> ## Documentation Index
> Fetch the complete documentation index at: https://kawax.biz/llms.txt
> Use this file to discover all available pages before exploring further.

# From Correct Code to Idiomatic Laravel: What's Next for Boost Benchmarks

> Based on the official Laravel blog post "AI agents pass tests. Can they write idiomatic Laravel?", this article explains how Boost Benchmarks is shifting from measuring test passes to measuring cost-per-token correctness and idiomatic Laravel style.

## Overview

On July 24, 2026, the official Laravel blog published [AI agents pass tests. Can they write idiomatic Laravel?](https://laravel.com/blog/idiomatic-laravel-ai-coding-agents). It describes a major shift in how **Boost Benchmarks**, the evaluation suite behind [Laravel Boost](https://laravel.com/docs/13.x/boost), assesses AI coding agents.

<Info>
  Boost Benchmarks is an evaluation suite that checks whether AI coding agents can correctly implement Laravel applications, using [Pest](https://pestphp.com/) tests and architecture tests.
</Info>

## Where things stood: 17 evals, nearly a perfect score

In the earlier post [which AI model is best for Laravel](https://laravel.com/blog/which-ai-model-is-best-for-laravel), leading models cleared 16 of 17 evals using Boost, with 99.4% test accuracy. Even so, the hardest tasks still tripped them up.

According to the new post, today's frontier models — GPT-5.6, Claude Fable 5 / Mythos 5, Gemini 3.x, and others — now clear those same 17 tasks at or near 100%. The question "can an AI agent write correct Laravel code?" has a clear answer: yes.

This isn't unique to Boost. The entire coding-benchmark landscape is converging:

* **SWE-bench Verified** — leading models now sit within a point or two of the human-expert baseline, effectively saturated
* **HumanEval** — has long stopped differentiating between leading models

## Why passing tests isn't enough

The article also points out that a green test suite is an imperfect proxy for "good code."

* OpenAI [stopped reporting SWE-bench Verified](https://openai.com/index/why-we-no-longer-evaluate-swe-bench-verified/) after an audit found many problems had flawed test cases
* [Cursor's research](https://cursor.com/blog/reward-hacking-coding-benchmarks) showed models sometimes found answers via `.git` history rather than solving the problem, with scores dropping sharply once that history was sealed off

Boost Benchmarks is less exposed to these issues since every run starts from a fresh Laravel app with no answers tucked into `.git`. Still, passing the Pest suite doesn't prove a model writes Laravel the way an Artisan would. That gap between "it works" and "it belongs in this codebase" is the next frontier.

## The new axes: correct code per token, and idiomatic Laravel

The post lays out two new focus areas going forward.

<Steps>
  <Step title="Correct code per token">
    Once every model can reach the right answer, what matters is the cost to get there. The spread can be enormous — one model may spend an order of magnitude more tokens than another for nearly identical results. Boost already records tokens and cost per run, and this is becoming a first-class metric rather than a footnote.
  </Step>

  <Step title="Idiomatic Laravel">
    Code that passes tests but uses raw queries, skips `Route::resource()`, hand-rolls validation instead of using [form requests](https://laravel.com/docs/13.x/validation), or omits `$fillable` isn't "good Laravel." Reference-free LLM-as-judge scoring is now considered mature enough to evaluate "is this idiomatic?" without needing a golden answer to diff against.
  </Step>
</Steps>

The reference for this scoring is Boost's [best-practices skill](https://github.com/laravel/boost/blob/main/.ai/laravel/skill/laravel-best-practices/SKILL.md), which documents 19 concrete conventions, including:

* Using `Cache::remember()` instead of hand-rolled caching
* Using form requests instead of inline validation
* Using `with()` for eager loading
* Using `Route::resource()` for routes

Its very first rule is "Consistency First" — before reaching for a supposedly better pattern, match the style already used in sibling files.

## What's planned for Boost Benchmarks

The article outlines several upcoming additions:

* **Token and cost as a headline metric**, reported alongside pass rate for each eval and model
* **Idiomatic scoring**, layered on top of existing Pest and architecture checks
* **Context tightening** as a core engineering focus for Boost itself — trimming redundant guidelines, tightening tool outputs, and indexing only what's actually needed

The existing 17 evals remain in place as a regression suite, with this new layer built on top.

## Summary

The question of whether AI agents can write correct Laravel code has been answered. The next question is how efficiently, and how idiomatically, they can do it. If you haven't tried Laravel Boost yet, check the [official documentation](https://laravel.com/docs/13.x/boost) to get started, or see the [Laravel Boost](/en/boost) page for installation details.

<Card title="Read the original article" icon="link" href="https://laravel.com/blog/idiomatic-laravel-ai-coding-agents">
  AI agents pass tests. Can they write idiomatic Laravel? (laravel.com)
</Card>


## Related topics

- [What is Laravel](/en/introduction.md)
- [Laravel and AI development](/en/ai.md)
- [Laravel Boost](/en/boost.md)
- [Introducing the Blaze package](/en/blog/blaze-introduction.md)
- [Task scheduling](/en/scheduling.md)