| name | cursor-cost-optimization |
| description | Reduce Cursor AI spending without sacrificing developer productivity. Use when asked about cost optimization, budget management, model selection strategy, or spend reduction for Cursor Enterprise teams. |
Cursor Cost Optimization
You have access to Cursor Enterprise usage data through the cursor-usage MCP server. This skill teaches you how to identify cost-saving opportunities and recommend actionable changes.
Cost Optimization Framework
Step 1: Understand the Spend Profile
Call get_team_overview to get the baseline, then:
-
Identify the spend distribution — Is spend concentrated in a few users or spread evenly?
- If top 10% of users account for >50% of spend → focus on those users
- If spend is evenly distributed → focus on model selection policies
-
Identify the cost driver — Is it model choice, volume, or both?
- Call
get_model_usage to see which models dominate
- Premium models (Opus, GPT-5) at 10-50x the cost of standard models (Sonnet, GPT-4o)
- A team of 50 where 5 people use Opus can spend more than the other 45 combined
Step 2: Model Selection Optimization
The single highest-impact cost lever is model selection.
| Task Type | Recommended Model Tier | Why |
|---|
| Code completion / tabs | Budget (Flash) | High volume, low complexity, latency-sensitive |
| Inline edits (Cmd+K) | Standard (Sonnet, GPT-4o) | Good balance of quality and cost |
| Chat conversations | Standard | Most questions don't need frontier models |
| Agent mode (complex tasks) | Premium (Opus) only when needed | Reserve for genuinely complex multi-step work |
| Code review | Standard | Pattern matching, not creative generation |
Key insight: Most developers default to the "best" model out of habit, not necessity. 80%+ of requests can be handled by standard-tier models with no noticeable quality difference.
Step 3: Spend Limits
Use set_spend_limit to set guardrails:
- Soft approach: Set limits at 2-3x the team median spend. This catches runaway usage without blocking normal work.
- Hard approach: Set limits at a fixed dollar amount per cycle. Good for budget-constrained teams.
- Per-group approach: Use billing groups to set different limits for different teams based on their needs.
Warning: Setting limits too low causes developer frustration and workarounds. Start generous and tighten based on data.
Step 4: Usage Pattern Optimization
- Agent mode loops: Check
get_usage_events for users with many consecutive agent requests. Long agent loops are the #1 cause of unexpected spend spikes.
- Headless requests: Filter events where
isHeadless: true. These are background processes (Bugbot, indexing) that may be running unnecessarily.
- Low acceptance rates: Call
get_agent_edits and get_tabs. If acceptance rates are below 30%, the team may need better prompting practices, not more AI requests.
Cost Benchmarks
These are rough benchmarks based on typical enterprise teams:
| Team Size | Monthly Spend (healthy) | Monthly Spend (high) | Monthly Spend (alarm) |
|---|
| 10 devs | $200-500 | $500-1,500 | >$2,000 |
| 50 devs | $1,000-3,000 | $3,000-8,000 | >$10,000 |
| 100 devs | $2,000-6,000 | $6,000-15,000 | >$20,000 |
| 500 devs | $10,000-30,000 | $30,000-75,000 | >$100,000 |
These assume a mix of standard and premium model usage. Teams exclusively using premium models will be 3-5x higher.
Presenting Recommendations
When presenting cost optimization findings:
- Lead with the dollar impact — "Switching 3 users from Opus to Sonnet would save ~$X/month"
- Show the data — Reference specific users, models, and spend figures
- Acknowledge tradeoffs — Premium models ARE better for complex tasks; the goal is right-sizing, not downgrading
- Suggest incremental changes — Don't recommend sweeping policy changes; suggest a pilot with willing users first
For Deeper Analysis
This skill covers quick, data-driven cost optimization. For ongoing monitoring with:
- Automated anomaly detection (statistical outlier detection)
- Slack/email alerts when spend spikes
- Historical trend analysis over months
- Incident lifecycle tracking (MTTD/MTTI/MTTR)
See cursor-usage-tracker — the open-source dashboard built for exactly this.
