Sometimes—but there is no reliable, universal speedup. A controlled GitHub exercise and three company field experiments found gains on their respective measures. In contrast, a 2025 randomized trial found that experienced developers took longer with early-2025 AI tools on real issues in codebases they knew well. Those findings are not direct contradictions: they measured different kinds of work, with different developers, tools, and outcomes.
What does “faster” mean in these studies?
A developer can finish one specified task sooner without completing more work across a month; a team can complete more tasks without reducing the time spent on each one. Neither result, by itself, establishes better code or lower total project cost. The studies use several distinct measures:
As an Amazon Associate I earn from qualifying purchases.
- Elapsed completion time: how long it took a participant to finish a bounded task or real issue.
- Task throughput: how many tasks were completed over a period.
- Self-report: whether developers felt more productive, focused, or satisfied.
- Deployment evidence: what happened when an organization made an assistant available in day-to-day work, which can include surveys and usage data without isolating a causal effect.
These outcomes answer related, but not interchangeable, questions. A percentage increase in completed tasks is not the same as the same percentage reduction in time per task.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
What the studies found
| Study | Setting and participants | Reported result | What the result measures |
|---|---|---|---|
| GitHub Copilot controlled experiment (2022) | 95 professional developers randomly assigned to Copilot access or no Copilot for a JavaScript HTTP-server exercise. | Copilot users averaged 1 hour 11 minutes; the comparison group averaged 2 hours 41 minutes. GitHub reported the Copilot group was 55% faster (p=.0017; 95% confidence interval for speed gain: 21% to 89%). Completion rates were 78% and 70%, respectively. | Time and completion on one bounded coding task—not a general estimate for software projects. |
| Microsoft Research pooled field experiments (2025) | Three randomized experiments at Microsoft, Accenture, and an anonymous Fortune 100 company, covering 4,867 developers. | The combined estimate was a 26.08% increase in completed tasks (standard error 10.3%). The authors describe individual experiments as noisy and report higher adoption and greater gains among less experienced developers. | Task throughput across field experiments—not a 26.08% reduction in time per task. |
| METR randomized trial (2025) | 16 experienced developers completed 246 real issues in mature open-source projects they had contributed to for an average of five years. The tools were available during February–June 2025; participants primarily used Cursor Pro with Claude 3.5 or 3.7 Sonnet. | Allowing AI increased measured completion time by 19% in this study. | Elapsed time on real issues in familiar, complex repositories for this particular group and generation of tools. |
| METR later experiment update (February 2026) | A later experiment with returning and newly recruited participants; METR warned that participant selection made the result unreliable as an estimate of current productivity effects. | Reported speedup estimates were -18% for returning participants (95% confidence interval: -38% to +9%) and -4% for newly recruited participants (95% confidence interval: -15% to +9%). Both intervals include no effect. | Uncertain estimates from an experiment METR says may be biased downward because more developers declined to participate if they had to work without AI. |
Why did the results point in different directions?
The studies were not replications of one another. Their differences matter when deciding how much any result says about your own work:
#1 Best Overall
- Task realism and scope: GitHub measured a specified JavaScript exercise. METR measured work on real issues in mature repositories. The Microsoft experiments measured task completion in company settings.
- Codebase familiarity: METR participants were experienced contributors working in projects they already knew. That is a particularly relevant setting for senior maintainers, but it is not representative of every developer or assignment.
- Participants and experience: GitHub recruited professional developers; METR recruited a small group of experienced open-source contributors; Microsoft Research pooled thousands of developers across three organizations. Microsoft Research reported larger gains among less experienced developers, but that does not establish that AI invariably helps beginners more.
- Tools and timing: The GitHub result dates to 2022. METR tested tools available in February–June 2025, primarily Cursor Pro with Claude 3.5/3.7 Sonnet. Results from one period should not be treated as a permanent rating of later systems.
- Measurement and workflow: A timed task, completed tasks over a period, and a survey response capture different parts of development. Integration into ordinary workflows and the effort needed to review generated code can also matter, but these studies do not establish either factor as the cause of the differing results.
Taken together, the findings support a conditional answer—not one percentage that can be applied to every developer, task, or tool.
Can developers feel faster without finishing faster?
Yes. In METR’s 2025 trial, participants forecast a 24% time reduction before starting. After doing the tasks, they estimated that AI had reduced their time by 20%, even though measured completion time increased by 19%. Their forecasts and retrospective estimates are perceptions, not observed speedups.
GitHub also surveyed more than 2,000 developers about their experience with Copilot. It reported that 60–75% agreed with statements about greater fulfillment, less frustration, and more focus; 73% said Copilot helped them stay in flow, and 87% said it preserved mental effort during repetitive tasks. Those responses help describe how developers experienced the tool. They do not show that all respondents completed work faster.
Crashes, No Sound, or Screen Glitches?
Random freezes, missing sound and display glitches usually trace back to one bad driver. Find and replace yours safely.Free scan · under a minuteWindows Errors? Fix Them Before They Spread
Repair common Windows errors and clear accumulated junk for a smoother, more stable PC - no reinstall needed.Free scan · no reinstallWhat does the public-sector trial establish?
The UK Government Digital Service ran a three-month AI coding assistant trial from November 2024 to February 2025. It distributed 2,500 licenses across more than 50 public-sector organizations, with 1,900 licenses assigned. The report’s main analysis used 424 survey responses from 31 departments; 73% of those respondents had at least five years of coding experience.
This is useful deployment evidence about public-sector use, combining survey responses and telemetry. It is not a clean randomized estimate of how much faster the assistants made developers: the report describes little public-sector-specific research, and the figures above do not establish a causal time saving.
What should a developer or team conclude?
Use the studies as reasons to measure impact in your own workflow, not as a promise of a fixed return. If you are evaluating an assistant, decide in advance what “faster” means for the work you care about and compare like with like:
Rank #4
- Choose a relevant outcome. For a repeated task, record elapsed time and completion. For a team, track completed work over a defined period. Do not substitute perceived productivity for either measure.
- Compare similar work. Account for task type, repository familiarity, developer experience, and tool version; a short new exercise is not equivalent to maintaining a large codebase.
- Include the whole workflow. Generated code still needs review and integration. These studies do not settle end-to-end quality, maintenance costs, or organizational outcomes, so faster initial code production alone is not a complete productivity result.
- Interpret uncertain results cautiously. METR’s February 2026 update says its later experiment was not a reliable estimate of current productivity effects. More developers declined to participate if required to work without AI, which METR says likely biased estimated speedup downward; it also notes that the true speedup could be higher among developers and tasks that did not enter the experiment.
The evidence does not support a universal speedup
AI coding tools can help on some tasks and in some workflows, but the best available figures here describe particular settings, not a general law. GitHub’s 55% result belongs to a controlled JavaScript exercise, Microsoft’s 26.08% to pooled task throughput in three field experiments, and METR’s 19% slowdown to a small trial of experienced contributors on real repository issues with early-2025 tools. None settles what a different team will achieve—or whether faster code production improves quality and long-term project outcomes.
Quick Recap
Best Value
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
