swe-cost-estimator ⚡

Instant, zero-dependency CLI to estimate multi-turn inference costs and effective cost-per-resolved-issue for autonomous software engineering agents across leading frontier AI models.

Why This Exists

Most AI pricing comparisons only measure price-per-token ($/1M tokens). But in autonomous software engineering, accuracy determines the true cost.

If a cheaper model fails 40% of tasks, you pay for repeated retry loops, failed test runs, and developer review overhead.

$$\text{Effective Cost per Solved Issue} = \frac{\text{Total Multi-Turn Inference Cost}}{\text{DeepSWE v1.1 Pass Rate}}$$

This tool scans your repository context and simulates realistic agentic workloads over complex SWE benchmarks.

Interactive Web Calculator & 1-Click Presets

Prefer a visual simulator? Try the live interactive calculator with 1-click team presets: 👉 https://optimist29.github.io/swe-cost-estimator/

• 📦 Small Service / MVP: 10k LOC (~30k tokens), 20 issues, 3 turns
• 🚀 Active Team Repo: 40k LOC (~120k tokens), 100 issues, 5 turns, git rebase contention
• 🏢 Enterprise Monorepo: 100k+ LOC (~350k tokens), 250 issues, 8 turns, heavy branch conflict churn

Quickstart (Zero Install)

Run directly in any code repository using npx:

npx swe-cost-estimator

The CLI automatically:

1. Scans codebase size and estimates total token context (skipping .git, node_modules, lockfiles).
2. Inspects 14-day Git commit velocity to dynamically calculate team rebase contention & rework multiplier.
3. Auto-detects local agent session logs (~/.claude) to extract empirical turn counts.

CLI Options & Customization

npx swe-cost-estimator [path] [options]

| Flag | Description | Default | | :--- | :--- | :--- | | [path] | Target directory or codebase to scan | . (current directory) | | --issues <num> | Number of tasks or issues to simulate | 100 | | --turns <num> | Average agent iterations/turns per issue | 4 (or auto-detected from local sessions) | | --rebase-factor <float> | Multiplier for branch conflicts and rework | Auto-detected from 14-day git churn (1.0x–1.45x) | | --cache-ratio <float> | Context caching fraction per turn | 0.4 (40%) | | --output-tokens <num> | Output tokens per turn (reasoning + git diff) | 3500 | | --no-churn | Disable automatic git commit velocity detection | false | | --json | Output raw JSON data for CI/CD or scripting | false | | -h, --help | Show help and options | |

Examples

# Auto-detect context, git churn contention, and local session turns:
npx swe-cost-estimator

# Simulate a sprint of 50 complex issues with custom turns:
npx swe-cost-estimator . --issues 50 --turns 6

# Explicitly model a high-contention team repo (1.5x rework multiplier):
npx swe-cost-estimator . --rebase-factor 1.5

# Pipe JSON into jq for CI/CD cost auditing:
npx swe-cost-estimator . --issues 100 --json | jq .

Empirical Verification (EXPERIMENT.md)

Want to see how these benchmarks hold up against real agent CLI executions? Check out EXPERIMENT.md for our reproducible "State X to State Y" protocol measuring end-to-end task cost on real bugs across Claude Code CLI and Gemini Flash.

Supported Models & Benchmark Data

All benchmarks are sourced from verified evaluations (DeepSWE v1.1):

| Model | Input / 1M | Output / 1M | DeepSWE v1.1 | Terminal-bench 2.1 | Context Window | | :--- | :--- | :--- | :--- | :--- | :--- | | Gemini 3.8 Flash | $0.75 | $3.75 | 73.7% | 89.4% | 1,048,576 | | Claude Opus 5 | $5.00 | $25.00 | 74.0% | 89.1% | 200,000 | | GPT-5.6 Sol | $5.00 | $30.00 | 73.0% | 88.8% | 1,100,000 | | Claude Sonnet 5 | $2.00 | $10.00 | 54.0% | 80.4% | 1,000,000 |

License

MIT © Praveen