Cursor vs Copilot Seat Audit: 12 Steps, 75 Min No verified, dated dataset measures Cursor against GitHub Copilot on seat share or active-developer share as of November 19, 2026, according to a Shattered audit of published figures. The audit shows Stack Overflow's 2025 Developer Survey numbers — 17.9% of respondents using Cursor as a development IDE and 67.9% of 8,323 respondents who answered that question having used GitHub Copilot — rest on different denominators, so placing 18% beside 68% does not establish a market lead. The audit instead offers a roughly 200-line standard-library Python project that joins a Copilot seat export, a Cursor member export and a team roster in 12 steps over about 75 minutes to report assigned share, active share, utilization, overlap, idle spend, orphan seats and cost per active developer. Stack Overflow’s 2025 developer survey put Cursor at 18% of developers and GitHub Copilot at 68% of developers who use out-of-the-box AI assistants, according to the survey’s press release https://stackoverflow.co/company/press/archive/stack-overflow-2025-developer-survey/ . Those two figures look like a market-share verdict for the Cursor vs Copilot fight. They are not one. The first counts everyone who answered, the second counts only a subset, and neither says a word about how many people on your team open the tool on a Tuesday. As of November 19, 2026, no verified, dated dataset measures Cursor against Copilot on seat share or active-developer share. The vendors publish growth claims, the survey publishes usage rates, and nobody publishes the one ratio a team lead needs at renewal time. A public-survey share audit https://shattered.io/cursor-vs-github-copilot-market-share-audit-12-steps-2026/ shows how to read the published numbers. The project below answers a different question: what share do the two tools hold inside your own organization, and what does each active developer cost? The build is a standard-library Python project about 200 lines that joins a Copilot seat export, a Cursor member export, and your roster. It reports assigned share, active share, utilization, overlap, idle spend, orphan seats, and cost per active developer. Twelve steps, about 75 minutes, no paid services, and no data leaves your machine. Don't miss new tech stories on Google Add Shattered once in the Google app and our stories show up in your news suggestions. Add now https://www.google.com/preferences/source?hl=en&q=https%3A%2F%2Fshattered.io%2F Why Published Cursor vs Copilot Share Numbers Cannot Settle Your Seat Decision Every public number in this market counts something different. That is the whole problem. A share figure only means something when numerator and denominator are both stated, and most headlines state neither. Take the Stack Overflow data. The 2025 survey reports that 17.9% of respondents used Cursor as a development IDE. It also reports that 67.9% of 8,323 respondents who answered the relevant question had used GitHub Copilot. The press materials then add a third framing: 68% of developers using out-of-the-box AI assistants selected Copilot. Stack Overflow wrote in its press release for that survey: “In the 2025 Stack Overflow Developer Survey, 18% of developers reported using Cursor, while 68% of developers using out-of-the-box AI assistants selected GitHub Copilot.” Read that sentence twice. Cursor’s 18% is a share of all developers. Copilot’s 68% is a share of one filtered group. Stack Overflow’s own September 30, 2026 retrospective on the 2025 findings https://stackoverflow.blog/2026/09/30/getting-ready-for-2026-results-a-look-back-on-developer-survey-findings repeats the 68% figure with the same filter. Putting 18% beside 68% and calling the gap a market lead mixes two denominators. For a method to compare survey rates fairly, the Stack Overflow survey walkthrough in pandas https://shattered.io/cursor-vs-copilot-stack-overflow-survey-pandas-12-steps-2026/ covers the weighting questions in detail. | Source | Reported figure | What the denominator is | What it cannot tell you | |---|---|---|---| | Stack Overflow 2025 survey | 17.9% used Cursor as a development IDE | Survey respondents | Seats, spend, or frequency of use | | Stack Overflow 2025 survey | 67.9% used GitHub Copilot | 8,323 respondents who answered that question | Whether they still use it today | | Stack Overflow 2025 press release | 68% chose Copilot | Developers using out-of-the-box AI assistants | Share among all developers | | GitHub, May 2026 | 140,000 organizations use Copilot | Organizations, not developers | Developers per organization | The vendors add their own units. GitHub said Copilot served 140,000 organizations, nearly triple the number a year earlier, with overall growth above 100% year over year, in a May 2026 post about its Gartner Magic Quadrant placement. That counts organizations. A ten-person startup and a 40,000-seat bank each add one. Cursor’s enterprise page claims use by 64% of Fortune 500 companies, which is a marketing claim about logos, not about developers. OpenAI reportedly counts more than 5 million weekly active Codex users, a third unit again weekly actives on a different product . Four sources, four denominators: respondents, filtered respondents, organizations, and Fortune 500 logos. None converts into the number that drives your budget. Sampling error also hides inside every survey percentage, and a team of 40 developers is a tiny sample on its own. Your own exports have none of these problems. They are a census, not a sample. You know the headcount, you know who holds a seat, and both vendors record when each seat was last used. Prerequisites and Versions for the Seat Audit The project needs almost nothing. It runs on a laptop, reads three CSV files, and writes a markdown report. Budget about 75 minutes if you already have admin access to both tools, and add a day of waiting if you need someone else to approve exports. | Requirement | Version or access | Why you need it | |---|---|---| | Python | 3.10 or newer | Standard library only, no pip installs | | Operating system | Linux, macOS, or Windows with WSL | Any shell that runs Python works | | GitHub access | Read-only token, org owner or billing manager rights | Lists Copilot seat assignments | | GitHub REST API version | Not listed | Pins the response format | | Cursor access | Team admin rights | Provides members and last-active dates | | Roster | CSV with email, GitHub login, team | Defines the headcount denominator | | Disk | Under 10 MB | CSV files and one report | Time budget by phase: Phase 1 define metrics takes 10 minutes, Phase 2 collect exports takes 25, Phase 3 join and count takes 20, and Phase 4 price, report, test takes 20. The remaining time goes to reading the output. Three rules before you start. First, keep the project folder outside any shared repository, because the CSV files contain employee identifiers. Second, ask HR or your privacy officer whether per-person activity data may be processed for license management. Third, use a read-only token. The audit never needs write access. You also need to decide what active means before you see any numbers. Writing that definition down first stops the common failure where a team picks the window that makes its favorite tool look best. Phase 1: Define the Metrics Before You Touch Data Seat audits go wrong at the definition stage far more often than in the code. These three steps take ten minutes and save an argument later. Step 1: Write down the seven numbers Every metric below uses your headcount as the common denominator, which is the fix for the survey problem above. Copy this table into your project notes. | Metric | Formula | What it answers | |---|---|---| | Assigned share | Matched seats / headcount | How widely licensed is each tool? | | Active share | Developers active in the window / headcount | How widely used is each tool? | | Utilization | Active developers / seats | How much of what you pay for gets used? | | Idle spend | Seats minus active x seat price | What does unused capacity cost per month? | | Cost per active developer | Seat spend + variable spend / active | What does one working user really cost? | | Overlap | Developers active on both tools | Are you paying twice for the same people? | | Orphan seats | Seats with no roster match | Who holds a seat but does not work here? | Assigned share and active share are the internal versions of the market-share numbers you cannot find publicly. Utilization and cost per active developer are the numbers that decide renewals. Step 2: Fix the as-of date and the activity window Pick one as-of date and one window length, then freeze both in a config file. This project uses November 19, 2026 and a 30-day window. A developer counts as active when the tool’s last-activity date falls on or after the cutoff, which is 30 days before the as-of date. The boundary is inclusive, and a test later proves it. A 30-day window suits most teams. A window that spans late-November holidays or a company shutdown understates activity for every tool equally, so compare the same window across tools and across months. Never compare a 30-day Copilot number with a 90-day Cursor number. Step 3: Create the project layout and config Make the folder and the config file. The prices in the config are placeholders, not list prices. Replace them with the per-seat amounts on your own invoices, after discounts and annual billing. mkdir -p seat-audit/data seat-audit/tests seat-audit/out cd seat-audit python3 --version printf 'data/\nout/\n' .gitignore The config maps each tool to its input file, the column that identifies a person, the column holding the last-active date, and an optional column for variable spend. Mapping columns in config instead of in code matters because export formats change. { "as of": "2026-11-19", "window days": 30, "min group size": 5, "roster": "data/roster.csv", "tools": { "copilot": { "file": "data/copilot seats.csv", "key": "github login", "roster key": "github login", "last active": "last activity at", "variable usd": "variable usd", "seat usd per month": 20.0 }, "cursor": { "file": "data/cursor members.csv", "key": "email", "roster key": "email", "last active": "last seen", "variable usd": "api spend usd", "seat usd per month": 30.0 } } } Variable spend deserves its own column because both products now meter usage beyond the seat. GitHub began moving Copilot usage to GitHub AI Credits on June 1, 2026, and its pricing page states that 1 AI credit equals $0.01. Pro, Pro+, and Max list monthly total credits of $15, $70, and $200 respectively. Cursor’s newer plans bill third-party models Claude, GPT, and Gemini at API list prices from prepaid credit on Premium and above, according to a November 14, 2026 pricing breakdown https://vdf.ai/blog/cursor-pricing/ . A seat price alone no longer predicts a bill, which is why the project tracks both parts. For the plan-by-plan view of limits, see the pricing and limits comparison of Claude Code, Cursor, and Copilot https://shattered.io/claude-code-vs-cursor-vs-github-copilot-pricing-limits-2026/ . Phase 2: Collect the Copilot and Cursor Exports Three CSV files feed the audit: Copilot seats, Cursor members, and the roster. Collect them in this order, because the roster decides which identifiers to look for in the other two. Step 4: Build the roster The roster is your headcount denominator, so it should list everyone who could plausibly hold a seat and nobody else. Export it from your identity provider or HR system with three columns: email , github login , and team . Exclude people on long leave only if you also exclude them from billing conversations. Keep contractors in the roster if they hold seats, tagged with their own team name. Two details trip people up. GitHub logins and emails are different identifiers, so the roster carries both. And the roster must be current. A stale roster turns every departed employee into an orphan seat and every new hire into a developer with no seat. If you want to dry-run the pipeline before touching real data, this generator writes a synthetic 57-person team with a fixed random seed, so your output matches the sample later in this article. python import csv import random from datetime import date, timedelta random.seed 7 AS OF = date 2026, 11, 19 TEAMS = {"platform": 14, "payments": 12, "mobile": 10, "data": 9, "growth": 8, "security": 4} roster, copilot, cursor = , , n = 0 for team, size in TEAMS.items : for in range size : n += 1 email = f"dev{n:02d}@example.com" login = f"dev{n:02d}-gh" roster.append {"email": email, "github login": login, "team": team} if random.random < 0.85: gap = random.choice 1, 2, 3, 5, 8, 12, 20, 29, 45, 70 copilot.append { "github login": login, "last activity at": AS OF - timedelta days=gap .isoformat + "T09:30:00Z", "variable usd": round random.choice 0, 0, 0, 4.5, 11.0, 27.5 , 2 , } if random.random < 0.55: gap = random.choice 0, 1, 2, 4, 6, 9, 15, 25, 33, 60 cursor.append { "email": email.upper if n % 7 == 0 else email, "last seen": AS OF - timedelta days=gap .isoformat , "api spend usd": round random.choice 0, 0, 8.25, 19.0, 64.4, 140.0 , 2 , } copilot.append {"github login": "former-contractor-gh", "last activity at": "2026-08-02T10:00:00Z", "variable usd": 0} cursor.append {"email": " email protected ", "last seen": "never", "api spend usd": 0} def dump path, rows : with open path, "w", newline="", encoding="utf-8" as fh: w = csv.DictWriter fh, fieldnames=list rows 0 w.writeheader w.writerows rows dump "data/roster.csv", roster dump "data/copilot seats.csv", copilot dump "data/cursor members.csv", cursor print len roster , len copilot , len cursor Step 5: Export Copilot seats with a read-only token GitHub’s REST API lists Copilot seat assignments per organization, and each seat record includes the assignee’s login and a last-activity timestamp. The script below paginates through the list and writes a CSV. It reads the token from an environment variable, never from a file, and it sends the pinned API version header. bash /usr/bin/env python3 """Export Copilot seat assignments for one org to CSV read-only .""" import csv import json import os import sys import urllib.error import urllib.request API = "https://api.github.com/orgs/{org}/copilot/billing/seats?per page=100&page={page}" def fetch page org, page, token : req = urllib.request.Request API.format org=org, page=page , headers={ "Authorization": f"Bearer {token}", "Accept": "application/vnd.github+json", "X-GitHub-Api-Version": "2022-11-28", }, try: with urllib.request.urlopen req, timeout=30 as resp: return json.load resp except urllib.error.HTTPError as err: sys.exit f"GitHub returned {err.code} for page {page}: {err.reason}" def main org, out path : token = os.environ.get "GITHUB TOKEN" if not token: sys.exit "Set GITHUB TOKEN to a read-only token first." rows, page = , 1 while True: seats = fetch page org, page, token .get "seats", if not seats: break for seat in seats: rows.append { "github login": seat "assignee" "login" .lower , "last activity at": seat.get "last activity at" or "never", "last activity editor": seat.get "last activity editor" or "", "plan type": seat.get "plan type" or "", "variable usd": 0, } page += 1 with open out path, "w", newline="", encoding="utf-8" as fh: w = csv.DictWriter fh, fieldnames=list rows 0 if rows else "github login" w.writeheader w.writerows rows print f"wrote {len rows } seats to {out path}" if name == " main ": if len sys.argv = 3: sys.exit "usage: fetch copilot.py ORG OUTPUT.csv" main sys.argv 1 , sys.argv 2 Run it with GITHUB TOKEN=... python3 fetch copilot.py your-org data/copilot seats.csv . GitHub’s documentation lists the token scopes the endpoint accepts, so check the current page and choose the narrowest read-only option. A seat that has never been used comes back with an empty activity field, which the script writes as never . Packaging is shifting here. According to GitHub’s plans page and documentation, enterprise packaging is moving toward a unified GitHub Enterprise plus AI-credit model rather than a separate Copilot subscription. If your licenses come through an enterprise account, the organization endpoint may not show every seat, so fall back to the billing export in your enterprise settings and map its columns in the config. Teams weighing Copilot against an agent-first rival instead of Cursor can adapt the same audit to the pairing in GitHub Copilot vs Claude Code https://shattered.io/github-copilot-vs-claude-code-2026/ . Step 6: Export Cursor members and map the columns Cursor’s team dashboard is the source for member data. If your plan exposes a member or usage export, download it. If it only shows a table, copy the table into a spreadsheet and save it as CSV. Column names differ by plan and change over time, so this project never hard-codes them. Open the file, find the column that holds each person’s email and the column that holds the last-active date, and put those names in config.json . Dates must arrive in ISO format 2026-11-18 . If the export uses another format, convert the column in your spreadsheet before saving. Add an API spend column if your export has one, or leave the variable usd mapping out and the audit treats variable spend as zero for that tool. Check the file’s first rows before moving on. A header with a leading blank, a date column full of relative text such as 2 days ago , or a members list that includes pending invitations will each distort the counts, and it is faster to fix them now than to explain them later. Phase 3: Join, Count, and Classify Developers The core logic is three functions. Each one is small enough to read in a minute, and each one does exactly one job. Step 7: Load the files and normalize identifiers The loaders read each CSV, lowercase every identifier, and parse dates. Lowercasing fixes the most common cause of false orphans: an export that writes \ email protected\ https://shattered.io/cdn-cgi/l/email-protection \ email protected\ https://shattered.io/cdn-cgi/l/email-protection utf-8-sig strips the invisible byte-order mark that Excel adds to CSV files, which otherwise corrupts the first column name. If a person appears twice in one export, the loader keeps the latest activity date. bash /usr/bin/env python3 """Seat audit: Cursor vs Copilot from your own exports.""" import argparse import csv import json from datetime import date, datetime, timedelta from pathlib import Path TOOLS = "copilot", "cursor" def parse day raw : raw = raw or "" .strip if not raw or raw.lower in {"never", "null", "none", "n/a"}: return None return datetime.fromisoformat raw.replace "Z", "+00:00" .date def credits to usd credits, usd per credit=0.01 : return round float credits usd per credit, 2 def load roster path : with open path, newline="", encoding="utf-8-sig" as fh: rows = list csv.DictReader fh for row in rows: row "team" = row.get "team" or "" .strip or "unassigned" return rows def load seats path, spec : seats = {} with open path, newline="", encoding="utf-8-sig" as fh: for row in csv.DictReader fh : key = row.get spec "key" or "" .strip .lower if not key: continue seat = { "last active": parse day row.get spec "last active" , "variable usd": float row.get spec.get "variable usd", "" , "" or 0 , } old = seats.get key if old and old "last active" and not seat "last active" or old "last active" seat "last active" : seat "last active" = old "last active" seats key = seat return seats Step 8: Join seats to people and flag activity The join walks the roster once. For each person and each tool it looks up the seat, records whether the seat exists, and records whether the last activity falls inside the window. Seats that no roster row claims become orphans, and their variable spend travels with them so it still shows up in the total. python def join roster, seats by tool, cfg : as of = date.fromisoformat cfg "as of" cutoff = as of - timedelta days=cfg "window days" rows, matched = , {t: set for t in TOOLS} for person in roster: row = {"email": person "email" , "team": person "team" } for tool in TOOLS: key = person.get cfg "tools" tool "roster key" or "" .strip .lower seat = seats by tool tool .get key if key else None if seat: matched tool .add key last = seat "last active" if seat else None row f"{tool} seat" = seat is not None row f"{tool} active" = bool last and last = cutoff row f"{tool} variable usd" = seat "variable usd" if seat else 0.0 rows.append row orphans = {} for tool in TOOLS: keys = sorted set seats by tool tool - matched tool orphans tool = { "count": len keys , "variable usd": sum seats by tool tool k "variable usd" for k in keys , "keys": keys, } return rows, orphans Two design choices matter. The cutoff comparison uses greater-than-or-equal, so a developer whose last activity landed exactly 30 days before the as-of date still counts. And the function returns orphans separately instead of dropping them, because an orphan seat is both a cost leak and a security finding. Step 9: Summarize shares, overlap, and team splits The summary function turns per-person rows into the metrics from Step 1. The overlap block classifies every developer into one of four groups: active on both tools, Copilot only, Cursor only, or no AI activity. The share-of-AI-users figures divide by developers active on at least one tool, which is the closest internal analogue to the survey’s filter on developers who use AI assistants. Reporting both views side by side headcount denominator and AI-user denominator shows exactly how much the choice of denominator moves the answer. python def summarize rows, orphans, cfg : n = len rows out = {"headcount": n, "tools": {}} for tool in TOOLS: price = cfg "tools" tool "seat usd per month" matched = sum r f"{tool} seat" for r in rows extra = orphans tool "count" if orphans else 0 extra var = orphans tool "variable usd" if orphans else 0.0 active = sum r f"{tool} active" for r in rows seats = matched + extra variable = sum r f"{tool} variable usd" for r in rows + extra var total = seats price + variable out "tools" tool = { "seats": seats, "orphan seats": extra, "assigned share": round matched / n, 4 if n else 0.0, "active": active, "active share": round active / n, 4 if n else 0.0, "utilization": round active / seats, 4 if seats else 0.0, "idle seats": seats - active, "idle spend usd": round seats - active price, 2 , "seat spend usd": round seats price, 2 , "variable spend usd": round variable, 2 , "total spend usd": round total, 2 , "cost per active usd": round total / active, 2 if active else None, } both = sum r "copilot active" and r "cursor active" for r in rows cop only = sum r "copilot active" and not r "cursor active" for r in rows cur only = sum r "cursor active" and not r "copilot active" for r in rows any active = both + cop only + cur only out "overlap" = { "both active": both, "copilot only": cop only, "cursor only": cur only, "any active": any active, "no ai activity": n - any active, "copilot share of ai users": round both + cop only / any active, 4 if any active else 0.0, "cursor share of ai users": round both + cur only / any active, 4 if any active else 0.0, } return out def by team rows, cfg : groups = {} for r in rows: groups.setdefault r "team" , .append r floor = cfg "min group size" big = {t: g for t, g in groups.items if len g = floor} small = r for t, g in groups.items if len g < floor for r in g if small: big "Other small teams " = small if "Other small teams " in big and len big "Other small teams " < floor: del big "Other small teams " return {t: summarize g, None, cfg for t, g in sorted big.items } The team breakdown enforces a privacy floor. Any team with fewer than min group size people is merged into an Other bucket, and if that bucket is still too small it is dropped. Otherwise a four-person team’s report would reveal exactly who is using what. The floor also protects you from reading noise as signal. Percentages from five or six people swing by 15 to 20 points when a single person changes habits, a problem the margin-of-error walkthrough for Cursor vs Copilot https://shattered.io/cursor-vs-copilot-margin-of-error-wilson-interval-12-steps-2026/ quantifies with confidence intervals. Phase 4: Price the Seats, Write the Report, and Test It The last three steps turn counts into money, money into a readable report, and the report into something you can trust. Step 10: Apply the cost model Cost lives inside summarize : seats times the configured price, plus variable spend, divided by active developers. Idle spend multiplies the idle seat count by the seat price. When a tool has zero active developers, cost per active returns None and the report prints n/a instead of dividing by zero or showing a misleading zero. The credits to usd helper converts Copilot AI credits to dollars at the documented $0.01 per credit, so a usage export that lists credits becomes a variable usd column with one line of spreadsheet math. A $10 budget covers 1,000 credits, and GitHub’s settings page alerts at 75%, 90%, and 100% of any configured budget, which makes a useful cross-check against your own total. Keep the seat price honest. If you pay annually, use the effective monthly amount. If you hold a mix of plans, run the audit once per plan and add the results, or extend the config with a per-seat price column. A single blended price hides the difference between a $15-credit seat and a $200-credit seat. Step 11: Render the report and wire up the command line The renderer builds a markdown report with the headline table, the overlap paragraph, the orphan list, and the team table. The main function reads the config, runs the pipeline, and writes three files: report.md , summary.json , and orphans.csv . python def pct x : return f"{x 100:.1f}%" def usd x : return "n/a" if x is None else f"${x:,.2f}" def render summary, teams, orphans, cfg : t = summary "tools" o = summary "overlap" lines = f" Cursor vs Copilot seat audit, as of {cfg 'as of' }", "", f"Activity window: {cfg 'window days' } days. Headcount in roster: {summary 'headcount' }.", "", "| Metric | Copilot | Cursor |", "|---|---|---|", f"| Seats incl. orphans | {t 'copilot' 'seats' } | {t 'cursor' 'seats' } |", f"| Assigned share of headcount | {pct t 'copilot' 'assigned share' } | {pct t 'cursor' 'assigned share' } |", f"| Active developers | {t 'copilot' 'active' } | {t 'cursor' 'active' } |", f"| Active share of headcount | {pct t 'copilot' 'active share' } | {pct t 'cursor' 'active share' } |", f"| Seat utilization | {pct t 'copilot' 'utilization' } | {pct t 'cursor' 'utilization' } |", f"| Idle seat spend / month | {usd t 'copilot' 'idle spend usd' } | {usd t 'cursor' 'idle spend usd' } |", f"| Total spend / month | {usd t 'copilot' 'total spend usd' } | {usd t 'cursor' 'total spend usd' } |", f"| Cost per active developer | {usd t 'copilot' 'cost per active usd' } | {usd t 'cursor' 'cost per active usd' } |", "", " Overlap", "", f"Active on both: {o 'both active' }. Copilot only: {o 'copilot only' }. " f"Cursor only: {o 'cursor only' }. No AI activity: {o 'no ai activity' }.", f"Among developers active on at least one tool, Copilot reaches {pct o 'copilot share of ai users' } " f"and Cursor reaches {pct o 'cursor share of ai users' }.", "", " Orphan seats no roster match ", "", f"Copilot: {orphans 'copilot' 'count' }. Cursor: {orphans 'cursor' 'count' }. " "Check these against your offboarding list first.", "", " By team groups under the privacy floor are merged ", "", "| Team | Headcount | Copilot active | Cursor active | Both |", "|---|---|---|---|---|", for name, s in teams.items : lines.append f"| {name} | {s 'headcount' } | {pct s 'tools' 'copilot' 'active share' } | " f"{pct s 'tools' 'cursor' 'active share' } | {s 'overlap' 'both active' } |" return "\n".join lines + "\n" def main : ap = argparse.ArgumentParser ap.add argument "--config", default="config.json" ap.add argument "--out", default="out" args = ap.parse args cfg = json.loads Path args.config .read text base = Path args.config .resolve .parent roster = load roster base / cfg "roster" seats = {t: load seats base / cfg "tools" t "file" , cfg "tools" t for t in TOOLS} rows, orphans = join roster, seats, cfg summary = summarize rows, orphans, cfg teams = by team rows, cfg out = Path args.out out.mkdir parents=True, exist ok=True out / "summary.json" .write text json.dumps {"overall": summary, "teams": teams}, indent=2 out / "report.md" .write text render summary, teams, orphans, cfg with open out / "orphans.csv", "w", newline="", encoding="utf-8" as fh: w = csv.writer fh w.writerow "tool", "key" for tool in TOOLS: for key in orphans tool "keys" : w.writerow tool, key print out / "report.md" .read text if name == " main ": main Run it from the project folder: python3 make sample.py python3 seataudit.py --config config.json --out out cat out/orphans.csv Step 12: Write tests that pin the edge cases Six tests cover the places where seat audits quietly go wrong: the inclusive window boundary, a seat one day outside the window, orphan pricing, variable spend in cost per active, the small-team merge, and the credit conversion. Put them in tests/test seataudit.py . python import unittest from datetime import date import seataudit as sa CFG = { "as of": "2026-11-19", "window days": 30, "min group size": 2, "tools": { "copilot": {"roster key": "github login", "seat usd per month": 20.0}, "cursor": {"roster key": "email", "seat usd per month": 30.0}, }, } ROSTER = {"email": " email protected ", "github login": "a-gh", "team": "t1"}, {"email": " email protected ", "github login": "b-gh", "team": "t1"}, {"email": " email protected ", "github login": "c-gh", "team": "t2"}, def seat day, var=0.0 : return {"last active": date.fromisoformat day if day else None, "variable usd": var} class SeatAuditTests unittest.TestCase : def test window edge is inclusive self : seats = {"copilot": {"a-gh": seat "2026-10-20" }, "cursor": {}} rows, = sa.join ROSTER, seats, CFG self.assertTrue rows 0 "copilot active" def test one day outside window is idle self : seats = {"copilot": {"a-gh": seat "2026-10-19" }, "cursor": {}} rows, = sa.join ROSTER, seats, CFG self.assertFalse rows 0 "copilot active" self.assertTrue rows 0 "copilot seat" def test orphans are counted and priced self : seats = {"copilot": {"ghost-gh": seat "2026-11-01", 5.0 }, "cursor": {}} rows, orphans = sa.join ROSTER, seats, CFG s = sa.summarize rows, orphans, CFG self.assertEqual orphans "copilot" "count" , 1 self.assertEqual s "tools" "copilot" "total spend usd" , 25.0 self.assertIsNone s "tools" "copilot" "cost per active usd" def test cost per active includes variable spend self : seats = { "copilot": {"a-gh": seat "2026-11-10", 10.0 , "b-gh": seat None }, "cursor": {}, } rows, orphans = sa.join ROSTER, seats, CFG s = sa.summarize rows, orphans, CFG "tools" "copilot" self.assertEqual s "total spend usd" , 50.0 self.assertEqual s "cost per active usd" , 50.0 self.assertEqual s "idle seats" , 1 def test small team is merged self : rows, = sa.join ROSTER, {"copilot": {}, "cursor": {}}, CFG teams = sa.by team rows, CFG self.assertIn "t1", teams self.assertNotIn "t2", teams def test credit conversion self : self.assertEqual sa.credits to usd 1000 , 10.0 if name == " main ": unittest.main Run the suite with PYTHONPATH=. python3 -m unittest discover -s tests -v . All six should pass in well under a second. If you later change the cost model or the window logic, a failing test tells you before a wrong number reaches a budget meeting. Sample Output From the Synthetic 57-Person Team Running the command on the generated data produces the report below. Every number here comes from synthetic data with a fixed seed and placeholder prices of $20 per Copilot seat and $30 per Cursor seat. Treat them as a demonstration of the format, never as a statement about either product. Cursor vs Copilot seat audit, as of 2026-11-19 Activity window: 30 days. Headcount in roster: 57. | Metric | Copilot | Cursor | |---|---|---| | Seats incl. orphans | 55 | 29 | | Assigned share of headcount | 94.7% | 49.1% | | Active developers | 44 | 23 | | Active share of headcount | 77.2% | 40.4% | | Seat utilization | 80.0% | 79.3% | | Idle seat spend / month | $220.00 | $180.00 | | Total spend / month | $1,394.00 | $2,214.10 | | Cost per active developer | $31.68 | $96.27 | Overlap Active on both: 18. Copilot only: 26. Cursor only: 5. No AI activity: 8. Among developers active on at least one tool, Copilot reaches 89.8% and Cursor reaches 46.9%. Orphan seats no roster match Copilot: 1. Cursor: 1. Check these against your offboarding list first. By team groups under the privacy floor are merged | Team | Headcount | Copilot active | Cursor active | Both | |---|---|---|---|---| | data | 9 | 77.8% | 33.3% | 3 | | growth | 8 | 87.5% | 50.0% | 3 | | mobile | 10 | 80.0% | 50.0% | 4 | | payments | 12 | 66.7% | 41.7% | 4 | | platform | 14 | 92.9% | 28.6% | 3 | Read the output from the top. For each tool, compare its matched seats plus any orphan seats against your headcount to get its assigned share. Then compare active share and utilization. In this synthetic team the two tools are about equally well used per seat, but Cursor costs far more per active developer because its sample variable spend is heavier. That pattern is an artifact of the generator. In a real audit you would check whether the variable spend is concentrated in a few heavy users, since a few people running expensive agent sessions can make a tool look costly while delivering the most value. The overlap line shows 18 developers active on both tools and 8 with no AI activity at all. Those 8 are the first conversation to have, before anyone debates which tool wins. The share-of-AI-users line is the closest analogue to a market-share number, and it differs sharply from the headcount view. That gap is the denominator effect from the survey section, reproduced inside one small team. The security team 4 people does not appear in the team table. That is the privacy floor working as designed: with a floor of 5, a four-person team merges into the Other bucket, which is itself too small and gets dropped. Usage numbers also say nothing about output quality. When the audit shows a close call, run a controlled task comparison such as the Cursor vs GitHub Copilot benchmark https://shattered.io/benchmark-cursor-vs-github-copilot-2026/ before moving any licenses. Common Pitfalls These seven mistakes account for most wrong audit numbers. Each one is cheap to avoid once you know it exists. 1. Dividing by the wrong denominator. If you divide active developers by seats, you measure utilization. If you divide by headcount, you measure reach. Both are useful, but they answer different questions, and mixing them in one sentence reproduces the survey confusion. Label every percentage with its denominator. 2. Treating an assigned seat as a used seat. A seat stays billable until someone removes it. A seat with an empty or stale last-activity field is spend without usage. The report separates assigned share from active share for this reason. 3. Joining on the wrong identifier. Copilot identifies people by GitHub login, Cursor by email, and your roster by both. Personal GitHub accounts, plus-addressed emails, and mixed-case addresses all create false orphans. If your orphan count looks high, inspect orphans.csv before trusting any other number. 4. Ignoring variable spend. A seat price tells you the floor, not the bill. With Copilot usage metered in AI credits and Cursor billing third-party models at API list prices on higher tiers, two teams on the same plan can have very different invoices. Include the variable column or state clearly that your cost-per-developer figure is a floor. 5. Comparing different windows. A 30-day Copilot figure next to a 90-day Cursor figure favors whichever tool got the longer window. Freeze one window in the config and apply it to every tool. 6. Publishing per-person data. The audit exists to size licenses, not to rank people. Report at team level, keep the privacy floor, and keep the row-level join out of slide decks. If a manager asks who is not using a tool, point them to the license-reclamation process instead of a name list. 7. Trusting list prices. Your negotiated rate, annual discount, and promotional credits change the real cost. Pull the numbers from invoices, and recheck them after every plan change. Cursor’s recent restructuring is a good example, as the state-of-play section below explains. Troubleshooting Seat Audit Problems Most failures show up as a Python traceback or as a number that looks too good. This table maps the ten symptoms people hit most often to their causes and fixes. | Symptom | Likely cause | Fix | |---|---|---| | KeyError naming a column such as email | Config column name does not match the export header, or the file has a byte-order mark | Open the CSV, copy the exact header into config.json, and keep the utf-8-sig loader | | ValueError: Invalid isoformat string | Dates arrive as 11/18/2026 or as relative text | Convert the column to ISO year-month-day dates in a spreadsheet before saving | | fetch copilot.py exits with an authentication error | Token missing, expired, or lacking the needed scope | Create a fresh read-only token and confirm the endpoint’s accepted scopes in GitHub’s docs | | fetch copilot.py exits with an authorization or not-found error | Token not authorized for the org’s SSO, wrong org name, or Copilot billed at enterprise level | Authorize the token for SSO, check the org slug, or use the enterprise billing export | | Every seat shows as an orphan | roster key points at the wrong roster column, or identifiers differ in form | Compare one login or email in the roster and the export by eye and fix the config mapping | | Active count is zero for a tool people clearly use | as of date set in the past, window too short, or last-activity fields empty | Check the newest date in the export against as of and widen the window if needed | | Cost per active developer prints n/a | No developer was active in the window | Verify date parsing, then confirm the window and the as of date | | A team is missing from the team table | The team has fewer people than min group size | Keep the floor, or lower it only with written approval from your privacy owner | | Total spend differs from the invoice | Placeholder prices, proration, discounts, or variable spend not exported | Replace config prices with invoice amounts and add the variable column | | Copilot seat count is lower than billing shows | Seats managed through an enterprise account are not in the organization list | Export from the enterprise billing page and map its columns | When none of these fits, print the first five rows of each loaded dictionary in an interactive Python session. A wrong key shape is visible within seconds. Advanced Tips for Better Cursor vs Copilot Audits The base project answers one snapshot question. These upgrades turn it into a monitoring habit. Schedule it. Run the audit on the first of each month and keep the summary.json files. Three months of trend lines beat one dramatic snapshot, and they reveal whether a drop in active share is a real decline or a holiday dip. If your team already automates agents in CI, the walkthrough on running Cursor’s CLI in GitHub Actions https://shattered.io/cursor-cli-github-actions-headless-agent-2026/ shows the pattern for scheduled jobs. Keep the CSV exports out of the repository, and upload only the aggregate summary as an artifact. Pin the action versions to the current majors or to commit SHAs when you adapt this. name: seat-audit on: schedule: - cron: "0 6 1 " workflow dispatch: permissions: contents: read jobs: audit: runs-on: ubuntu-latest steps: - uses: actions/checkout@v4 - uses: actions/setup-python@v5 with: python-version: "3.12" Replace this step with a secure fetch of the roster and the Cursor export - name: Fetch Copilot seats env: GITHUB TOKEN: ${{ secrets.COPILOT READ TOKEN }} run: python3 fetch copilot.py your-org data/copilot seats.csv - name: Run audit run: python3 seataudit.py --config config.json --out out - uses: actions/upload-artifact@v4 with: name: seat-audit-summary path: out/summary.json retention-days: 90 Add a third tool. The TOOLS tuple and the config drive everything, so adding another assistant takes one config block and one tuple entry. OpenAI’s own rate card estimates Codex at roughly $100 to $200 per developer per month on average, with real cost varying by model, concurrent instances, automations, and fast mode, which makes it a good candidate for the variable-spend column. Weight by depth of use. A last-activity date says a developer touched the tool, not how much. If an export offers request counts or agent sessions, add a second threshold such as active on at least eight separate days, and report light users and heavy users separately. Light users often justify a cheaper plan, and heavy users often justify an upgrade. Run multiple windows. Compute 7-day, 30-day, and 90-day active shares in one pass. A tool with strong 90-day reach and weak 7-day reach is losing people. The reverse pattern signals a recent rollout that is still ramping. Reclaim idle seats on a schedule. Send each team owner the list of seats idle for 60 days, with a two-week notice before removal. Reclaiming idle seats is the highest-return action in the whole project, and it needs no debate about which tool is better. Chart the denominator drift. Store headcount alongside every snapshot. A share that falls because hiring outpaced seat purchases tells a different story from one that falls because developers stopped logging in. Keeping the Audit Data Safe Seat exports look harmless, but they are employee activity records with names and timestamps. Treat them like any other HR-adjacent dataset. Start with access. Keep the working folder on an encrypted disk, restrict it to the two or three people running the audit, and add data/ and out/ to .gitignore on day one. A token with read-only scope limits the damage if the script ever leaks, and storing it in an environment variable keeps it out of config files and commits. Next, limit what leaves the project. The report shows team-level percentages and an orphan list. Share the report, not the row-level join. If you need to circulate the orphan list, send it only to the IT and security owners who handle offboarding. Orphan seats deserve a security review, not just a billing one. A seat held by someone who left the company is a live account tied to your tenant. In the synthetic sample, the audit flags one former contractor’s Copilot login and one alumni address on Cursor. In a real organization, each orphan should trigger the same offboarding checklist you use for other SaaS accounts. Finally, keep the toolchain small. This project installs nothing, so no package can add a surprise dependency. That choice has a security payoff. Cursor itself posted a security notice about malicious versions of an Anthropic Claude Code extension versions 2.1.312 and 2.1.313 in the Cursor VS Code Extension Marketplace, a reminder that developer tooling is an active supply-chain target. A script with zero third-party packages leaves nothing to compromise. Set a retention rule too. Delete raw exports after the report is approved, keep only the aggregated summary.json , and write the deletion date in your notes. Activity data you no longer hold cannot leak. Turning the Numbers Into a Renewal Decision A report changes nothing until it triggers a decision. Use simple rules and write them down before the next audit so the thresholds are not negotiable after the fact. First, fix the waste. Idle seats cost money whichever tool you prefer. If utilization falls below your threshold many teams pick 60% to 70% , reclaim seats before you renew anything. The idle-spend row gives you the monthly figure to put in the proposal. Second, check the overlap. Developers active on both tools are paying for two assistants. That is sometimes justified, because one tool may handle inline completion while the other runs longer agent tasks. It is wasteful when both are used the same way. Ask the overlap group which tool they would keep, and weigh the answer more heavily than any survey. Third, compare cost per active developer with a cost-and-value mindset. A tool that costs more per active developer can still win if it removes review cycles or lets a team ship faster. The audit supplies the cost side. The value side needs an outcome measure such as cycle time or review turnaround, which no seat export can provide. Fourth, plan switches deliberately. Moving a team from one tool to the other costs configuration time, not just license fees. The guide on migrating from Copilot instructions to Cursor rules https://shattered.io/cursor-rules-migrate-github-copilot-2026/ lists what carries over and what needs rewriting. Run a two-month pilot with a single team, rerun the audit at the end, and compare active share before and after. Fifth, set a recheck date. Both vendors have changed packaging within the past few months, so a decision that is right today can be wrong by next quarter. Put the next audit on the calendar when you sign. Where Cursor and Copilot Stand on November 19, 2026 Both products changed quickly this autumn, and the changes affect how you read an audit. Treat what follows as the current state, and check each vendor’s pages before you sign anything. GitHub added a run of models to Copilot in under ten days. According to GitHub’s November 2026 changelog https://github.blog/changelog/month/11-2026/page/2/ , Gemini 4 Carbon reached general availability on November 10 and Gemini 4 Hydrogen on November 13. OpenAI’s GPT-6.2 Astra and GPT-6.2 Sol arrived on November 17, alongside MAI Code 2 and MAI Code 2 Flash. Gemini 4 Neon followed on November 18 for Pro+, Max, Business, and Enterprise users, billed at provider list prices under usage-based billing. GitHub also deprecated unspecified Copilot models on November 16. For auditors, the lesson is that usage-based billing makes variable spend a first-class number. Cursor restructured its plans in the same period. Since October 12, new customers have reportedly bought through a structure called The X Subscription, with Free, Plus at $8 a month, Premium at $30, Super at $100, and Ultra at $200. The older Pro plan at $20 and Pro+ plan at $60 are now treated as legacy offerings, and a pricing breakdown reports that moving from a legacy plan to the new structure cannot be reversed. For your audit, that means a team may hold seats on two different Cursor structures at once, so run the cost model once per plan, as Step 10 suggests. On the product side, Cursor said cloud agents start three times faster with builds announced November 10 and added Voice Chat with Agents on November 6 for paid plans with cloud agents enabled. A headline on Cursor’s own blog says its recurring revenue doubled in three months to $2 billion, which is a company claim and not an independent measurement. Two cautions apply. A report that a tool is growing quickly says nothing about whether your team uses it. And no verified number puts Cursor against Copilot on seat share, so any article that quotes one deserves a hard look at its denominator. The model churn on both sides is covered in the comparison of Copilot and Cursor model releases https://shattered.io/github-copilot-vs-cursor-4-models-8-days-2026/ , and the wider coding and developer tools coverage https://shattered.io/category/software/ tracks pricing changes as they land. Frequently Asked Questions Is there a reliable Cursor vs Copilot market share figure? Not as of November 19, 2026. Published numbers use different denominators: survey respondents, filtered respondents, organizations, and company logos. For a decision about your own licenses, your seat exports beat any public statistic. Which tool costs less per active developer? It depends on your plan mix and usage. The audit computes it from your own seat prices and variable spend. The sample figures in this article come from synthetic data and say nothing about either product. What counts as an active developer? Anyone whose last-activity date falls inside your chosen window, 30 days in this project. Pick the window before you look at results and apply it to every tool. Can I run the audit without the GitHub API? Yes. Export Copilot seats from the billing page, save them as CSV, and map the columns in config.json. The API script only saves typing. Enterprise-billed seats often need the billing export anyway. How often should I rerun it? Monthly is a good default. Rerun it after any plan change, rollout, or reorganization, and keep the summary files so you can chart trends. Is tracking per-developer activity allowed? That depends on your jurisdiction and your employment policies. Check with HR and your privacy officer before you start, report at team level, apply a privacy floor, and delete raw exports once the report is approved. Will Cursor’s plan changes break the audit? The code does not care about plan names, only about seats, dates, and prices in the config. If you hold seats on both legacy and new structures, run the cost model per plan and add the totals. Can the project handle a third assistant? Yes. Add the tool to the TOOLS tuple, add a config block, and extend the report table. The join and summary logic loops over tools instead of naming them.