Perplexity cuts Agent API research costs and adds an Ultra effort level

Perplexity says its Agent API now costs less per query across every effort level, while reporting higher or unchanged research quality on its evaluations. The new Ultra preset runs multiple agents in parallel for broad research, giving developers a ready-made choice between speed, cost and depth rather than another afternoon spent tuning knobs.

Perplexity Watch analysis

What happened

The changes, announced on 5 November, cover the Agent API’s named effort presets. Perplexity says Medium’s scores on BrowseComp and DSQA rose from 55% to 77% as its cost fell by almost 90%. At xhigh, it reports a WANDR score of 76.5% at about a quarter of the August cost.

The new Ultra preset is aimed at broad research and data collection. Perplexity says three agents gather evidence independently, then a fourth consolidates their results. It reports an 85.9% WANDR score at $21.71 per task. The Unite.AI account of the changes also describes features including adjustable step and tool-call budgets, context compaction, and extracting relevant passages from web pages.

Why it matters

For developers, these presets bundle model choice, tool access and research limits into a selectable setting. Lower reported costs could make longer research runs more practical, while Ultra offers a route to wider coverage without building an orchestration system from scratch. The trade-off is that higher effort can mean more time and spending, not a free lunch with a JSON wrapper.

The performance figures are Perplexity’s reported benchmark results, not a guarantee that every application will improve by the same amount. Its own account notes that a model which scored well on public tests performed worse on some internal enterprise, legal and finance tasks.

Our read

This is a meaningful API change because it combines lower stated costs with a new way to buy more research depth. Developers should compare the presets on their own tasks, including latency and evidence quality, before treating a benchmark win as a budget forecast. The appeal is real; the receipts still need checking against the work that matters to you.

What to watch

• Whether developers can reproduce the reported quality and cost changes on their own workloads.
• How Ultra’s parallel-agent approach affects response time and cost in practice.
• Whether later model or tool changes alter the balance between dynamic presets and pinned configurations.

Discussion spark: Would you trust a multi-agent preset to do broad research for your application, or would you rather control each model, tool and spending limit yourself?

Independent WittyWires tracker for public updates about Perplexity. Not affiliated with or endorsed by Perplexity; this is not an official account.