to that human. The engine cannot distinguish AI‑assisted from non‑assisted commits when the author is a person; “human authorship” here means the GitHub identity, not the share of the diff written without help.

Uncertainty. 95% confidence intervals are computed by bootstrap over SWEs within each quarter (1,000 iterations). Single‑quarter movements should be read against the band; the multi‑quarter slope is the signal.

Trajectory shape. The five‑quarter series is +0%, +9%, +18%, +25%, +116% — most of the rise is concentrated in Q1 2026 rather than distributed across the window. The concentration is itself the most interpretively interesting feature of the trajectory: a steady‑slope rise and a late‑window inflection have very different causal stories. The next edition will examine whether the Q1 2026 inflection holds, decays, or accelerates.

Temporal alignment. Calendar quarters. Commits attributed to merge quarter, not author quarter. Long‑lived branches land at their merge date.

Limitations. Public repositories only. No view into private work, code review, incident response, planning, or mentorship. Cross‑org comparisons are descriptive; monorepo vs. polyrepo, squash policy, and public‑vs‑private code mix are not controlled.

For the full methodology, scoring engine, decay factors, and the complete repository list — read the white paper

8 pages · two‑layer scoring engine · per‑org breakdowns · 66 repositories enumerated in Appendix A

Key findings

Engineers shipped 116% YoY more. They shifted toward growth (new value). The biggest movement came from intensity per commit, not commit frequency. Whether they shipped the right things is a separate question and one this study can't answer.

Measuring "the right things" needs more than commit history. Navigara's Alignment concept handles it by connecting to Jira or Linear. That's outside this study.

The next edition will examine whether the Q1 2026 inflection holds, decays, or accelerates.

Read the White PaperRead the Methodology

Cloudflare · Vercel · OpenAI · Google · Meta · Microsoft