The Agent’s Tab: What 600 AI Coding Runs Reveal About Ruby vs Python
Claude Code just ran the same task 600 times across 13 languages — and quietly settled an old argument. Run it yourself below, then step through the AI-governed release pipeline that ships the result.
01 Run the benchmark
In April 2026, developer mame’s ai-coding-lang-bench had Claude Code (Opus 4.6) build a mini Git implementation from scratch, 40 times, in 13 different languages — then InfoQ and others picked it apart. Press play to replay a compressed version of that run.
| Language ▼ | Avg time | Avg cost | Lines (v2) | Tests |
|---|
02 What should you pick?
The “right” language depends on what you’re optimizing for. Drag the sliders to weight what matters for your team’s AI-assisted workflow — the recommendation below recalculates from the real benchmark deltas (cost, speed, and the type-checker tax).
03 Where the code actually goes
Writing the code is the cheap part now. Here’s what happens to an AI agent’s commit on its way to production in a 2026-grade pipeline — click through each stage.
04 One config to steal
GitHub Actions made Node 24 the default runtime for JS actions this quarter, with no opt-out. If an AI agent is opening PRs against your workflows, pin this now so an unmaintained marketplace action doesn’t quietly break your pipeline.
# pin runtime + action versions so agent-authored PRs # can't silently break CI when upstream actions move jobs: build: runs-on: ubuntu-latest steps: - uses: actions/checkout@v5 - uses: actions/setup-node@v5 with: node-version: '24' # explicit — don't inherit the new default blind - run: npm ci --ignore-scripts - run: npm audit --audit-level=high
If an AI agent is writing most of your code now, the deciding cost isn’t “which language do I like” — it’s which language lets the agent iterate fastest without a type-checker tax, and whether your pipeline can review and gate that output before it ships. On this benchmark, Ruby and Python are statistically tied and both beat typed languages on cost; the real gap opened up at the release stage, not the coding stage.