Perplexity announced an update to every Agent API effort level on November 5, 2026, saying research quality is equal or better at a lower cost per query. It also introduced an Ultra level for broader research using several agents. The announcement concerns the configurations behind the service, rather than a new downloadable model.
An effort level is a complete research configuration
Perplexity describes each level as a bundle of model selection, search settings, tools, and research budget. The update includes changes to step and tool-call limits, context handling, source extraction, and workflow selection. Its performance comparisons are vendor evaluations under the stated tasks and settings.
Requests selecting a level by name receive the new configuration automatically. Developers who need a stable comparison can pin a version. The November 6 changelog confirms the update and describes Ultra as running three agents in parallel before combining their findings.
The cost unit is a completed research request
Our engineering analysis is that a lower inference price and a lower research cost are different outcomes. A research system may save money by reading fewer irrelevant passages, avoiding repeated searches, or choosing a shorter workflow for a simple question. Those changes can matter even when the underlying model is unchanged.
A useful evaluation should therefore compare complete requests, including failures and retries. Keep the questions, evidence cutoff, expected outputs, and verification criteria fixed. Record whether the answer is supported and complete, how long it takes, and what the whole request costs.
Broader research needs an evidence handoff
Parallel agents can investigate different parts of a question, but additional sessions do not automatically produce additional coverage. They may duplicate the same searches or repeat a shared misunderstanding. The final synthesis should identify which parts were checked and resolve conflicting evidence explicitly.
Test a research question with several independent dimensions and at least one misleading source. Review whether the system finds the relevant primary material and preserves uncertainty where the record is incomplete. A long answer with many links should not receive credit merely for looking comprehensive.
Manage automatic updates deliberately
A named configuration is convenient when a team wants improvements as they arrive. A pinned configuration is useful when reproducing an evaluation or diagnosing a change. Decide which behavior the application needs and retain the configuration identifier with the result.
After an update, sample ordinary production-shaped requests before revising a budget or quality target. Compare a mix of simple lookups, multi-step questions, and broad investigations. Aggregate scores can hide a regression in the category most important to a particular customer.
The November release makes research orchestration a visible part of the API’s value proposition. Its significance is the attempt to improve the whole path from question to supported answer, with cost and quality assessed together rather than inferred from one model benchmark.