Perplexity Cuts Agent API Effort-Level Costs and Adds Ultra: Medium Nearly 90% Cheaper; Ultra ~85.9% WANDR

Core Background & Industry Pain Points

Research Agent APIs often bundle model, search and tool budgets into depth tiers, but higher tiers can be costly and hard to tune. On November 5 Perplexity said every Agent API effort level now costs less per query with equal or better research quality, and added an ultra level for broad, long-running coverage.

Architecture Highlights & Internals

Effort levels are named presets (model, step/tool budgets, search behavior). Calling by name (e.g. effort="medium") picks up the revised config automatically. Officially, ultra runs three parallel research agents then a fourth consolidator; background mode is supported and it runs on anthropic/claude-opus-5-5. Related changes include budget notices, context compaction and fetch_url extraction to cut unused context.

Authoritative Benchmarks & Measured Scores

Vendor-reported (via official blog / Unite.AI summary; not AICoder re-runs): medium rose from about 55% to about 77% on BrowseComp and DSQA with nearly 90% lower cost; xhigh held about 76.5% on WANDR at roughly 25% of its August 2026 cost; new ultra scored about 85.9% on WANDR at about $21.71 per task, highest among configs they compared. Changelog lists [email protected].

Developer Hands-on Guide

Set Agent API effort to fast|low|medium|high|xhigh|ultra; pin a version (e.g. [email protected]) for reproducibility. Prefer background for long runs. See Effort docs and Changelog. Note: WANDR/BrowseComp figures are vendor-reported; ultra has no full frozen schema—call by name or pin.