← Retour aux guides

Cut Pull-Request Review Costs with LLM Automation

Why LLM Pull Request Reviews Need Cost Controls

LLMs can catch missing null checks, unsafe SQL construction, broken edge cases, inconsistent tests, and undocumented behavior before a human reviewer opens a pull request. The risk is cost creep: sending every changed file, dependency lockfile, generated asset, and full repository convention document to a premium model can turn a useful reviewer into an expensive one.

A cost-efficient system reviews the changed code and relevant context, routes simple work to a fast model, and escalates only high-risk pull requests. 59API is a practical option for this workflow because it provides pay-as-you-go access to native official-quality Claude models, including Opus, Sonnet, Haiku, and Fable, plus GPT models. Its OpenAI-compatible endpoint at https://api.59api.com also means an existing OpenAI SDK integration can usually be adapted by changing the base URL and API key rather than rebuilding the reviewer.

Set a Token Budget Before You Automate

Start with a hard budget per pull request. For routine reviews, target 8,000 input tokens and 800 output tokens. That is usually enough for a focused diff, a short repository guide, and structured findings. Reserve 20,000 to 40,000 input tokens for security-sensitive or architecture-changing pull requests.

Use this simple planning formula: monthly review cost = pull requests per month × average input tokens × input rate + pull requests per month × average output tokens × output rate. For example, a team processing 600 pull requests monthly at 8,000 input tokens and 800 output tokens consumes 4.8 million input tokens and 480,000 output tokens. If a selected model costs $1 per million input tokens and $5 per million output tokens, the estimated monthly total is $7.20: $4.80 for input and $2.40 for output. Check the current 59API model rate before setting production budgets, since pricing varies by model.

Send Better Context, Not More Context

Do not submit an entire repository by default. Build a review payload from the pull request title, description, changed-file list, unified diff, and a compact engineering policy of 300 to 600 tokens. Exclude vendor directories, package lockfiles, minified files, snapshots, generated clients, binaries, and files larger than 100 KB unless they are directly relevant to the risk being reviewed.

• Cap the diff at 12,000 to 15,000 tokens for the first pass.
• For a changed function, include up to 150 lines before and after the edit when needed for control-flow context.
• Retrieve only two to five related symbols, tests, or API contracts instead of attaching whole files.
• Ask the model to return no more than five findings, reducing output tokens and reviewer noise.
• Skip review when a pull request changes only documentation, comments, or approved generated output.

These rules often cut prompt size by 50% or more without sacrificing useful findings. They also reduce latency, which matters when a review comment must appear while the author is still working on the branch.

Route Models by Risk and Expected Value

Use a low-cost, fast model such as Claude Haiku for the first pass on ordinary pull requests. Ask it to classify the change as low, medium, or high risk and identify files requiring deeper inspection. Escalate to a stronger model such as Sonnet only when the classifier detects authentication, authorization, payments, database migrations, public API changes, concurrency, cryptography, or a diff larger than 500 changed lines.

A concrete routing policy might send 75% of pull requests to the fast model, 20% to Sonnet, and 5% to Opus for complex security or cross-service changes. If your premium model is five times the cost of the fast model, this routing approach is dramatically cheaper than using the premium model on 100% of reviews. Use Opus as an exception path, not the default.

Make Output Actionable and Cheap

Require structured output with severity, file path, line range, explanation, and suggested fix. Tell the model: report only issues that are likely to cause a bug, security flaw, data loss, or meaningful maintenance problem; do not praise code or restate the diff. This avoids long, low-value prose and makes it easier for your GitHub or GitLab bot to post inline comments.

Measure acceptance rate weekly. If fewer than 20% of AI findings lead to code changes, tighten the prompt or reduce review scope. Track cost per accepted finding, not just cost per pull request. For example, $30 per month that prevents 12 confirmed defects costs $2.50 per accepted finding; that is a much more useful operating metric than raw token volume.

Deploy with a Safe, Incremental Rollout

Run the bot in comment-only mode for two weeks. Do not block merges initially. Compare its findings with human review outcomes, tune false-positive rules, then enable merge-blocking only for high-confidence critical findings. Store no secrets in prompts, redact environment variables, and restrict the bot token to read pull requests and write review comments.

Because 59API works with Claude Code, Codex, and OpenAI SDK-compatible applications, teams can test model routing without locking themselves into a new client stack. If you want lower-cost access to capable models for an automated review pipeline, sign up for 59API and begin with a capped budget and one repository.

Tarifs · Docs