AI Coding Assistant Statistics 2026 — “AI Always Makes Developers Faster” Is Not What the Data Says
AI coding assistants are now standard infrastructure for software teams, but the productivity story is more conditional than the adoption curve suggests. GitHub Copilot has crossed 26 million users, Claude Code is a multibillion-dollar product, and Cursor is embedded across major engineering organizations—yet controlled field research has also found experienced developers becoming slower on familiar codebases.
This report separates adoption from outcomes, comparing Cursor statistics, GitHub Copilot statistics, Claude Code statistics, pricing, enterprise deployment, developer trust, agent usage and the organizational conditions that determine whether AI-assisted software development actually pays off.
Public GitHub Copilot user count disclosed by Microsoft; this is not a paid-user or MAU figure.
Paid GitHub Copilot subscribers in Microsoft’s January 2026 earnings call, up 75% year over year.
Claude Code run-rate revenue disclosed by Anthropic in February 2026; more than doubled since January 1.
Cursor annualized revenue disclosed in November 2025, alongside “millions” of developers.
Technology professionals in DORA’s 2025 study who reported using AI at work.
Experienced open-source developers in METR’s field RCT took longer with early-2025 AI tools.
// Table of Contents
The Productivity Headline Is Real—and Still Incomplete
The most repeated AI coding claim is that assistants make developers roughly 55% faster. That result is real: GitHub’s controlled experiment found a 55.8% speed improvement on a JavaScript HTTP-server task. But treating it as a universal productivity multiplier ignores newer field evidence, different developer populations and downstream verification costs.
The 2026 picture is better described as a range of outcomes. DORA finds near-universal workplace adoption and strong perceived productivity gains. Cursor’s organizational study reports materially higher pull-request throughput. At the same time, METR’s randomized field study found experienced maintainers working in codebases they knew well were 19% slower when AI was available. The question is no longer “does AI coding work?” It is “for which developer, on which task, inside which system?”
GitHub Copilot users completed the controlled coding task faster than the non-Copilot group.
Average completion time for developers using Copilot in GitHub’s experiment.
Average completion time for the control group without Copilot.
Task completion rate for the Copilot group.
Task completion rate for the non-Copilot group.
Lower bound of GitHub’s 95% confidence interval for the speed gain.
Upper bound of the same 95% confidence interval.
Experienced open-source developers in METR’s randomized field study.
Real tasks completed in mature projects in the METR study.
Average prior experience participants had with the repositories they worked on.
Before the study, developers predicted AI would reduce task time by 24%.
After the study, developers still estimated AI had sped them up by 20%.
Perceived speed and measured speed can diverge. METR participants believed AI had made them about 20% faster after using it, while the randomized measurement showed completion time increased 19%.
This does not “disprove” GitHub’s 55.8% finding. The GitHub study used a bounded greenfield task; METR used experienced maintainers doing authentic work in mature repositories. Those contexts impose very different costs for context gathering, checking suggestions, reverting mistakes and maintaining architectural consistency.
DORA respondents who believed AI increased their productivity.
DORA respondents reporting little or no trust in AI-generated code.
Stack Overflow respondents who said AI tools or agents had a positive productivity effect.
Developers frustrated by AI solutions that are “almost right, but not quite.”
Developers saying debugging AI-generated code is more time-consuming.
For operators, the takeaway is straightforward: use AI coding statistics as a portfolio of signals, not a single headline multiplier. Throughput, cycle time, defect rate, review burden, developer satisfaction and spend can all move in different directions at the same time.
AI Coding Adoption Has Shifted From Optional Tool to Default Workflow
The adoption curve is steepest on GitHub, where Copilot moved from a paid add-on to a default expectation for new developers. Microsoft’s disclosures show the installed base continuing to expand while paid adoption deepens.
| Milestone | Metric | What it means | Source |
|---|---|---|---|
| Apr. 2025 | 15M+ Copilot users | More than 4× year-over-year. | Microsoft · FY25 Q3 Earnings Call · Apr. 2025 |
| Jul. 2025 | 20M Copilot users | Five million additional all-time users in roughly three months. | Microsoft · FY25 Q4 Earnings Call · Jul. 2025 |
| Oct. 2025 | 26M+ Copilot users | Microsoft called Copilot the most popular AI pair programmer. | Microsoft · FY26 Q1 Earnings Call · Oct. 2025 |
| Jan. 2026 | 4.7M+ paid subscribers | Paid subs up 75% year over year. | Microsoft · FY26 Q2 Earnings Call · Jan. 2026 |
| Apr. 2026 | ~140K organizations | Enterprise subscriber count nearly tripled year over year. | Microsoft · FY26 Q3 Earnings Call · Apr. 2026 |
The fastest structural change is at the start of the developer journey. GitHub reported that nearly four in five new developers used Copilot within their first week in 2025. That matters more than a one-time user count: it means AI assistance is being normalized before newcomers have built a long pre-AI workflow to compare against.
Share of new GitHub developers using Copilot within their first week.
Year-over-year increase represented by those new GitHub developers.
Pull requests created by GitHub Copilot coding agent between May and September 2025.
Accepted GitHub Copilot code suggestions reported by GitHub by October 2025.
Pull requests Copilot was contributing each month by October 2025.
Copilot’s most important 2026 statistic may be the 80% first-week adoption rate among new GitHub developers. That is a habit-formation signal, not merely an installed-base figure.
Claude Code shows a different kind of growth: monetization and depth of use. Anthropic says it reached a $1 billion run-rate only six months after public availability, exceeded $2.5 billion by February 2026, doubled weekly active users from January 1, quadrupled business subscriptions over that period and shifted more than half of Claude Code revenue to enterprise use.
Claude Code run-rate revenue reached in November 2025, six months after public availability.
Time Anthropic says Claude Code took to grow from public launch to a $1B run-rate.
Weekly active Claude Code users doubled from January 1 to Anthropic’s February 2026 disclosure.
Business subscriptions to Claude Code quadrupled from the start of 2026.
Enterprise use represented more than half of Claude Code revenue by February 2026.
External analysis cited by Anthropic estimated Claude Code authored 4% of public GitHub commits.
That estimated public-commit share had doubled in one month.
Why Developers Keep Adopting Coding Assistants Despite Trust Problems
The adoption drivers are not limited to faster code generation. The strongest data points cluster around reduced friction, agentic delegation, codebase context, learning and the ability to move work outside the editor.
Stack Overflow respondents using or planning to use AI tools in their development process.
Professional developers using AI tools daily in Stack Overflow’s 2025 survey.
Agent users agreeing agents reduced time spent on specific development tasks.
Agent users agreeing agents increased their productivity.
Increase in merged-PR rate after Cursor’s Agent became default, relative to baseline time trends.
Average active-running time Claude Code users spend with the tool, per Anthropic’s 2026 research.
The root cause is workflow compression. Older copilots reduced keystrokes. The newer generation reads codebases, writes and runs tests, opens pull requests, searches documentation, edits across files and works asynchronously. That changes the unit of delegation from “complete this line” to “take this task.”
Share of conversation-starting Cursor Agent requests in a 1,000-user sample that asked for implementation.
Increase in Agent acceptance rate per standard-deviation increase in developer experience.
Average Claude Code actions triggered by each user prompt in Anthropic’s session data.
Typical number of user-agent turns in a Claude Code session.
Average Claude output per turn in the studied sessions.
Sessions averaging more than 100 Claude actions per prompt.
Sessions averaging more than 200 actions per prompt.
Sessions averaging more than 500 actions per prompt.
Autocomplete made code cheaper to type. Agents make execution loops cheaper to delegate. That is why time spent planning, scoping and verifying is becoming more important than raw typing speed.
Claude Code’s division-of-labor data makes that shift measurable: humans make most planning decisions, while Claude makes most execution decisions. The tool is therefore not eliminating human judgment so much as moving it earlier in the workflow.
Planning decisions made by people in a typical Claude Code session.
Execution decisions made by people in a typical session; Claude makes the other ~80%.
Actions per prompt in typical novice-rated Claude Code sessions.
Actions per prompt in typical expert-rated sessions.
Approximate output words per prompt in novice-rated sessions.
Approximate output words per prompt in expert-rated sessions.
Cursor, Copilot and Claude Code Are Converging From Different Starting Points
Copilot began as editor-native autocomplete attached to the world’s largest code host. Cursor began as an AI-native IDE. Claude Code began as a terminal agent backed by a frontier model provider. By 2026 all three are expanding into agents, code review, cloud execution and enterprise controls—but their adoption signatures remain different.
| Platform | Public scale signal | Commercial signal | Enterprise signal | Source |
|---|---|---|---|---|
| GitHub Copilot | 26M+ users | 4.7M+ paid subscribers | ~140K organizations | Microsoft · AI in Action · 2026 / Microsoft · FY26 Q2 Earnings Call · Jan. 2026 / Microsoft · FY26 Q3 Earnings Call · Apr. 2026 |
| Cursor | Millions of developers | $1B+ annualized revenue | 50%+ of Fortune 500 by Jun. 2025 | Cursor · Series D / Past, Present, and Future · Nov. 2025 / TechCrunch · Cursor Series C / ARR · Jun. 2025 |
| Claude Code | 20 hrs/week average active-running use | $2.5B+ run-rate revenue | 50%+ of revenue from enterprise | Anthropic Research · Agentic Coding & Expertise · Jun. 2026 / Anthropic · Series G / Claude Code Milestones · Feb. 2026 |
The comparison is intentionally asymmetric because the companies disclose different denominators. Copilot publishes broad user and organization counts. Cursor’s clearest company-level figures are revenue and enterprise penetration. Anthropic publishes unusually detailed behavioral research and revenue mix. Converting those into a fake “market share” table would create precision that the source data does not support.
Fortune 500 companies using Cursor by the June 2025 Series C disclosure.
Cursor team size disclosed in November 2025.
Fortune 100 companies using GitHub Copilot in Microsoft’s public AI metrics.
GitHub Copilot enterprise subscribers nearly tripled year over year by April 2026.
Quarter-over-quarter growth in individual Copilot Pro+ subscriptions in FY26 Q2.
Siemens developers in a Copilot rollout cited by Microsoft.
Enterprise outcomes also differ by environment. Cursor’s 2026 customer studies report high adoption at Dropbox, Box, Salesforce, NVIDIA and Coinbase, while GitHub’s advantage is its existing repository, pull-request and security surface. Claude Code’s advantage is depth of model-agent integration and a usage pattern that increasingly extends beyond pure software occupations.
Cursor-generated lines accepted by Dropbox engineers each month.
Files indexed in Dropbox’s monorepo for Cursor.
Coinbase developers using Cursor in regular workflows.
The AI Coding Market Is Moving From Seat Pricing to Consumption Economics
The old pricing model was simple: one seat, one monthly fee. Agentic coding makes cost more variable because a single task can trigger long model runs, tool calls and multiple agents. All three major platforms now expose some version of usage limits, credits or tiered capacity.
| Plan | Price | Included usage signal | Source |
|---|---|---|---|
| GitHub Copilot Free | $0 | 2,000 completions/mo | GitHub · Copilot Plans & Pricing · Aug. 2026 |
| GitHub Copilot Pro | $10/mo | 1,000 base AI credits | GitHub Docs · Copilot Plans · Aug. 2026 |
| GitHub Copilot Pro+ | $39/mo | 3,900 base credits | GitHub Docs · Copilot Plans · Aug. 2026 |
| GitHub Copilot Max | $100/mo | 10,000 base credits | GitHub Docs · Copilot Plans · Aug. 2026 |
| GitHub Copilot Business | $19/seat/mo | 1,900 credits/user/mo | GitHub Docs · Copilot Plans · Aug. 2026 |
| GitHub Copilot Enterprise | $39/seat/mo | 3,900 credits/user/mo | GitHub Docs · Copilot Plans · Aug. 2026 |
| Cursor Pro | $20/mo | Individual agent plan | Cursor Docs · Pricing and Plans · Aug. 2026 |
| Cursor Pro+ | $60/mo | Higher agent usage | Cursor Docs · Pricing and Plans · Aug. 2026 |
| Cursor Ultra | $200/mo | Power-user tier | Cursor Docs · Pricing and Plans · Aug. 2026 |
| Cursor Teams Standard | $40/user/mo | Admin + team controls | Cursor Docs · Pricing and Plans · Aug. 2026 |
| Cursor Teams Premium | $120/user/mo | Higher team usage | Cursor Docs · Pricing and Plans · Aug. 2026 |
| Claude Pro | $20/mo | Includes Claude Code | Anthropic · Claude Plans & Pricing · Aug. 2026 |
| Claude Max | from $100/mo | 5× or 20× Pro usage | Anthropic · Claude Plans & Pricing · Aug. 2026 |
| Claude Team Standard | $25/seat/mo | $20 annual-billing equivalent | Anthropic · Claude Plans & Pricing · Aug. 2026 |
| Claude Team Premium | $125/seat/mo | $100 annual-billing equivalent; 5× Standard usage | Anthropic · Claude Plans & Pricing · Aug. 2026 |
The pricing spread is now wide enough that “we pay $20 per developer” is no longer a credible total-cost assumption for agent-heavy teams. GitHub’s newer credit model explicitly meters chat, agents and CLI usage, while completions remain unlimited on paid plans. Cursor uses included model pools plus on-demand usage. Claude’s Max and Premium tiers similarly price for sustained agent workloads.
Claude Code crossed $2.5B in run-rate revenue, Cursor crossed $1B annualized revenue, and Copilot had 4.7M+ paid subscribers by January 2026. AI coding is no longer a feature category; it is a standalone software market with enterprise-scale budgets.
Cursor Series D financing announced in November 2025.
Post-money valuation attached to Cursor’s Series D.
Cursor Series C financing announced in June 2025.
Cursor ARR reported in June 2025 before crossing $1B later that year.
Claude Code’s run-rate revenue more than doubled from January 1 to February 2026.
Anthropic doubled Claude Code’s five-hour rate limits for Pro, Max, Team and seat-based Enterprise plans in May 2026.
GitHub value assigned to one AI credit under its 2026 usage-based billing model.
Coding Assistants Are Becoming Coding Agents
The adjacent-technology story is autonomy. AI coding tools increasingly run commands, inspect repositories, open pull requests and work asynchronously. Anthropic’s Claude Code dataset is the richest current view of how that changes the composition of work.
Interactive Claude Code sessions analyzed by Anthropic between October 2025 and April 2026.
People represented in the privacy-preserving Claude Code analysis.
Claude Code sessions primarily focused on building or writing code.
Sessions primarily focused on fixing broken code.
Sessions primarily focused on testing or orchestrating code/agents.
Sessions focused on operating software: deploy, configure, run or monitor.
Sessions focused on planning or understanding systems.
Sessions producing analysis or prose where code was incidental.
That composition is already shifting. Between October 2025 and April 2026, fixing broken code fell sharply while operating software and analysis-oriented work grew. The agent is moving upstream and downstream from code generation into the rest of the software lifecycle.
Share of Claude Code sessions spent fixing broken code in October 2025.
Share spent fixing broken code in April 2026.
Operating-software share in October 2025.
Operating-software share in April 2026.
Combined writing/data-analysis share at the start of the period.
Combined writing/data-analysis share by April 2026.
Increase in Anthropic’s estimated average session task value from October to April.
Estimated value increase for building-type Claude Code tasks.
Estimated value increase for operating-type tasks.
Estimated value increase for fixing-type tasks.
Claude Code’s debugging share fell from 33% to 19% in seven months while operating software rose from 14% to 21%. The frontier is shifting from “write code for me” toward “run more of the workflow for me.”
GitHub is making the same architectural bet with Coding Agent and Agent HQ; Cursor is doing it with Cloud Agents, Slack-triggered work and agentic code review. The competitive moat is therefore increasingly the harness: context management, tool permissions, observability, review UX and the ability to coordinate multiple models—not just which model produces the next token.
“Cursor offers the best harness, and the platform has the smoothest developer experience of anything I’ve tried.”Joe Haddad, Distinguished Engineer at Vercel · Cursor customer case study, July 2026
Adoption Is High. Trust Is Not.
The human data is the core constraint on agentic coding. Developers increasingly use AI every day, yet trust, verification burden and task complexity still determine where they are willing to delegate.
Stack Overflow respondents using or planning to use AI tools in development.
Comparable adoption/planned-adoption figure cited for the prior year.
Positive sentiment reported in both 2023 and 2024 before declining in 2025.
Professional developers with a favorable stance toward AI tools.
Professional developers saying AI tools struggle with complex tasks in 2025.
Comparable professional-developer figure cited for 2024.
Developers are using AI more while trusting it less. Stack Overflow found 46% distrust accuracy versus 33% who trust it, and 66% say “almost right” answers are their biggest frustration.
Resistance rises with responsibility. Seventy-six percent of developers said they do not plan to use AI for deployment and monitoring, while 69% said the same for project planning. Those are system-level activities where an error can have wider consequences than a local code edit.
Developers who do not plan to use AI for deployment and monitoring.
Developers who do not plan to use AI for project planning.
Top reason developers would still ask a person for help: not trusting AI answers.
Respondents saying vibe coding is not part of their professional workflow.
Additional respondents emphatically rejecting vibe coding as part of their work.
Developers who either do not use agents or stick to simpler AI tools.
Software developers using agents at work who use them for software development.
Respondents concerned about agent security and data privacy.
Agent users saying agents improved team collaboration—the lowest-rated impact in the survey.
Expertise appears to improve outcomes rather than become irrelevant. Anthropic’s Claude Code study found expert users elicited more actions per prompt, recovered from trouble more successfully and achieved stricter verified-success criteria more often. Cursor’s research similarly found senior developers more likely to accept agent-written code than junior developers.
Verified-success rate for novice-rated Claude Code sessions.
Verified-success range for intermediate-or-higher sessions.
Novice sessions reaching at least partial success.
Intermediate-or-higher sessions reaching at least partial success.
Troubled novice sessions that still reached verified success.
Troubled expert sessions that recovered to verified success.
Troubled novice sessions abandoned with zero lines written.
Comparable abandonment rate for everyone above novice level.
AI-Assisted Development Is Expanding With the Global Developer Base
Public tool vendors do not release comparable country-level active-user tables for Cursor, Copilot and Claude Code. GitHub’s Octoverse data is therefore the most defensible geographic proxy for where the next wave of AI-assisted software production is likely to grow.
United States
~12.8M generative-AI repository contributions, representing 31.8% of the total in GitHub’s 2025 dataset.
India
~5M generative-AI repository contributions, 12.5% of the total; India led in distinct AI repositories.
APAC
13M net new GitHub developers from 2024 to 2025, led by India, Japan and Indonesia.
Europe
6.3M net new GitHub developers from 2024 to 2025.
LATAM
3.2M net new GitHub developers from 2024 to 2025.
Africa & Middle East
3.4M net new GitHub developers from 2024 to 2025.
The raw developer populations reinforce the same direction. India added more than five million GitHub developers in 2025 and is projected by GitHub’s models to become the largest developer community by 2030. The implication for coding-assistant vendors is obvious: local pricing, payment methods, latency, education and language support will increasingly shape growth outside North America.
Share of GitHub’s new 2025 developers accounted for by India.
Distinct generative-AI repositories associated with U.S. contributors in GitHub’s comparison.
Distinct generative-AI repositories associated with Indian contributors.
APAC added 13M GitHub developers in one year—more than Europe, LATAM, and Africa/Middle East combined. Tool adoption strategies that assume a U.S.-centric developer market will age badly.
Anthropic’s research also shows coding agents spreading beyond software engineering. In a 2026 survey of quantitative social scientists, 20% were already using coding agents at least weekly; Claude Code was the dominant agent among adopters. That is an early signal that the addressable market is expanding from “developers” toward anyone whose work can be expressed as data, files and executable workflows.
Quantitative social scientists in Anthropic’s early-2026 coding-agent survey.
Researchers who had tried generative AI in the research process.
Researchers regularly using coding agents more than once a week.
Coding-agent users in the survey reporting Claude Code use.
Coding-agent users reporting Codex use, the second-most common named tool.
Economists regularly using coding agents in the survey.
Political scientists regularly using coding agents.
Public-health researchers regularly using coding agents.
Education researchers regularly using coding agents.
Communications researchers regularly using coding agents.
The Highest-ROI Teams Change the System Around the Tool
DORA’s 2025 conclusion is deliberately unglamorous: AI is an amplifier. Strong platforms, fast feedback loops and clear workflows compound the benefit; weak systems simply generate more output into the same bottlenecks. The 2026 enterprise case studies support that interpretation.
Coinbase is the most extreme example because it redesigned planning and delegation rather than dropping an assistant into an unchanged process. Some teams reduced idea-to-production time from 20 days to under two days; the company says engineers are running multiple agents in parallel and smaller teams are taking on work that previously required larger groups.
Idea-to-production time reported for some Coinbase teams after the agent-first redesign.
Idea-to-first-PR time reported by Coinbase.
Manual coding time saved per Coinbase developer.
Asynchronous agents many Coinbase engineers run in parallel.
Small teams now taking on features that previously required full teams.
Decrease in resources allocated per feature at Vercel.
Time saved per Vercel developer annually.
Reduction in migration effort reported by Box.
Increase in Cursor usage after Box’s six-week mentorship program.
Increase in Cursor power users during that program.
Time Money Forward reports saving per engineer.
Reduction in QA test-generation time at Money Forward.
Increase in committed code for NVIDIA developers using Cursor, with bug rates reported flat.
IMDEX-reported pace improvement on a major legacy migration versus its pre-AI model.
Time IMDEX reports taking to consolidate two flagship platforms.
Pre-AI estimate for the same consolidation, according to IMDEX’s case study.
The biggest reported gains occur when teams redesign planning, testing, review, context and parallelism—not when they measure “lines of AI code.” Coinbase, Vercel and Box all changed workflow architecture alongside tool adoption.
The caution is source quality. Customer case studies are useful operating evidence, but they are vendor-published and usually feature successful deployments. They should guide hypotheses, not substitute for an internal baseline. The safer practice is to run controlled rollouts and watch delivery speed and quality together.
Greater likelihood of passing all 10 unit tests in GitHub’s Copilot code-quality RCT.
More lines written per readability error with Copilot in the blind-review phase.
Readability improvement in GitHub’s quality scoring.
Reliability improvement in the same study.
Maintainability improvement in the same study.
Conciseness improvement in the same study.
Higher likelihood reviewers approved Copilot-authored code.
Experienced developers recruited into the GitHub code-quality RCT.
Developers reporting greater confidence in code quality with Copilot and Copilot Chat.
Faster code reviews in GitHub’s Copilot Chat research.
Developers reporting they maintained flow state with Copilot Chat.
Three Charts That Explain the 2026 AI Coding Market
These are the comparisons where charts add context without pretending incompatible metrics are the same: Copilot’s disclosed user milestones, Claude Code’s session composition, and the developer sentiment gap around AI accuracy and productivity.
GitHub Copilot disclosed user milestones
Public “users” count; not monthly active users. Paid subscriber count is shown separately in the article.
Claude Code session mix
Anthropic privacy-preserving analysis of ~400K interactive sessions, Oct. 2025–Apr. 2026.
Developer AI sentiment
Selected Stack Overflow 2025 survey signals; percentages answer different questions.
Chart note: sentiment bars are not a single 100% distribution. They summarize separate survey questions and should be interpreted independently.
A second procurement implication follows from the speed data: teams should expect the preferred tool to vary by repository and role. A greenfield web team may optimize for rapid generation, while a platform team may care more about repository context, review quality and safe tool execution. Data-science and research users may value terminal autonomy over IDE polish. Standardizing governance, logging and evaluation criteria can therefore create more leverage than forcing every developer onto the same assistant. The shared control plane should be consistent even when the coding interface is not.
10 Evidence-Based Moves for Engineering Leaders in 2026
The winning strategy is not “pick the model with the highest benchmark.” It is to match tools to work, redesign the development system, preserve verification, and measure value in delivery outcomes.
METR found a 19% slowdown on familiar mature repositories while GitHub measured a 55.8% speedup on a bounded task. Your workload determines which result is closer. Evidence: S006 Evidence: S007
Copilot’s public installed scale is 26M+ users, but Microsoft separately reported 4.7M+ paid subscribers. Track the denominator that maps to spend. Evidence: S001 Evidence: S002
Cursor’s organizational study found a 39% increase in merged PRs, while Vercel reports a 54% PR-throughput gain and an 89% cycle-time reduction for power users. Evidence: S047 Evidence: S198
Claude Code users make ~70% of planning decisions while Claude makes ~80% of execution decisions; experts elicit 12 actions per prompt versus five for novices. Evidence: S057 Evidence: S060
Copilot now meters AI credits, Cursor has $60 and $200 individual tiers, and Claude Max begins at $100/month. Agent-heavy users create variable compute cost. Evidence: S088 Evidence: S099 Evidence: S103
46% of Stack Overflow respondents distrust AI accuracy, 66% are frustrated by almost-correct solutions, and 81% worry about agent security/privacy. Evidence: S138 Evidence: S022 Evidence: S154
Coinbase reports 75% of PRs created by agents and idea-to-first-PR falling from eight days to under 30 minutes after process redesign. Evidence: S188 Evidence: S194
Anthropic found verified success rising from 15% in novice sessions to 28–33% for intermediate-or-higher sessions; experienced Cursor users also accepted agent edits more often. Evidence: S156 Evidence: S050
GitHub’s RCT found higher unit-test success and modest gains in readability, reliability and maintainability, showing quality can be measured alongside throughput. Evidence: S211 Evidence: S215
APAC added 13M GitHub developers in one year, while 20% of surveyed quantitative social scientists already use coding agents weekly. Evidence: S166 Evidence: S180
None of these actions requires committing to one vendor across the entire organization. In 2026, heterogeneous stacks are normal: an engineer may use Copilot for repository-native completion and review, Cursor for IDE-centered agent work, and Claude Code for terminal-heavy or long-horizon tasks. The operating discipline is to define where each tool is allowed to act, what evidence is required before merge or deployment, and which business metric determines renewal. That keeps experimentation fast without turning AI adoption into an unmeasured entitlement.
How This Report Was Built
Data is current as of August 26, 2026, and the article uses US English to match SEOScaleUp’s existing statistics pages. We prioritized primary company disclosures and research, then major survey/research organizations and reputable financial/technology reporting where a company did not publish the figure itself. Older 2022–2024 GitHub studies are included only because they are landmark controlled experiments frequently cited in the AI-coding productivity debate; their original dates are stated so they are not mistaken for 2026 measurements.
Sources used: Microsoft FY25 Q3 and Q4 earnings calls; Microsoft FY26 Q1, Q2 and Q3 earnings calls; Microsoft AI in Action (2026); GitHub Octoverse 2025; GitHub Copilot: Faster, Smarter, and Built for How You Work Now (2025); GitHub Copilot productivity experiment (2022); GitHub Copilot Chat quality research (2023); GitHub Copilot code-quality randomized controlled trial (2024, updated 2025); GitHub Copilot plans and pricing documentation (2026); Cursor Series C disclosure and TechCrunch Series C/ARR coverage (2025); Cursor Series D disclosure (2025); Cursor Productivity Impact of Coding Agents (2025); Cursor pricing documentation (2026); Cursor 2026 customer case studies for Coinbase, Vercel, NVIDIA, Dropbox, Box, Salesforce, Money Forward and IMDEX; Anthropic Claude Code $1B milestone / Labs announcement (2025–2026); Anthropic Series G / Claude Code milestones (2026); Anthropic Agentic Coding and Persistent Returns to Expertise (2026); Anthropic Coding Agents in the Social Sciences (2026); Anthropic Economic Index: Learning Curves (2026); Anthropic Higher Claude Code Limits (2026); Anthropic Claude plans and pricing (2026); Google Cloud / DORA State of AI-Assisted Software Development 2025 and DORA’s March 2026 follow-up; Stack Overflow Developer Survey 2025; and METR Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity (2025). Conflicting productivity figures are not averaged because they measure different tasks and populations. GitHub’s 55.8% speed gain is a controlled greenfield task; METR’s 19% slowdown is a randomized field experiment on experienced maintainers working in familiar repositories; DORA and Stack Overflow include self-reported outcomes; Cursor customer stories are vendor-published deployment case studies. Likewise, Copilot “users,” Copilot paid subscribers, Cursor annualized revenue and Claude Code run-rate revenue are kept as separate measurement bases rather than converted into unsupported market-share estimates.