Claude Code vs Cursor vs Copilot for Enterprise Engineering Teams

Experienced open-source developers using Cursor Pro took 19% longer in a METR randomized trial, despite forecasting they would finish 24% faster. That mismatch between forecast and outcome, not the headline productivity numbers, changes how to evaluate standardizing a thousand AI coding seats. The three tools competing for that budget (Claude Code, Cursor, and GitHub Copilot) diverge on seat cost, security posture, and measurable effect on code quality in ways that matter before the first seat ships.

What a Thousand Seats Actually Costs

Seat prices look close on paper and diverge quickly in practice. GitHub Copilot Business lists at $19 per user per month; Copilot Enterprise is $39 and requires GitHub Enterprise Cloud underneath, which lifts the effective floor nearer to $60. Cursor Teams and Business sit around $40 per seat; the Enterprise tier is quoted individually with SCIM, audit logs, and pooled credits. Claude Code Enterprise lists a $20 seat and bills token usage at API rates on top. In real deployments, teams report $100 to $200 per active developer per month once agents are the primary interface.

For a thousand seats, the spread from roughly $228,000 to north of $2 million per year is easy to hit by varying plan and usage. Three questions resolve most of it:

• How heavy is agent use. Token spend dominates once agents handle multi-step work.
• How much pooled headroom the team needs. GitHub's June 2026 shift to AI Credits moved features from the flat fee to metered consumption.
• Whether VPC or in-tenant deployment is required. That filter cuts the shortlist before the unit math starts.

A disciplined procurement treats seat price as the floor, not the budget. Published usage caps and overage rates belong in the model before contracting, especially if the org is standardizing AI development services across several product teams.

Enterprise Security Controls Decide the Shortlist

Compliance now sorts the three tools cleanly. All three carry SOC 2 Type II. Only GitHub Copilot ships FedRAMP Moderate hosting and US + EU data residency, added April 2026 with a 10% model-multiplier premium when FedRAMP or residency routing is in effect. Only Claude Code Enterprise offers VPC deployment, where prompts and code stay inside the customer's network. Cursor sits between them with SOC 2 Type II, Privacy Mode, SCIM, and audit logs at the Business tier, Enterprise custom.

Prompt injection is now a procurement question. CurXecute (CVE-2025-54135) gave an attacker remote code execution in Cursor by having the agent silently rewrite ~/.cursor/mcp.json through a poisoned MCP server such as Slack. Severity was CVSS 8.6; Cursor patched in version 1.3 on July 29, 2025. The takeaway is less about Cursor specifically and more about MCP as blast radius. Any assistant that auto-runs MCP entries needs the same approval, sandbox, and rollback discipline you would place around CI/CD, which is exactly the frame used when running AI agents in production.

Measurable Code Quality Varies by Developer and Task

A controlled GitHub experiment with 95 professional developers found Copilot users completed an HTTP server task 55% faster, averaging 1 hour 11 minutes against 2 hours 41 minutes without the tool, with a higher completion rate of 78% versus 70%. That is the win column, and it is real for greenfield, well-scoped work.

The counterweight is METR's randomized trial with 16 experienced open-source developers across 246 real tasks: developers forecasted a 24% reduction in completion time and actually took 19% longer. Context explains most of the mismatch between forecast and outcome. Mature repos, subtle invariants, and senior engineers who already know where the code lives are exactly where AI assistance underperforms the expectation, because reading an unfamiliar suggestion costs more than writing the familiar line.

The procurement implication is unglamorous: measure in your own repo, with your own developers, on your own tickets. Then layer in review discipline before extrapolating. The team that audits AI-generated code before production catches the regressions that make raw speed meaningless.

Adoption Signals Are Not Interchangeable

Three headline numbers tell three different stories. GitHub Copilot crossed 20 million all-time users by July 2025. Anthropic reported Claude Code at over $2.5 billion in run-rate revenue in early 2026, with enterprise accounting for more than half and roughly 4% of all public GitHub commits authored by the tool. Cursor hit $1 billion in ARR 24 months after launch with about 300 employees.

Loyalty separates differently from adoption. The Pragmatic Engineer 2026 survey of roughly 1,000 engineers named Claude Code most-loved by 46%, Cursor by 19%, and GitHub Copilot by 9%. Workplace use is also shifting faster than most procurement cycles. JetBrains data across the first half of 2026 shows Claude Code climbing from 18% to 39%, Codex from 3% to 16%, Copilot falling from 29% to 21%, and Cursor from 18% to 12%.

A Procurement Framework That Survives Contact

Four steps keep this from becoming a demo-driven decision reversed in year two.

1. Gate on compliance first. FedRAMP or EU residency requirements thin the shortlist before productivity debates begin.
2. Pilot small, measure what matters. Twenty engineers for four weeks, same repo, with PR cycle time and change-failure rate as the yardsticks. One tool can make a task 55% faster in a lab and another can slow senior developers by 19% on their own code. Both results are real.
3. Harden the MCP boundary. Treat the agent's ability to edit local config the way you would treat a service account: approvals, logs, rollback, and no silent auto-run.
4. Expect multi-tool. About 70% of surveyed engineers already use two to four assistants. Standardize one; do not forbid the other two by policy and then discover them in expense reports.

Governance scales badly if the standard tool and the shadow tool answer to different owners. Pick one for the enterprise agreement, name an owner for the others, and keep the audit trail consistent across all three.

What to Watch at Sixty Days

The number worth watching sixty days after a rollout is PR cycle time against the pre-rollout baseline, held to the same class of work. A credible assistant shaves it by at least 10% on repeatable tickets without lifting change-failure rate; a 5% dip paired with a review backlog means autocomplete is replacing thought, not rework. The METR mismatch between a 24% forecast and a 19% slowdown is the reminder worth pinning above the dashboard. Confidence in AI output is not throughput, and buying a thousand seats of the wrong one is an expensive way to find that out.