Larridin vs. DX: 2026 Engineering Benchmarks Compared

DX’s Core 4 benchmarks show engineering activity increasing as AI adoption grows. Larridin’s preliminary data tells a more nuanced story: engineers are producing more valuable work and addressing harder problems, but AI costs, review delays, and weak agent readiness are creating new bottlenecks.

Key Takeaways

• Raw engineering activity has increased approximately 12% since the beginning of 2026, while Larridin’s complexity- and quality-adjusted Engineering Output has increased 22%.
• AI token costs have grown 10x in six months and are now equivalent to approximately 2% to 10% of engineering headcount costs, depending on the organization.
• Token waste remains high at 28%, although it has improved from 32% in Q1 2026.
• Code review times have increased approximately 50%, while the rate of rubber-stamped reviews has risen 30%.
• Engineers are tackling more complex problems, including a growing amount of technical-debt remediation, but most codebases remain poorly prepared for autonomous agents.

DX’s 2026 Core 4 benchmarks show pull request throughput increasing nearly 20%, defect ratio improving 16.9%, the Developer Experience Index rising 2 points, and innovation ratio declining 10%.

DX also reports that self-reported AI-authored code increased 140% and weekly AI time savings rose from 3.9 to 6.2 hours per developer. At the same time, review turnaround and incremental delivery declined across most benchmark segments.

The overall picture is clear: AI is making it easier to generate code, but organizations have not yet redesigned the software development lifecycle to absorb that additional output efficiently.

Larridin sees the same tension. Engineering output is growing, and engineers are taking on more complex work. However, AI costs, token waste, codebase readiness, and review quality are becoming increasingly important constraints.

How Larridin Compares With the 2026 DX Core 4 Benchmarks

Methodology note: The Larridin numbers below are illustrative draft figures based on a placeholder cohort of approximately 3,200 engineers across 25 organizations, measured from January 1 through July 31, 2026. They must be replaced or validated before publication. DX and Larridin may use different definitions, cohorts, and measurement methods, so comparisons should be treated as directional.

Engineering Activity Is Up 12%, but Engineering Output Is Up 22%

Larridin’s preliminary data shows raw engineering activity increasing approximately 12% from January through July 2026. This includes measures such as pull requests, commits, and completed engineering work.

Activity, however, does not account for the difficulty or durability of that work. A small dependency update and a major architectural change should not contribute equally to an engineering productivity benchmark.

Larridin’s Engineering Output metric adjusts for both complexity and quality. Using this measure, output increased approximately 22% over the same period.

The gap between the 12% increase in activity and the 22% increase in adjusted output suggests that engineers are not simply producing more changes. They are completing harder, more valuable work.

This is particularly important as AI takes over more routine coding tasks. The benefit may not appear exclusively as a dramatic increase in pull request count. Instead, AI can give engineers more capacity to work on complex migrations, architectural improvements, reliability initiatives, and technical debt.

Engineers Are Addressing More Complex Problems

DX reported a 10% decline in innovation ratio, which measures the share of work allocated to feature development. That result could be interpreted as a negative signal, but feature allocation is only one dimension of engineering impact.

Larridin’s data suggests that engineers are allocating more capacity to complex work that was previously postponed. This includes platform modernization, refactoring, infrastructure improvements, test coverage, reliability work, and technical-debt remediation.

Technical debt, in particular, is being addressed more rapidly. AI makes it easier to understand unfamiliar code, generate tests, update repetitive patterns, and execute migrations that previously competed with roadmap work.

This shift may reduce short-term feature allocation while improving the organization’s ability to deliver safely and quickly in the future. Engineering leaders should distinguish between low-value maintenance and strategic technical-debt reduction before concluding that a decline in feature work represents a decline in impact.

AI Token Costs Have Increased 10x

The economics of AI-assisted engineering are changing quickly. Across Larridin’s illustrative cohort, token costs increased approximately 10x during the six months ending July 31, 2026.

For some organizations, AI token spending remains relatively small. For others, it is already equivalent to approximately 2% to 10% of engineering headcount costs.

This is becoming a material budget category, particularly as companies move from individual coding assistants to agents that plan tasks, inspect repositories, run tests, review code, and iterate autonomously.

The opportunity is substantial, but unmanaged usage can erase part of the productivity benefit. Engineering leaders need to understand which tools, models, teams, repositories, and workflows generate useful output and which consume tokens without producing durable engineering value.

Token Waste Has Improved, but Remains at 28%

Larridin’s Token Waste metric estimates the share of AI spending associated with abandoned, duplicated, unnecessary, or low-value agent activity.

Token waste decreased from approximately 32% in Q1 2026 to 28% in the latest measurement period. This improvement suggests that teams are getting better at selecting models, constructing prompts, managing context, and choosing appropriate tasks for agents.

Nevertheless, a 28% waste rate means that more than one-quarter of token expenditure may still be avoidable.

Common sources of waste include agents repeatedly rediscovering codebase context, excessive context-window usage, duplicated work across tools, failed task loops, poorly scoped requests, and using expensive models for routine operations.

AI costs are growing rapidly, but organizations still have significant opportunities to reduce those costs without reducing adoption.

Agent Readiness Remains Around 2 out of 5

AI agents are only as effective as the environment in which they operate. Larridin’s Agent Readiness assessment evaluates factors such as documentation, test reliability, code structure, dependency clarity, development-environment consistency, and the availability of machine-readable context.

The average codebase currently scores approximately 2 out of 5.

This means that many organizations are deploying increasingly capable agents into environments that were designed exclusively for human developers. Agents spend too much time locating information, resolving inconsistent setup instructions, navigating unclear ownership, and recovering from unreliable tests.

Improving Agent Readiness can increase successful task completion while reducing token consumption and human supervision. Teams can evaluate their current readiness at code.larridin.com.

Code Review Is Becoming the New Bottleneck

Code generation is accelerating faster than code review. Larridin’s preliminary benchmarks show code review times increasing approximately 50% across organizations and team sizes.

This does not necessarily mean reviewers are working more slowly. The review system is receiving more code, from more sources, at a higher rate than existing processes were designed to handle.

At the same time, the rate of rubber-stamped reviews has increased approximately 30%. These are reviews that are approved unusually quickly or with limited evidence of substantive inspection.

The combination is concerning. Some pull requests are waiting longer for attention, while others may be passing through the review process without sufficient scrutiny.

Organizations should not treat an approval event as proof that a meaningful review occurred. Review quality should be evaluated alongside review time, comment depth, change complexity, rework, code turnover, reverts, and production outcomes.

The Software Development Lifecycle Needs to Be Rethought

Most software development processes assume that writing code is one of the primary constraints. AI changes that assumption.

When code can be generated much faster, bottlenecks move into planning, context gathering, validation, review, testing, deployment, and production monitoring. Adding more code to an unchanged workflow can increase queues and risk rather than improve delivery.

Organizations need to redesign the software development lifecycle around a combination of human and agent work. That includes clearer task specifications, better codebase documentation, automated validation, complexity-aware review routing, stronger test environments, and more deliberate cost controls.

The goal should not be to maximize AI-generated code. It should be to maximize durable engineering outcomes per dollar and per hour of human attention.

What Engineering Leaders Should Take From These Results

1. Measure output, not just activity. A 12% increase in engineering activity becomes a 22% increase when complexity and quality are considered. Pull request count alone can understate the value of harder work.
2. Treat AI spending as an engineering efficiency metric. Token costs have increased 10x, but 28% of that usage may still be waste. Cost attribution and outcome measurement are becoming essential.
3. Prepare the codebase for agents. An average Agent Readiness score of 2 out of 5 indicates that the surrounding engineering environment is limiting agent effectiveness.
4. Redesign review workflows. Review times are up 50%, while rubber-stamping is up 30%. Faster code generation cannot produce durable gains if review capacity and quality do not scale with it.
5. Recognize the value of technical-debt remediation. Engineers are using AI to address more complex and previously deferred work. A decline in feature allocation does not automatically mean that engineering impact has declined.

Benchmark AI’s Engineering Impact, Not Just Its Activity

AI is increasing engineering capacity, but the gains are not automatic. Costs are growing, review systems are under pressure, and most codebases are not yet ready for dependable autonomous development.

At the same time, engineers are producing more complexity- and quality-adjusted output and tackling problems that were previously too expensive or time-consuming to address.

The next phase of AI adoption will not be defined by how much code organizations generate. It will be defined by how effectively they convert AI investment into durable, high-quality engineering outcomes.

Talk to an Expert