Summary: GitHub Copilot charges $10/month with an estimated cost-to-serve of ~$30/user. This article analyzes the unit economics data and what it indicates about pricing AI-powered products.
The Unit Economics
GitHub Copilot, one of the most widely adopted AI products, shows the following unit economics:
- Revenue: $10/user/month (subscription price)
- Estimated Cost: ~$30/user/month (OpenAI API costs)
- Margin: -$20/user/month
Each additional customer adds to the aggregate margin pressure.
This is Microsoft—one of the most well-resourced technology companies—operating an AI product with negative per-user margins.
This pattern is instructive for understanding AI product economics broadly.
Relevance for AI Products
The factors affecting GitHub Copilot's margins appear in most AI products:
- High usage variance — Some users generate 100x more API costs than others
- Pricing before cost visibility — Prices set before cost-to-serve is understood
- Aggregate-only tracking — Cost-per-customer not visible until significant scale
Factor 1: Usage Variance Affects Flat Pricing
Not all customers cost the same.
GitHub Copilot users distribute into three cohorts:
- Light users (20%): Generate a few completions per day. Cost: ~$5/month.
- Average users (60%): Moderate usage. Cost: ~$15-20/month.
- Power users (20%): All-day usage. Cost: ~$50-100/month.
With $10/month flat pricing:
| Cohort | Cost | Revenue | Margin |
|---|---|---|---|
| Light users | $5 | $10 | +$5 |
| Average users | $15 | $10 | -$5 |
| Power users | $50 | $10 | -$40 |
The positive margin from light users does not offset the negative margin from power users and average users.
Related: For detailed cohort margin analysis across the AI industry, see Usage Variance in AI Products.
Cursor's Pricing Evolution
Cursor followed a similar path. They launched with flat pricing, observed high-usage customers with 10x average costs, and adjusted pricing four times in 2024:
- January 2024: $20/month unlimited
- May 2024: $20/month with soft limits (500 completions)
- September 2024: $20/month with hard limits + overage charges
- December 2024: Hybrid model (base + usage tiers)
Each adjustment involved customer communication and retention considerations.
Related: For migration approaches, see From Flat-Rate to Usage-Based Pricing Migration
Factor 2: Pricing Before Cost Visibility
A typical AI product launch sequence:
- Month 1: Launch with $29/month pricing (based on estimates)
- Month 2: Acquire 50 customers
- Month 3: First complete API bill: $15,000
- Month 4: Calculate actual cost-per-customer: $300
By the time cost-to-serve becomes visible, customers are on established pricing.
The GitHub Copilot Timeline
- 2021: Launch at $10/month (based on initial estimates)
- 2022: Scale to 1M+ users
- 2023: Cost-per-user data shows -$20/user/month
- 2024: Margins remain negative
The lag between pricing decisions and cost visibility was approximately two years.
Factor 3: Aggregate-Only Cost Tracking
Most AI founders can answer:
- "What's our total OpenAI bill this month?" — Yes ($15,000)
- "How many customers do we have?" — Yes (50)
- "What's our MRR?" — Yes ($1,450)
They often cannot answer:
- "Which customer costs us the most?" — Unknown
- "What's our margin on Customer A?" — Unknown
- "What's our gross margin per customer?" — Approximate
Visibility at the aggregate level doesn't reveal per-customer margin variance.
Example: A High-Cost Customer
We observed a customer on a $200/month plan who appeared nominal in dashboards—paying customer, regular usage.
Cost reconciliation showed $12,000/month in OpenAI API costs.
Margin: -$11,800/month on one customer.
This was identified through manual reconciliation. Without that process, the cost would have continued untracked.
Action Steps
1. Identify Highest-Cost Customers
Export OpenAI/Anthropic usage data. Match to customer IDs. Sort by cost.
Questions to answer:
- Which customer has the highest cost-to-serve?
- Is their plan price above or below their cost?
- Would usage limits or tier changes address the gap?
Time required: ~2 hours
2. Model Pricing Scenarios
Test scenarios with your actual data:
- Pure usage-based ($X per token)
- Hybrid (base fee + usage)
- Tiered (different limits per tier)
- Volume discounts (committed use)
Output:
- Margin by scenario
- Retention estimates per scenario
- Revenue impact projections
The Pattern: Early Visibility
Compare two approaches:
Approach A (Copilot path):
- Launch with flat pricing
- Scale to significant user base
- Observe negative margins
- Pricing adjustments affect large customer base
Approach B (Early visibility):
- Launch with flat pricing (acceptable at low volume)
- Track cost-per-customer from launch
- At 50 customers, observe usage distribution
- Adjust to hybrid pricing before significant scale
The difference is when cost visibility becomes available.
Approach B identifies the pattern at 50 customers. Approach A identifies it at 1 million.
What Copilot Demonstrates
GitHub Copilot has:
- Microsoft's engineering resources
- Substantial capital
- 1M+ paying customers
- Strong product-market fit
And operates with negative margins because per-customer cost tracking wasn't prioritized early.
This indicates: per-customer cost visibility is essential regardless of scale or resources.
Bear Billing
We built Bear Billing to provide per-customer cost visibility for AI products:
- Track costs per customer automatically (OpenAI, Anthropic, AWS)
- Identify highest-cost customers
- Model pricing scenarios with real usage data
- Support usage-based billing implementation
Discussion
What are your per-customer margins? We're interested in hearing from founders building AI products.
About the Author: Blaise is building Bear Billing to provide per-customer cost visibility for AI startups. This work is informed by direct experience with a customer showing $12K/month cost on a $200/month plan.