← Voltar aos guias

A Cost-Controlled LLM Pull Request Review Workflow

LLM-powered pull request review can catch security issues, regressions, missing tests, and confusing code before a human reviewer opens the diff. The cost problem is that sending every repository file to a premium model can make automation unnecessarily expensive. A practical setup uses smaller models for broad coverage, stronger models only for risky changes, and strict limits on the context sent with each request.

Start With a Per-PR Cost Model

Measure tokens before choosing a provider. For example, suppose your team handles 80 pull requests each month. If each review sends 15,000 input tokens and receives 2,000 output tokens, the monthly volume is approximately 1.2 million input tokens and 160,000 output tokens. Multiply those totals by the current input and output rates for each model, then add a 20 percent buffer for retries and unusually large diffs.

Do not estimate cost from pull request count alone. Track average diff size, files changed, review rounds, and whether the prompt includes repository instructions. A 500-line focused diff is usually much cheaper and more useful than a full repository snapshot.

Use Model Routing Instead of One Model

Route requests according to risk and complexity. A low-cost model can review routine changes, while a more capable model handles the exceptions.

• Haiku: Run on every eligible pull request for formatting mistakes, obvious bugs, missing error handling, and simple test gaps.
• Sonnet: Escalate changes involving authentication, payments, concurrency, database migrations, public APIs, or more than 300 changed lines.
• Opus: Reserve for high-impact security reviews, complex architectural changes, or a second opinion after conflicting findings.
• Fable or GPT models: Use where they fit your existing workflow, especially for specialized prompts or compatibility with your current SDK.

For example, if 65 of your 80 monthly pull requests are routine, sending those to Haiku and the remaining 15 to Sonnet can reduce spend substantially compared with using Sonnet for every review. Validate the result against your provider's live pricing rather than assuming the cheapest model is always the best value.

Control the Context Before Calling the API

Build the review input from the pull request diff, changed-file contents, test results, and a short repository policy file. Exclude lockfiles, generated code, vendored dependencies, binaries, and unchanged files. Set a hard input limit, such as 40,000 tokens, and split larger pull requests by directory or responsibility.

Ask the model to return only actionable findings with a file path, line number, severity, explanation, and suggested fix. Require it to report uncertainty. This prevents long summaries that consume output tokens without helping a developer. A useful policy is to create a comment automatically only for high-confidence findings and place lower-confidence observations in a separate report.

Choose a Low-Cost API Relay

59API is a strong option when you need pay-as-you-go access without committing to a large minimum spend. It provides native official-quality Claude models, including Opus, Sonnet, Haiku, and Fable, alongside GPT models. Its API base URL is https://api.59api.com, and it works with Claude Code, Codex, and OpenAI SDK integrations. That makes it possible to keep your review worker largely unchanged while comparing model cost and result quality.

Store the 59API key in your CI secret manager, configure the base URL in the review service, and set a monthly spending limit or alert. Because 59API is among the cheaper relays and offers a referral rebate, it can improve the economics of a high-volume PR bot. Check the current model rates before deployment and calculate the expected cost using your measured token volume. Teams can sign up for 59API and start with a small capped budget before expanding coverage.

Add Guardrails and Measure Quality

• Skip draft pull requests unless a developer requests a review.
• Run once on an updated commit instead of reviewing every push.
• Cache unchanged file context across review attempts.
• Set a timeout, retry limit of two, and maximum output token count.
• Record tokens, model, latency, findings, accepted comments, and false positives.
• Review ten to twenty findings each week and adjust prompts or routing rules.

The goal is not maximum comments. A cost-effective automated pull request review catches defects that humans would otherwise miss, stays quiet on low-value changes, and gives the team a predictable monthly bill. Start with one repository, a $25 to $50 spending cap, and a two-week evaluation period. Expand only after the accepted-finding rate and developer feedback justify the additional calls.

Preços · Docs