How do new agent effort tiers and URL fetching costs impact grounded AI development?

Perplexity updated its API changelog to include direct URL fetching capabilities and new agent effort levels. These changes allow developers to control how the system retrieves and processes web content before generating answers.Perplexity added features to let developers fetch specific web pages and control agent complexity.The API now supports direct URL fetching with web grounding and expanded effort level configurations. The update also clarifies pricing for these additional retrieval steps.

The announcement

The Decisions API allows passing a page URL as state when mode is set to web_grounding. The fetched text replaces the URL before search. The response provides the retrieved page in fetched_results and search evidence in search_results. Developers can configure max_staleness to specify the maximum age of a page in seconds since its last crawl. If the stored copy is older, the system triggers a fresh crawl. Each successful URL fetch costs $0.001, separate from search and input token fees.reported

The context

The Agent API effort levels received significant updates. The fast tier defaults to openai/gpt-5.6-luna when openai/gpt-6-luna is not available.reported Low and medium efforts increased max output tokens to 128,000 and raised web_search token limits to 4,000.reported The high effort level uses openai/gpt-6.1-terra, increasing max steps to 40.reported The xhigh setting retains anthropic/claude-opus-5-5 as its primary model, with openai/gpt-6.1-sol serving as the fallback, while increasing the maximum step count to 300.reported A new ultra-preview effort level targets broad research tasks.

Reading between the lines

The API renamed the preset parameter to effort, though preset remains a deprecated alias. Responses now return both fields to ease migration. The undocumented ultra effort name is removed and now returns an error. The wide-research alias now selects xhigh. The maximum value for max_steps has increased from 100 to 500.reported Users of frozen configurations should update their values to align with the public API fields for each effort level.

The business case

Platform architects must account for the $0.001 per fetch cost when designing grounded workflows.reported The increased step limits in high and xhigh tiers allow for more complex reasoning chains, potentially increasing latency and token usage. The removal of the undocumented ultra effort simplifies versioning but requires code updates. The fallback models in fast and xhigh tiers provide resilience but may alter output characteristics.