What happened
Perplexity’s update changes the bundled configurations behind every effort level in its Agent API. The company describes each level as a package that sets the model, step budget, search settings and tools for a particular type of task. Requests that select a level by name, such as effort="fast", receive the updated configuration automatically. Existing integrations therefore do not need to change their effort names to pick up the new settings.
The post says the revisions combine lower costs with research quality that is higher or unchanged. Perplexity attributes the changes to two broad sets of adjustments: budgets, context handling and extraction reduced the cost of a run, while skills, model selection and parallel agents improved or maintained answer quality. It frames the update around a practical challenge in research agents: a fixed budget can force an early stop, while an unconstrained agent can keep taking costly actions that add little value and fill its context with irrelevant material.
For medium, Perplexity reports an improvement from 55% to 77% on BrowseComp and DSQA, with costs down almost 90%. It describes BrowseComp and DSQA as public research benchmarks. For xhigh, the company says its score on WANDR stayed at 76.5% while costs fell to about 25% of the August 2026 configuration. WANDR is Perplexity’s own benchmark for long, multi-entity research tasks, so this result is not an independent benchmark claim.
The update also introduces ultra, the highest effort level for the most wide-ranging and comprehensive research. The supplied announcement describes it as suited to coverage-critical tasks. It does not give a price for ultra or specify a maximum run length, request budget or access restriction. The announcement’s central operational promise is that effort presets are actively evaluated and updated over time to deliver research at or below their existing cost; the evidence supplied here is the company’s account of this update.
What the archive adds
This is a new fact on a known story, not a first look at effort economics. LlamaIndex’s max-effort setting and compute-based pricing showed another vendor exposing higher effort with variable compute costs. Our earlier note on choosing effort levels covered the risk that extra reasoning can cost more without helping. Perplexity’s update adds vendor-reported evidence of a cheaper preset and an additional top tier.
Trend signal
This supports the pattern that agent economics is measured at the task level, not just by token price. Perplexity says it lowered per-run costs by changing budgets, context handling and extraction while maintaining or improving research results. That is consistent with agent economics moving to completed work, although the announcement does not provide independently verified cost per completed task.
Why now
Perplexity presents the changes as the outcome of ongoing evaluation and tuning of preset configurations. That makes the update relevant now to users who choose an effort level: named settings receive the new configuration automatically, so their existing calls may behave differently in cost or quality without an integration change.
Why it matters
Effort settings can hide important operational choices behind a short parameter. If the model, step budget, search behavior and tools are bundled together, a developer choosing medium is selecting a workflow as well as a reasoning level. Perplexity’s reported reductions suggest that the cost of a research agent may depend as much on controlling context and action budgets as on choosing a cheaper model.
For teams already using the API, the low-friction response is to remeasure representative tasks rather than assume the old cost-quality profile still applies. Compare total cost, completion rate, factual support and time to a usable answer—not just benchmark scores. The new ultra tier may suit investigations where missing a source is more costly than a longer run, but without its price and limits, teams should establish their own caps before applying it broadly.
What happens next
- Teams using named effort presets will retest cost and answer quality after the automatic configuration change (3 months, 70%). Watch for: published evaluations that separate workload-level costs from benchmark scores.
Limits and uncertainty
The numerical results come from Perplexity’s announcement. The supplied material does not provide independent replication, benchmark protocols, sample sizes, confidence intervals or task-level cost breakdowns. It gives no absolute dollar prices, and does not quantify savings for fast, low, high or other levels beyond the examples described. The WANDR result is on Perplexity’s own benchmark; results on BrowseComp and DSQA do not establish performance on a particular customer’s research tasks. The announcement also does not specify ultra’s price, access conditions, hard limits or how its quality compares with xhigh. Lower benchmark cost does not by itself show lower cost per verified, usable answer in production.
Skill to build now
Run a small before-and-after evaluation on representative research tasks. Record the named effort level, total API cost, completion time, sources returned and whether a reviewer accepts the answer. Include routine lookups and multi-source investigations, and set a spend ceiling before trying ultra. This will show whether the revised preset helps your workload rather than merely reproducing a vendor benchmark result.
Reader takeaway
Perplexity says existing Agent API effort names now map to cheaper, equally or more capable configurations, and adds ultra for broad research. Treat the reported gains as a reason to retest your workloads, not as a guarantee: the announcement does not establish your cost per accepted result or disclose ultra’s price and limits.