AI Search

Perplexity Cuts Agent API Costs and Adds an Ultra Effort Level

Perplexity announced on November 5, 2026 that every effort level of its Agent API now costs less per query, with research quality higher or unchanged. The medium effort level improved from 55% to 77% on the BrowseComp and DSQA research benchmarks while its cost fell almost 90%. A new ultra level sits at the top of the range for broad research, running three research agents in parallel plus a fourth that merges their findings. CEO Aravind Srinivas summed up the pitch on X the same day: "Every month is a new Pareto optimal curve for Perplexity."

Cheaper at every level, better in the middle

The Agent API lets developers send a research task to Perplexity and get back an answer built from live web sources. The effort level controls how much work the API does: low for quick lookups, medium for ordinary research, high and xhigh for deeper work. Perplexity says every level now costs less per query than it did before, and that quality went up or stayed the same at each level.

The headline number is the medium level. It improved from 55% to 77% on BrowseComp and DSQA — two benchmarks that test research quality — while its cost fell almost 90%. That is the combination every API buyer wants: a better answer for a tenth of the price. Perplexity attributes the improvement to a more efficient research pipeline, though it has not published a detailed breakdown of what changed inside it.

At the xhigh level, the WANDR score — Perplexity's own benchmark for long research tasks — held at 76.5% while the configuration ran at about 25% of its August 2026 cost. In other words, the same quality for a quarter of the price. For developers running research at scale, that kind of cost reduction changes what is worth automating: tasks that were too expensive to run on every item in a list become affordable.

The new ultra level

Ultra is the new top tier, built for broad research where a single agent is not enough. It runs three research agents in parallel, each exploring a different angle, plus a fourth agent that merges their findings into a single answer. The model behind it is anthropic/claude-opus-5-5, one of the most capable models available, and the price is $21.71 per task.

On WANDR, ultra scored 85.9% — the highest figure Perplexity has reported for any configuration. That is a strong number, but it comes with two caveats. The first is cost: at $21.71 per task, ultra is not a tool for routine lookups. It is for the research task where the answer is worth twenty dollars — a competitive analysis, a due-diligence check, a deep literature review. The second caveat is that WANDR is Perplexity's own benchmark, designed for long research, and the score has not been independently checked.

The design of ultra is itself interesting. Running three agents in parallel and merging their work is a bet that breadth beats depth: three researchers covering different ground will find more than one researcher going deep on a single path. The merger agent is the key piece — it has to resolve contradictions, drop duplicates and produce a coherent answer from three separate lines of inquiry. Whether that works as well in practice as it does on a benchmark is the question developers will answer by trying it.

More context per search, cheaper fast search

Two smaller changes matter for developers building on the API. Each search can now return up to 4,000 tokens of results rather than 2,000, doubling the amount of source material the model can work with in a single call. That means fewer follow-up searches to fill gaps, and richer answers for questions that need several sources to answer properly.

Fast Search, the cheapest tier for simple lookups, is priced at $1.00 per 1,000 requests. That is a rounding error for most applications, and it makes Perplexity's API cheap enough to use as a background enrichment step — checking a fact, fetching a price, verifying a claim — rather than only as the main event in a research workflow.

Why it matters for SEOs and link builders

For SEOs, the practical effect is that automated research gets cheaper and better. Competitor monitoring, content gap analysis, citation checking, and brand mention research are all tasks that an agent API can now do at a price that works for a small agency, not just an enterprise. The medium level's improvement to 77% on research benchmarks, at a tenth of the old cost, is the kind of change that turns a manual weekly task into an automated daily one.

For link builders, the more important shift is on the other side of the same coin. The AI tools that cite and recommend content are getting better at research, which means they are getting better at telling a strong source from a weak one. A page that earns a citation from an agent that runs three parallel researchers and merges their findings has to survive three independent checks, not one. Thin content, unsupported claims and borrowed statistics will fail that test more often than they pass it.

Sources

Sources: Perplexity blog, November 5, 2026; Unite.AI. Note: the benchmark scores in this article are Perplexity's own and have not been independently checked.

Related reading on this site: Claude Climbs to No. 25 as Perplexity Drops to No. 392 in Similarweb's October Rankings, What Gets Cited by ChatGPT: Ahrefs' 1.4 Million-Prompt Study, Three AI Models Shopped Shopify's Catalog and Each Picked a Different Winner, Google CEO Pichai Sees Search Absorbed Into an Agentic Layer, Google, Amazon, Microsoft and Meta Could Dominate AI Agents, Open Markets Warns, and Anthropic Reportedly Moving Ahead With IPO Plans.

More from Linkbaits

Turn this playbook into links

Linkbaits helps you create content that earns citations and backlinks. Start with the toolkit and build your first linkable asset this week.

Start free trial