Perplexity announced lower per-query costs across Agent API effort levels and introduced Ultra on November 5, 2026. Its cost and benchmark figures describe different configurations and units, including per-query experiments, a WANDR task and individual search requests. Perplexity’s announcement reports a specific step-budget test that makes one cost-quality tradeoff concrete, but the figures are not a single price schedule for every preset.

Keep the counters separate: an Agent API query, a WANDR task and a Fast Search request are different units.

What changed in the Agent API

Perplexity’s effort levels bundle a model, step budget, search settings and tools for a type of task. The company said the levels cost less per query after its updates, with research quality higher or unchanged. Selecting an effort level by name applies its updated configuration automatically.

The new ultra level is designed for broad research. Perplexity says it runs three agents in parallel, each researching the task, then uses a fourth to combine their findings, remove duplicates and write the answer.

For a separate part of the developer workflow, NeoTeo has covered how Perplexity’s Brain memory carries stored context into later tasks.

The reported figures measure different things

Perplexity reported that medium improved from 55% to 77% on BrowseComp and DSQA while its costs fell almost 90%. For xhigh, the company reported a 76.5% WANDR score at about 25% of the cost of its August 2026 configuration.

For ultra, Perplexity reported an 85.9% WANDR score and a cost of $21.71 per task, the highest score among the configurations it tested. That figure is tied to a WANDR task; the announcement separately reports query-level costs for other tests.

A further high-effort example compared extraction with no extraction. Perplexity reported about 50% fewer input tokens per query with extraction, alongside an 88.0% benchmark result at $0.073 per query. Without extraction, it reported 83.5% at $0.094 per query. These are results for that comparison, not a common price scale across all effort levels.

One step-budget example shows the cost tradeoff

In a Perplexity-reported openai/gpt-6-luna experiment, raising the permitted steps from 1 to 15 increased accuracy on BrowseComp and DSQA from 41% to 80%. Cost rose from $0.0017 to $0.0152 per query.

That is an increase of $0.0135 per query. The 15-step run cost about 8.9 times as much, while the reported accuracy value was about 1.95 times as high, a difference of 39 percentage points. Those calculations describe this experiment alone.

A step is not necessarily one tool call: Perplexity says a single step can issue multiple calls. Its effort levels set both step limits and a cap on total tool calls, so the step count by itself does not capture every action in a run.

Fast Search has a per-request rate

Perplexity stated a Fast Search rate of $1.00 per 1,000 requests. That amount is for search calls; the other examples describe Agent API query or task costs.