How to Configure Laravel Servers for Long-Running LLM API Calls

Laravel 13, released in March 2026, introduced first-party AI primitives, making it easier to integrate large language model features into web applications. However, default server configurations are poorly suited for LLM workloads, where requests can last 30 to 120 seconds and responses arrive as token streams rather than single payloads. Key bottlenecks include PHP's max_execution_time, Nginx timeout and buffering settings, and Laravel's HTTP client timeout, all of which can cause failures or unexpected costs during AI calls. A particularly costly issue involves queue retry_after settings, where a long-running job can be handed to a second worker and billed twice before the first completes. Developers are advised to offload LLM calls to queued jobs rather than handling them synchronously, preventing FPM worker exhaustion and keeping web requests fast.