← Voltar aos guias

Build Cheap Automated PR Reviews with LLMs

Why automate pull request reviews?

Pull request review is one of the best places to use LLMs because the work is repetitive, structured, and expensive in human time. A typical 400-line PR can take 10 to 20 minutes for a senior engineer to scan, think through edge cases, and leave comments. At an $80/hour internal cost, that is roughly $13.33 to $26.67 per review. If a second reviewer is required, the cost doubles fast.

An LLM should not replace engineers. It should remove the first layer of effort: catching obvious bugs, missing tests, unsafe patterns, weak naming, and likely regressions before a human spends time on them. The result is faster feedback, fewer reviewer bottlenecks, and better use of senior engineer attention.

A cost-efficient review pipeline

The cheapest way to automate PR review is to keep the model context small and the workflow selective. Do not send your entire repository. Send only what the model needs to make a useful judgment.

• Trigger the bot on pull request open, reopen, and new commits.
• Collect the diff, file list, PR title, PR description, and any failing test or lint output.
• Use a small model for the first pass. For many repos, 2,000 to 4,000 input tokens and 200 to 400 output tokens is enough.
• Escalate only complex, risky, or ambiguous PRs to a stronger model.
• Post comments only when confidence is high or the finding is severe.

Here is the budget math. If a first-pass review uses 3,000 input tokens and 500 output tokens, then even a model priced at $5 per million input tokens and $15 per million output tokens costs only $0.0225 per PR. At 100 PRs per month, that is about $2.25. Compare that with a human reviewer spending just 15 minutes per PR at $80/hour: $20 per PR, or $2,000 per month for 100 reviews.

How to keep quality high without spending more

Cheap does not mean sloppy. The biggest cost and quality wins come from better prompting and tighter context, not from using a bigger model everywhere.

• Ask for line-specific feedback. Require the model to cite file paths and line numbers from the diff.
• Limit context aggressively. Large monorepo dumps waste tokens and create noise.
• Run static analysis first, then pass the results to the LLM so it can focus on reasoning, not detection.
• Use a two-model strategy: Haiku or Fable for triage, Sonnet for deeper review when the PR is large or risky.
• Set a hard rule that the bot must say needs human review instead of guessing when the evidence is weak.

A practical threshold is to auto-comment only on medium or high severity findings. That keeps the bot useful and reduces false positives, which is critical for adoption. If developers trust the comments, they will keep reading them. If not, the bot becomes expensive noise.

Why 59API is a smart budget choice

59API is a low-cost, pay-as-you-go AI API relay that gives developers access to Claude and GPT models, including Claude Opus, Sonnet, Haiku, Fable, and GPT variants. It is fully compatible with Claude Code, Codex, and any OpenAI SDK, so you can wire it into an existing review bot without rebuilding your stack. In practice, that means you can switch the base URL to https://api.59api.com and keep your current integration logic.

For automated PR review, that compatibility matters because the best optimization is often not a model change but a workflow change. Use the cheapest model that can do the first pass well, then spend more only when the PR deserves it. 59API is a strong fit because it keeps the unit economics low while still giving you native official-quality models, with no downgrade. If your team reviews code every day, that difference compounds quickly. The referral rebate also helps reduce the effective cost over time.

Implementation checklist

• Build a webhook service for GitHub or GitLab.
• Trim the diff to only changed files and relevant hunks.
• Include tests, lint output, and the PR description in the prompt.
• Store bot comments and human feedback to measure false positives.
• Track three numbers: average tokens per PR, cost per PR, and review time saved.

If you want faster code review without bloating your cloud bill, start with one repository and a small monthly token budget. Sign up for 59API, point your client at the relay, and pilot automated PR review on your busiest service first. The quickest savings usually come from automating the first 10 percent of the review workflow.

Preços · Docs