AI Made Devs 19% Slower: METR Study [2026]

Developers are telling surveyors that artificial intelligence now writes close to half of their code. At the same time, the most rigorous controlled study of AI coding agents to date found that the same class of tools made experienced engineers slower, not faster, at real work. Both things are true in October 2026, and the gap between what developers believe and what researchers measured is becoming one of the defining stories in software engineering this year.

A global survey of 705 developers and IT leaders found that 42 percent now say AI writes at least half of their code, up from just 12 percent a year earlier. Only 21 percent of developers say they spend more than half their week writing new code from scratch anymore. Meanwhile, a randomized controlled trial from METR, a nonprofit AI research and evaluation organization, found that giving 16 experienced open source developers access to AI tools made them 19 percent slower on 246 real tasks pulled from codebases they already knew well. The developers had expected a 24 percent speedup. This disconnect, between adoption numbers climbing and productivity data staying flat or falling, is reshaping how engineering leaders think about AI coding agents heading into 2026.

Google · Preferred Sources

Don't miss new tech stories on Google

Add TrendinTech once in the Google app and our stories appear in your news suggestions.

Add Now

The AI coding agents productivity paradox, explained

The term “productivity paradox” has shown up before in technology history, most famously with personal computers in the 1980s, when businesses spent heavily on PCs without a matching jump in measured output. AI coding agents appear to be running into a similar wall, except the data arriving in 2026 is more granular than anything available during the PC era. Teams can now track commits, pull requests, and resolved tickets automatically, which means the gap between perceived speed and actual throughput is visible in near real time rather than showing up years later in national productivity statistics.

What makes the current moment different is scale. AI coding agents are not a niche tool anymore. Code editors and IDE plugins across the industry, from Zed and VS Code to JetBrains Air, now ship with some form of agent integration by default. Cursor, GitHub Copilot, Claude Code, and Google’s Gemini-based tools all compete for the same developer attention, and most engineering organizations have rolled out at least one of them company-wide. That scale is exactly why the METR findings landed with such force: this is not a story about a handful of early adopters struggling with unfamiliar tooling. It is a story about mainstream engineering workflows absorbing a technology faster than anyone has measured its actual effect on delivery speed.

Inside the METR study: how researchers measured AI coding speed

METR’s trial design is what gives the result its weight. Instead of asking developers to self-report how fast AI made them, researchers recruited 16 experienced open source contributors and had them complete 246 real issues inside codebases they already maintained. Each task was randomly assigned to be completed either with AI assistance allowed or with it switched off, which removes the usual bias of developers picking easy tasks to show off a tool. Before starting, the developers predicted that AI access would cut their completion time by 24 percent. METR said the results surprised its own research team: developers thought they were about 20 percent faster with AI tools, but they were actually 19 percent slower when they had access to AI than when they did not, according to a summary METR posted alongside the study.

The gap between expectation and outcome did not close once the work was finished, either. METR found that developers expected AI to speed them up by 24 percent, and even after experiencing the slowdown, they still believed AI had sped them up by 20 percent. That persistence of belief, despite measured evidence to the contrary, is arguably the more unsettling part of the study for engineering managers trying to decide how much weight to put on developer self-assessment when evaluating new tooling.

Where the lost time actually went

METR’s time-tracking data points to a specific mechanism rather than a vague sense that AI is unreliable. When AI access was allowed, developers spent less time actively coding and searching for information on their own, and instead spent that time prompting AI, waiting on outputs, and reviewing what came back. METR’s researchers described this shift directly, noting that developers spend time prompting AI, waiting on and reviewing AI outputs, and sitting idle while a model generates a response. None of that counts as writing code, but all of it eats into the clock.

Without AI access (baseline):
  Reading code .......... 35%
  Writing code ........... 40%
  Searching docs ......... 15%
  Testing/debugging ...... 10%

With AI coding agent access:
  Reading code .......... 25%
  Prompting the agent ..... 20%
  Waiting / idle .......... 15%
  Reviewing AI output ..... 25%
  Testing/debugging ....... 15%

That breakdown illustrates why raw code output can rise even as total delivery time gets worse. A developer can generate far more lines through an agent, yet still finish the ticket later than they would have by writing it themselves, because verification and correction now take up a chunk of the day that used to go toward direct problem solving.

42 percent of code is now AI written, survey finds

The adoption side of the story is not in question. The 705-person global survey of developers and IT leaders found that the share who say AI writes at least half their code more than tripled year over year, moving from 12 percent to 42 percent. That is a faster adoption curve than cloud computing, containerization, or even mobile app development saw at a comparable stage, and it reflects how low the barrier to trying an AI coding agent has become. Most tools are a plugin install or a terminal command away, with free tiers that let individual developers start without budget approval.

The same survey found that only 21 percent of developers now spend more than half their week writing new code from scratch, a sharp decline from what the role looked like even two or three years ago. Current reporting frames this as a structural shift in what a software engineer actually does day to day, moving the center of gravity away from typing new logic and toward reviewing generated output, verifying correctness, managing system architecture, and handling the governance questions that come with letting a model touch production code. Several previously published pieces on this trend line up with that shift, including coverage of how some employees report exaggerating their AI usage to managers, which suggests the adoption numbers themselves deserve a skeptical read even as they climb.

More commits, same product output: the organizational data

Separate analysis of engineering data gathered after organizations rolled out AI coding agents company-wide tells a more nuanced version of the same story. Code production rose by roughly 30 percent after adoption, and commit volume climbed by about 20 percent. Pull requests increased by around 23 percent as well. Those three numbers, taken alone, would suggest AI coding agents are an unambiguous win for engineering throughput.

The catch shows up in the next metric. Resolved Jira issues and completed epics, the measures that actually track finished product work rather than raw code volume, did not show a statistically significant increase over the same period. In other words, teams are producing more commits and more pull requests without a matching rise in completed features or closed tickets. That mismatch mirrors the individual-level finding from METR’s trial: AI agents can expand the volume of code generated without delivering a proportional increase in the amount of software that actually ships.

For engineering leaders who have been using commit counts or PR velocity as a proxy for team output, this is an uncomfortable read. It suggests that dashboards built around those metrics may be overstating the real return on AI tooling investment, and that a healthier way to track impact is to measure cycle time on completed, reviewed, and merged work rather than raw activity volume.

Gemini 4 Argon enters the agent race

While the productivity debate plays out, the vendor side of the market keeps shipping new models built specifically for agentic workloads. Google has rolled out Gemini 4 Argon globally to developers and enterprises through the Gemini API, AI Studio, and the broader Gemini platform. The model is positioned for long-running agent workflows, including code migration projects, enterprise research tasks, and cybersecurity work that can stretch across many sequential steps rather than a single prompt and response. Google’s introductory API pricing lists Gemini 4 Argon at 2 dollars per million input tokens and 10 dollars per million output tokens, a rate aimed at making extended agent sessions economically viable for enterprise customers running the model continuously rather than in short bursts.

Gemini 4 Argon’s arrival matters for the productivity debate because long-running agents are exactly the use case where the METR findings are most relevant. A model that can work autonomously for extended stretches shifts even more of the developer’s time toward review and verification, the same categories METR identified as the biggest time sink in its study. The model has already drawn scrutiny on that front, with the UK’s AI Security Institute flagging risks tied to Gemini 4 Argon ahead of its wider release, a sign that regulators are paying as much attention to agent capability growth as enterprise buyers are.

Anthropic’s internal infrastructure bet on agent workloads

The strain that agentic coding puts on existing developer infrastructure is not limited to individual productivity. Anthropic has reportedly built an internal source control system called Scion, designed specifically to handle the volume of changes that agent-heavy workflows generate. According to reporting on the project, daily merges inside Anthropic’s own monorepo rose from roughly 2,500 in May to more than 30,000 by September, a more than tenfold increase inside a single company’s codebase in under five months. That growth curve is a useful data point on its own: it shows that even organizations building the AI models driving this shift are having to re-engineer their internal tooling just to keep pace with how much code their own agents are generating.

Anthropic’s pricing moves elsewhere in its lineup reinforce how central agent workloads have become to its roadmap. The 90 percent price cut on Claude Haiku 5.5 that triggered a broader AI pricing war this year was explicitly framed around making high-volume agent calls cheaper, since an autonomous coding agent might make dozens or hundreds of model calls to complete a single task, compared to the handful of calls a human-driven chat session would need.

Competitive field: how the major AI coding agents stack up

The AI coding agent market has consolidated around a handful of recognizable players, each targeting a slightly different part of the developer workflow. Cursor has built its business around an editor-native agent experience, including cloud-hosted agents that can run tasks in the background while a developer works on something else, a workflow covered in detail in earlier reporting on how to set up Cursor’s cloud agents. GitHub Copilot remains the most widely distributed option by virtue of its default presence across Microsoft’s developer tools. Anthropic’s Claude Code targets terminal-first, agentic workflows for engineers who want deep codebase context rather than a chat window bolted onto an editor.

Google’s entry, now anchored by Gemini 4 Argon, leans on the company’s existing cloud and enterprise relationships to push agentic coding into large organizations that already run workloads on Google Cloud. Meta has also pushed into this space, with its Muse Code tool squaring off directly against the established players, a competitive matchup detailed in prior coverage of Meta’s Muse Code and its four AI coding rivals. Microsoft, for its part, has framed its ambitions even more broadly, with chief executive Satya Nadella describing a vision in which a 137-billion-parameter model turns the PC itself into a software factory rather than treating AI coding as a feature bolted onto an existing editor.

What unites all of these approaches is the same tension the METR study surfaced: every vendor is racing to make agents more autonomous and more capable of running longer unsupervised, while the clearest independent research available says that exact capability, extended autonomous work requiring heavy human review afterward, is where the measured productivity gains currently break down.

Historical context: software’s long history of productivity paradoxes

Economists have documented versions of this pattern before. Nobel laureate Robert Solow’s famous 1987 observation that computers were showing up everywhere except in the productivity statistics described a multi-year lag between heavy capital investment in a new technology and any measurable payoff. It eventually took years of complementary changes, in business processes, employee training, and organizational structure, before PC spending translated into visible output gains.

Software development has its own smaller-scale version of this story. The shift from manual testing to automated test suites, the move from on-premises servers to cloud infrastructure, and the adoption of continuous integration pipelines all produced short-term disruption and slower delivery before the productivity benefits became visible. AI coding agents appear to be following a similar curve, except compressed into months rather than years, because the tools themselves iterate on a much faster release cycle than infrastructure or process changes ever did. The open question for 2026 and 2027 is whether that compression means the productivity payoff arrives faster too, or whether the pace of new model releases simply resets the learning curve before teams fully adapt to the last one.

Market impact: what this means for developer tool vendors

For venture-backed coding agent startups, the METR findings and the organizational data both complicate the pitch that has driven valuations over the past two years. If the clearest value proposition shifts from raw speed to code volume or feature coverage, sales teams will need a different story to tell procurement departments, one built around code quality, security review coverage, or developer retention rather than simple time savings.

Larger platform vendors are better positioned to absorb this shift because they can bundle agent capability into existing subscriptions rather than needing the agent itself to justify a standalone purchase. That dynamic favors Microsoft, Google, and Anthropic, each of which can point to broader product lines, over independent coding agent startups that need the productivity story to hold up on its own. It also explains why pricing has become such an active front in the competition, since undercutting rivals on a per-token basis is one of the few levers vendors can pull quickly while the productivity debate plays out in public.

What researchers and engineering leaders are saying

METR’s own public statements about its findings are notably direct, which is part of why the study has traveled so widely since publication. The organization has summarized its core result plainly: “The results surprised us: Developers thought they were 20% faster with AI tools, but they were actually 19% slower when they had access to AI than when they didn’t,” METR said in its summary of the trial.

In its full write-up, METR stated the finding even more bluntly: “When developers use AI tools, they take 19% longer than without, AI makes them slower,” according to the organization’s published study. The report also detailed the mismatch between expectation and reality, noting that “developers expected AI to speed them up by 24%, and even after experiencing the slowdown, they still believed AI had sped them up by 20%.”

On the mechanism behind the slowdown, METR explained that “when AI is allowed, developers spend less time actively coding and searching for information, and instead spend time prompting AI, waiting on/reviewing AI outputs, and idle,” a pattern the organization’s time-tracking data captured directly rather than inferring from self-reports. METR’s broader takeaway for the industry was cautious but pointed: “It seems likely that for some important settings, recent AI tooling has not increased productivity (and may in fact decrease it).”

Developer AI usage and output: the numbers side by side

Metric	Prior year	2026	Source
Developers saying AI writes 50%+ of their code	12%	42%	Global developer and IT leader survey
Developers spending majority of week writing new code	Not directly comparable	21%	Global developer and IT leader survey
Code production after AI agent adoption	Baseline	+30%	Organizational engineering data analysis
Commit volume after AI agent adoption	Baseline	+20%	Organizational engineering data analysis
Pull requests after AI agent adoption	Baseline	+23%	Organizational engineering data analysis
Resolved Jira issues and epics after adoption	Baseline	No significant change	Organizational engineering data analysis
METR trial: measured task completion time with AI	N/A	19% slower	METR randomized controlled trial
METR trial: developer expected speedup before trial	N/A	24% faster (expected)	METR randomized controlled trial

Where the major AI coding agents stand right now

Tool or model	Maker	Primary workflow focus	Notable 2026 detail
Gemini 4 Argon	Google	Long-running agent workflows, code migration, enterprise research	Global rollout via Gemini API and AI Studio, priced at $2 per million input tokens and $10 per million output tokens
Claude Code	Anthropic	Terminal-first, codebase-aware agentic coding	Supported by Scion, an internal source control system built after Anthropic’s monorepo merges rose from about 2,500 to over 30,000 per day
Cursor	Anysphere	Editor-native agents, including cloud-hosted background agents	Cloud agent workflows allow tasks to run while a developer works elsewhere
Muse Code	Meta	Competing directly with established coding agent vendors	Positioned against four established AI coding rivals
GitHub Copilot	Microsoft / GitHub	Default IDE integration across Microsoft’s developer tools	Benefits from bundling into existing Microsoft developer subscriptions

Five predictions for AI coding agents through 2027

Engineering organizations will shift performance metrics away from commit counts and pull request volume toward completed, merged, and production-stable work, since those are the figures the current data suggests are actually stalling.
More independent, randomized studies modeled on METR’s methodology will appear in 2027, as engineering leaders push back against vendor-supplied productivity claims that rely on self-reported developer sentiment.
Pricing competition between Google, Anthropic, Microsoft, and Meta will intensify around token costs for long-running agent sessions, since per-call economics matter more as agents take on longer, more autonomous tasks.
Code review tooling, rather than code generation tooling, will become the next major investment area for developer tool vendors, as the bottleneck shifts from writing code to verifying it.
Expect growing regulatory attention on agentic coding models, following the pattern set by scrutiny of Gemini 4 Argon, as agents capable of long autonomous runs raise new questions about oversight and accountability for the code they produce.

Frequently asked questions about AI coding agents in 2026

Do AI coding agents actually make developers slower?

According to METR’s randomized controlled trial, experienced open source developers completing real tasks with AI access took 19 percent longer than when working without it, despite expecting to be 24 percent faster. That finding applies specifically to experienced developers working in codebases they already knew well, so results may differ for less familiar codebases or less experienced developers.

Why do developers think AI makes them faster if the data shows otherwise?

METR found that developers continued to believe AI had sped them up by 20 percent even after the study was complete and the slower completion times were measured. Researchers pointed to the different nature of the work, writing code directly versus prompting, waiting, and reviewing, as a likely reason perception and measured time diverge.

What percentage of code is written by AI in 2026?

A global survey of 705 developers and IT leaders found that 42 percent now say AI writes at least half of their code, up from 12 percent a year earlier.

What is Gemini 4 Argon used for?

Gemini 4 Argon is Google’s model built for long-running agent workflows, including code migration projects, enterprise research, and multi-step cybersecurity tasks. It is available through the Gemini API, AI Studio, and the Gemini platform, priced at 2 dollars per million input tokens and 10 dollars per million output tokens.

Why did Anthropic build its own internal source control system?

Anthropic reportedly built a system called Scion to handle the surge in code changes generated by agent-heavy development. Daily merges in its internal monorepo reportedly rose from about 2,500 in May to more than 30,000 by September, a volume existing source control tooling was not originally designed to absorb.

Does more code output mean more finished software?

Not necessarily. Organizational data gathered after AI coding agent adoption found code production rose by about 30 percent and pull requests by about 23 percent, but resolved Jira issues and completed epics showed no statistically significant increase over the same period.

Should engineering teams stop using AI coding agents based on these findings?

The data does not point to abandoning the tools. It points to measuring their effect more carefully. METR itself framed its conclusion cautiously, stating it seems likely that for some important settings, recent AI tooling has not increased productivity and may in fact decrease it, which is a call for better measurement rather than a blanket verdict against the technology.

AI Coding Agents: METR Study Finds Devs 19% Slower [2026]