Cursor, GitHub Copilot, Claude Code, and others: we ran the numbers across 200+ developer workflows. The results are higher than most teams expect.
Most engineering teams dramatically underestimate what AI coding tools actually cost once they move beyond the marketing claims.
We analyzed anonymized usage data from 200+ developers across startups and mid-size companies using Cursor, GitHub Copilot, Claude Projects, and several other tools throughout late 2025. Here's what we found.
The Sticker Price vs Reality Gap
Marketing pages love to show low per-user numbers. Reality is messier.
Tool
Advertised Price
Real Avg. Monthly Cost (Heavy Users)
Effective Cost per Engineer
Cursor Pro
0/mo
$47
$31
GitHub Copilot
0/mo (Ind)
8
9
Claude Team + API
$30/user + usage
$94
$61
Copilot + Custom GPTs
0 + OpenAI
$71
$44
The gap comes from three sources most teams don't model:
1. Context window inflation: Real sessions routinely use 40k-90k tokens per task once you include the full codebase context, not the 2-3k token examples in demos. 2. Background agents and autocomplete: These features run constantly in the background and are rarely disabled. 3. Evaluation and iteration: Developers rarely accept the first suggestion. 3-7 iterations is common on complex tasks.
Usage Patterns That Destroy Budgets
The highest spenders weren't the teams with the most engineers. They were the teams with:
Heavy reliance on agentic workflows (Cursor Composer, Aider, etc.)
Large monorepos (context loading becomes extremely expensive)
Junior-to-mid level engineers using AI as primary coding method (higher iteration counts)
One 14-person team we tracked was burning
,800/month on AI coding tools alone after six months, roughly
00 per engineer. That sits well above the per-tool numbers above because they had stacked multiple subscriptions on top of heavy agentic API usage, not because any single tool costs that much.
Where Teams Waste the Most Money
Over-provisioned context: Loading entire monorepos into every prompt when 15-20% of the codebase would suffice. This one pattern accounts for roughly 35% of wasted spend.
Default model selection: Many tools default to the most expensive model. Switching heavy users to Claude Haiku 4.5 or GPT-5.4 mini for routine tasks drops costs 60-75% with minimal quality loss on most tasks.
Duplicate subscriptions: We saw multiple teams where engineers had both Cursor + GitHub Copilot + personal Claude subscriptions active simultaneously.
The ROI Question Nobody Answers Honestly
Does $30-60 per engineer per month in AI tooling pay for itself?
Our data suggests:
High-skill senior engineers: Often yes, especially on greenfield work and debugging. Time savings of 25-40% are common in our interviews.
Junior and mid-level engineers: Much less clear. Speed increases, but code quality and architectural decisions frequently require more senior review time. Net productivity gain is often 10-18%.
Large existing codebases: ROI drops significantly. Context costs explode and suggestion quality decreases.
The uncomfortable truth is that many teams are paying for a productivity tool whose actual return they have never measured.
Recommendations for 2026
1. Measure before scaling. Run a 6-week controlled experiment with clear time-tracking before rolling AI tools out company-wide. 2. Default to cheaper models for 70-80% of work. Reserve Sonnet/GPT-5.4 for architecture and complex logic. 3. Set hard monthly budgets per engineer ($35-50 range is reasonable for most teams). 4. Audit subscriptions quarterly. Cancel anything not actively used.
AI coding tools are genuinely useful. They're also genuinely expensive once real usage patterns emerge. The teams winning right now are the ones treating this as a real line item with real optimization work, not an unlimited productivity magic button.
---
*Methodology: Aggregated self-reported usage + exported billing data from 14 companies between September 2025 and January 2026. All numbers anonymized and rounded.*
This article is part of ongoing research into real technology costs. Figures are based on public pricing at publication time and may change.