A Quickstart for Measuring the Return on Your Claude Code Investment

Executive summary

Coding tools like Claude Code promise to attach jetpacks to developers, and the data backs it up. With 79% of Claude Code conversations being automation tasks, teams are seeing real productivity gains that go far beyond simple code completion and that models can automate more than augment as they get better. During the June 12th cloud outage, I realized how dependent I'd become on Claude Code - even for fixing simple apostrophe errors in convoluted cURL commands.

Your team probably already understands the need for developer productivity tools, and with Claude Code accessing the Anthropic API directly (directly or through cloud providers), key security concerns have been quelled. What you need now is telemetry to answer the important questions: Are developers actually using it? Which teams are getting the most value? What's our real cost per feature or bug fix?

This guide walks you through setting up observability with Claude Code using OpenTelemetry metrics. You'll get the complete setup - from Prometheus configuration to automated reporting - so your engineering leadership can make data-driven decisions about your AI tooling investment.

Setting up your measurement infrastructure

Quick Verification

Before diving into Prometheus, let's verify telemetry is working:

1export CLAUDE_CODE_ENABLE_TELEMETRY=1
2export OTEL_METRICS_EXPORTER=console
3export OTEL_METRIC_EXPORT_INTERVAL=1000 
4claude -p "hello world"

You should see output like this:

1=== Resource Attributes ===
2{ 'service.name': 'claude-code', 'service.version': '1.0.17' }
3===========================
4
5{
6  descriptor: {
7    name: 'claude_code.cost.usage',
8    type: 'COUNTER',
9    description: 'Cost of the Claude Code session',
10    unit: 'USD',
11    valueType: 1
12  },
13  dataPointType: 3,
14  dataPoints: [
15    {
16      attributes: {
17        'user.id': '7b673f5715f2da49af2cdd341e0cad17fb3274f32d4ceed00d57d53da4e76fb2',
18        'session.id': '092d99fc-ac17-4ee7-b310-57d387777c91',
19        'model': 'claude-4-sonnet-20250514'
20      },
21      value: 0.000297
22    }
23  ]
24}

Prometheus Setup

For actual ROI measurement, you need Prometheus. Console output is just debugging - Prometheus gives you dashboards, historical data, and executive-ready visualizations.

First, grab the configuration files:

1# Clone the configuration repo
2git clone https://github.com/katchu11/claude-code-guide
3docker-compose up -d

Configure Claude Code to send metrics to Prometheus:

1# Enable telemetry
2export CLAUDE_CODE_ENABLE_TELEMETRY=1
3
4# Configure OTLP exporter
5export OTEL_METRICS_EXPORTER=otlp
6export OTEL_EXPORTER_OTLP_PROTOCOL=grpc
7export OTEL_EXPORTER_OTLP_ENDPOINT=http://localhost:4317
8
9# Optional: Set authentication if required
10# export OTEL_EXPORTER_OTLP_HEADERS="Authorization=Bearer your-token"
11
12# Run Claude Code
13claude

Common infrastructure endpoints:

• Local development: http://localhost:4317
• Kubernetes: http://otel-collector.monitoring.svc.cluster.local:4317
• AWS: Your Application Load Balancer endpoint
• DataDog: https://otlp.datadoghq.com
• Honeycomb: https://api.honeycomb.io:443

Key Prometheus Queries

Once data is flowing, these queries give you the insights you need:

1# Total cost across all sessions
2sum(claude_code_cost_usage_USD_total)
3{} 0.43
4
5# Token usage by type (input/output)
6sum(claude_code_token_usage_tokens_total) by (type)
7{type="input"} 751
8{type="output"} 1863
9{type="cacheCreation"} 2136
10{type="cacheRead"} 78465
11
12# Token usage by user email
13sum(claude_code_token_usage_tokens_total) by (user_email)
14{user_email="user@company.com"} 83215
15
16# Cost by model type
17sum(claude_code_cost_usage_USD_total) by (model)
18{model="claude-opus-4-20250514"} 0.429
19{model="claude-3-5-haiku-20241022"} 0.0006
20
21# Rate of token consumption over time (requires multiple data points)
22rate(claude_code_token_usage_tokens_total[5m])
23{type="input"} 2.5
24{type="output"} 6.2
25{type="cacheRead"} 261.5

Understanding Usage and Cost

Let's start with what Claude Code actually costs. Unlike fixed subscription tools, Claude Code has variable costs based on usage patterns.

Token Usage Patterns

The claude_code.token.usage metric breaks down by:

• Input tokens: Code you provide for analysis
• Output tokens: Code Claude generates
• Model type: Different models have different costs

Usage Patterns by Plan Type:

Subscription Users (Pro/Max):

• Claude Pro: predictable $20/month
• Claude Max 5x: $100/month for heavy development
• Claude Max 20x:$200/month for intensive workflows

API Users (Pay-per-token) - Real Usage Patterns:

Pro tip: "ultrathink" triggers Claude's maximum thinking budget for complex analysis. The hierarchy is: "think" < "think hard" < "think harder" < "ultrathink" - each level increases the amount of tokens set to the reasoning budget. See Claude Code's best practices doc for more info!

Simple Development Tasks (Haiku):

• Quick questions, simple functions, code snippets
• 200-500 input tokens, 10-50 output tokens
• $0.0001-0.0003 per simple query ‍

Real examples: Complex architectural analysis cost $0.34, simple "hello world" cost $0.0003

*Annual discounts may be available for subscription plans.

API Pricing by Model:

• Haiku 3.5: $0.80/$4.00 per million input/output tokens (fastest, cheapest)
• Sonnet 4: $3.00/$15.00 per million tokens (balanced performance)
• Opus 4: $15.00/$75.00 per million tokens (most capable) ‍

Pro tip: For regular Claude Code usage, Max subscriptions are often more predictable than API costs. Individual developers can't control API optimizations like prompt caching, making subscriptions simpler for daily development work.

Note: Telemetry may require a clean Claude Code installation if you experience hanging. See troubleshooting.md for solutions.

Session Duration Analysis

Understanding how long developers stay in Claude Code sessions helps identify engagement patterns:

Understanding Developer Productivity Impact

Measuring developer productivity isn't straightforward, but Claude Code provides several useful metrics for tracking impact on your development workflows.

Core Productivity Metrics

Additional Available Metrics

Claude Code provides these additional telemetry metrics:

Combining Metrics for Insights

While Claude Code provides specific telemetry, you can combine this with your existing development metrics:

• Session count + PR count: Measure conversion from Claude sessions to completed work
• Token usage patterns + commit frequency: Understand if high token usage correlates with code quality
• Code edit tool decisions: Track developer confidence in Claude's suggestions
• Cost per PR: Calculate by dividing total costs by PR count over time periods

Key insight: The built-in metrics work best when combined with your existing Git and project management data.

Based on Anthropic's research on software development impact:

• 79% of Claude Code conversations are automation tasks (vs 49% on Claude.ai)
• 35.8% are feedback loop interactions - iterative development and debugging
• 43.8% are directive conversations - direct task completion
• JavaScript/TypeScript dominates usage at 31% of queries, followed by HTML/CSS (28%) and Python (14%)

Important Note: The metrics and examples in this guide are generated from limited sample data and hypothetical scenarios. Development workflows vary significantly across:

• Team practices: Some teams prefer frequent small commits, others batch larger changes
• Project types: Frontend development patterns differ from backend API work
• Developer experience: Senior developers may use Claude for architecture review, while junior developers focus on implementation assistance
• Organizational culture: Code review practices, testing requirements, and deployment processes all influence usage patterns

Your organization should define what productivity means in your context. Claude Code provides the measurement tools - the interpretation depends on your team's goals and workflows.

Answering Sample Business Questions with Telemetry

Here are a few common questions and how to answer them. These examples demonstrate the insights you can extract from Claude Code telemetry specific to your organization:

Question 1: What's our ROI on Claude Code per developer?

1// Calculate cost-to-value ratio
2developerId = user.account_uuid
3totalCost = sum(claude_code.cost.usage) by (developerId)
4prCount = sum(claude_code.pull_request.count) by (developerId)
5commitCount = sum(claude_code.commit.count) by (developerId)
6linesAdded = sum(claude_code.lines_of_code.count{type="added"}) by (developerId)
7avgCostPerCommit = totalCost / commitCount
8avgCostPerLine = totalCost / linesAdded

Question 2: How does developer tenure correlate with effective Claude Code usage?

1// Assuming you have developer tenure data available in your analytics platform
2// Calculate tool acceptance rates per developer
3acceptanceRates = sum(claude_code.code_edit_tool.decision{decision="accept"}) by (developerId) /
4                 count(claude_code.code_edit_tool.decision) by (developerId)
5
6// Correlate with developer experience data
7// Experience categories: junior, mid, senior
8correlate(developerExperienceCategory, acceptanceRates)
9
10// Track progression over onboarding timeline
11timeSeriesAnalysis(acceptanceRates, timeWindow=90d, groupBy=experienceLevel)

Question 3: Should we switch to subscription plans instead of pay-as-you-go?

1// Calculate total cost under current pay-as-you-go model
2currentMonthlyCost = sum(claude_code.cost.usage) over last_30d
3
4// Project costs for different subscription tiers
5// Claude Pro: $20/mo per user with 1x capacity
6// Claude Max 5x: $100/mo per user with 5x capacity
7// Claude Max 20x: $200/mo per user with 20x capacity
8activeUsers = count(distinct user.account_uuid) where claude_code.session.count > 0
9proTierCost = activeUsers * 20
10maxTier5xCost = activeUsers * 100
11maxTier20xCost = activeUsers * 200
12
13// Analyze typical usage patterns to determine capacity needs
14avgTokensPerUserPerMonth = sum(claude_code.token.usage) by (user.account_uuid) / activeUsers
15// Determine if users would hit subscription limits based on usage patterns
16usersExceedingProTier = count(users where avgTokensPerUserPerMonth > proTierLimit)
17
18// Calculate projected cost savings or additional expense
19subscriptionSavings = currentMonthlyCost - recommendedTierCost

‍Question 4: Where are our developers getting stuck with Claude Code?

1// Detect problematic usage patterns
2problemPatterns = {
3  // Long sessions with no commits
4  longUnproductiveSessions = filter(
5    duration > 30min AND claude_code.commit.count == 0,
6    group by (developerId, sessionId)
7  ),
8  // High rejection rates
9  toolRejectionHotspots = filter(
10    claude_code.code_edit_tool.decision{decision="reject"} > 10,
11    group by (developerId, tool)
12  ),
13  // API errors
14  apiErrorPatterns = group by (error, model) {
15    count(claude_code.api_error)
16  }
17}

Question 5: How does Claude Code impact our MTTR for bug fixes?

1// Join bug resolution data with Claude Code usage
2bugData = loadFromJira("bugs_resolved.csv")
3claudeUsageForBugs = filter(
4  claude_code.user_prompt contains "bug" OR
5  claude_code.user_prompt contains "fix" OR
6  claude_code.session metadata.issue_id in bugData.issueIds,
7  group by (developerId, issue_id)
8)
9// Calculate MTTR with and without Claude assistance
10mttrWithClaude = avg(bugData.resolutionTime where bugData.issueId in claudeUsageForBugs.issue_id)
11mttrWithoutClaude = avg(bugData.resolutionTime where bugData.issueId not in claudeUsageForBugs.issue_id)
12improvementRatio = mttrWithoutClaude / mttrWithClaude

Using Claude Haiku for Telemetry Analysis

Claude Haiku can effectively analyze your telemetry data, particularly user prompts. Here's a technical implementation approach:

1// Extract user prompt data from OpenTelemetry logs
2// Note: requires setting OTEL_LOG_USER_PROMPTS=1 to capture actual prompt content
3userPromptEvents = query {
4  from: claude_code.user_prompt
5  select: [user.account_uuid, prompt, prompt_length, event.timestamp]
6  where: event.timestamp >= now() - 30d
7}
8
9// Process batches of prompts through Claude Haiku
10function analyzePromptPatterns(prompts) {
11  // Structure the analysis request
12  const request = {
13    model: "claude-3-5-haiku-20241022",
14    messages: [{
15      role: "user",
16      content: `Analyze these developer prompts and identify:
17      1. Common task categories (debugging, feature development, refactoring, etc.)
18      2. Prompt quality patterns (specificity, context provided, clarity)
19      3. Developer skill indicators in the prompts
20      4. Recommended prompt improvements for better results
21      5. Areas where developers need additional training
22
23      Prompts to analyze:
24      ${JSON.stringify(prompts)}
25
26      Format your response as structured JSON with the categories above.`
27    }]
28  };
29
30  // Process through Haiku (low-cost analysis)
31  return callClaudeAPI(request);
32}
33
34// Categorize developers by usage patterns
35developerSegments = analyzePromptPatterns(userPromptEvents)
36  .groupBy(user.account_uuid)
37  .calculateMetrics([
38    "avgPromptQuality",
39    "dominantTaskCategory",
40    "skillLevel",
41    "improvementAreas"
42  ]);
43
44// Generate customized training recommendations
45trainingRecommendations = developerSegments
46  .filter(segment => segment.improvementAreas.length > 0)
47  .map(segment => ({
48    userId: segment.user.account_uuid,
49    recommendations: segment.improvementAreas.map(area => trainingModules[area])
50  }));

This approach lets you cost-effectively analyze thousands of prompts using Haiku's lower pricing while extracting valuable insights to improve developer effectiveness.

Real-world telemetry snapshot:

• Total cost: $0.43 across 6 sessions (average $0.072 per session)
• Cache efficiency ratio: 39:1 (78k cache reads vs 2k cache creation)
• Model distribution: 99.9% Opus, 0.1% Haiku

Your actual usage patterns will depend on:

• Workflow complexity: Simple bug fixes vs architectural reviews
• Developer preferences: Some use "ultrathink" for complex analysis, others prefer iterative conversations
• Project phase: Initial development vs maintenance vs refactoring work
• Team collaboration patterns: Individual work vs pair programming with Claude

Use this telemetry data as input for your organization's specific value assessment framework.

Brownie Points: Automated Reporting

Here's where it gets fun. You can ask Claude to generate reports combining your telemetry data with project management metrics.

Setting up Linear MCP Integration

Claude Code has built-in MCP support. Set up the Linear MCP server using the CLI:

1# Add Linear MCP server to Claude Code
2claude mcp add linear -s user -- npx -y mcp-remote https://mcp.linear.app/sse
3
4# Verify it's working
5claude mcp list
6
7# Restart Claude Code to activate the integration
8Alternatively, you can configure it directly in your .claude.json file:
9{
10  "mcpServers": {
11    "linear": {
12      "type": "stdio",
13      "command": "npx",
14      "args": ["-y", "mcp-remote", "https://mcp.linear.app/sse"]
15    }
16  }
17}

Generating Automated Reports

We've created a comprehensive prompt template for generating reports. Here's a simplified example:

1claude -p "Using the Linear MCP, analyze our team's velocity for the last sprint and combine with these Claude Code metrics:
2
3{
4  \"claude_code_sessions\": 42,
5  \"total_cost\": 103.45,
6  \"avg_session_duration\": \"28.5 minutes\",
7  \"tool_acceptance_rates\": {\"Edit\": 0.81, \"MultiEdit\": 0.92}
8}
9
10Generate a comprehensive productivity report with visualizations."

Dynamic Metric Fetching Example:

1#!/bin/bash
2# fetch-claude-metrics.sh - Automate report generation with live data
3
4# Fetch metrics from Prometheus
5TOTAL_COST=$(curl -s "http://localhost:9090/api/v1/query?query=sum(claude_code_cost_usage_USD_total)" | jq -r '.data.result[0].value[1] // "0"')
6INPUT_TOKENS=$(curl -s "http://localhost:9090/api/v1/query?query=sum(claude_code_token_usage_tokens_total{type=\"input\"})" | jq -r '.data.result[0].value[1] // "0"')
7OUTPUT_TOKENS=$(curl -s "http://localhost:9090/api/v1/query?query=sum(claude_code_token_usage_tokens_total{type=\"output\"})" | jq -r '.data.result[0].value[1] // "0"')
8
9# Generate report with live data
10claude -p "Using the Linear MCP, analyze our team's velocity and combine with these live Claude Code metrics:
11
12{
13  \"total_cost\": $TOTAL_COST,
14  \"input_tokens\": $INPUT_TOKENS,
15  \"output_tokens\": $OUTPUT_TOKENS,
16  \"cost_per_token\": $(echo "$TOTAL_COST / ($INPUT_TOKENS + $OUTPUT_TOKENS)" | bc -l)
17}
18
19Generate a comprehensive productivity report with visualizations."

For the complete prompt with all metrics and detailed instructions, see report-generation-prompt.md.

Sample Report Output

Using the prompt template above, Claude generates comprehensive reports like this:

Executive Summary

This week's analysis shows a 17% improvement in commit velocity for teams using Claude Code effectively. Total cost was $103.45 across 15 active Linear issues.

Key Insights Generated

• Development velocity improved 17% for teams with high Claude Code adoption
• Average session duration of 28.5 minutes falls within optimal productivity range
• MultiEdit tool shows 92% acceptance rate but low usage - opportunity for training
• Bug fix tickets (KAS-10, KAS-14) would benefit from increased MultiEdit usage

What the Report Includes:

• Mermaid visualizations for tool usage, costs, and productivity metrics
• Real Linear ticket analysis (KAS-10 through KAS-15)
• Actionable recommendations based on actual usage patterns
• Cost breakdowns and ROI calculations

For the complete generated report with all visualizations and metrics, see sample-report-output.md.

Organization Deployment

Rolling out telemetry across your organization is straightforward with managed settings.

Managed Settings Configuration

Create a managed-settings.json file:

1{
2  "telemetry": {
3    "enabled": true,
4    "endpoint": "https://your-otel-collector.company.com:4317",
5    "headers": {
6      "Authorization": "Bearer ${OTEL_TOKEN}"
7    }
8  },
9  "exporters": {
10    "metrics": "otlp",
11    "logs": "otlp"
12  }
13}

Deploy via your MDM solution or configuration management tool. Users can't disable telemetry, ensuring consistent data collection.

For detailed deployment options, see the Claude Code deployment documentation.

Conclusion

Measuring Claude Code's ROI doesn't have to be complicated. The key steps are:

1. Enable telemetry immediately - Even console output gives you basic insights
2. Set up Prometheus for real measurement - Dashboards and queries unlock the full picture
3. Track both cost and productivity metrics - You need both sides of the ROI equation
4. Use automated reporting for regular reviews - Let Claude generate the insights from your data