Prompt caching — up to 90% off

Repeated context is cached across calls, cutting cached-input costs by up to 90% on supported models.