Perplexity says the medium effort level of its Agent API now scores 77% on the BrowseComp and DSQA research benchmarks, up from 55%, while its cost fell almost 90%, according to a company blog post. Chief executive Aravind Srinivas shared it on X at 23:54 UTC on 5 November: “Every month is a new Pareto optimal curve for Perplexity.”
The post says “every effort level now costs less per query, with research quality that is higher or unchanged”, and it adds a new ultra level that runs three research agents in parallel. By our arithmetic, the medium change gives roughly 12 to 14 times as much benchmark score per dollar.
Business Pill · Usage-Based Pricing
An evergreen Business Engineer explainer on usage-based pricing: paying for work, not for seats. It teaches the general idea only and does not discuss Perplexity or this update.
The key insight: As we read it, Perplexity is reporting that its Agent API now delivers higher or unchanged benchmark scores at lower cost per query at every effort level, by its own measurements: the medium level rose from 55% to 77% on BrowseComp and DSQA while its cost fell almost 90%. That is roughly 12 to 14 times as much benchmark score per dollar, by our arithmetic. The new ultra level costs $21.71 per WANDR task.
Effort levels let developers pick a bundled configuration of model, step budget, search settings and tools for each task, the post says; requests that call a level by name pick up the new settings automatically, and pinning a version keeps an older one.
The post gives these figures. The xhigh level held its score on WANDR, Perplexity’s own long-research benchmark, at 76.5% while running at about 25% of the cost of its August 2026 configuration. With the openai/gpt-6-lunamodel, raising the step limit from 1 to 15 lifted accuracy on BrowseComp and DSQA from 41% to 80% while cost per query rose from $0.0017 to $0.0152, “about 9x the cost for about 2x the score”.
A new extraction step in fetch_url sends each page to a smaller model and keeps only the relevant passages; in one evaluation a 5,561-token page was cut to 293 tokens, the post says. At the high level, extraction scored 88.0% at $0.073 per query, against 83.5% at $0.094 without it. Fast Search costs $1.00 per 1,000 requests.
The ultra level runs three agents on the whole task in parallel, and a fourth merges their findings and writes the answer. On WANDR it scored 85.9% at $21.71 per task, the highest of the configurations Perplexity tested, according to the post.
Perplexity chief executive Aravind Srinivas posted at 23:54 UTC on 5 November: โEvery month is a new Pareto optimal curve for Perplexity.โ The post links to Perplexityโs blog on its Agent API effort levels, the source of this pieceโs figures.
Reading the Numbers
By our arithmetic, extraction raised the high score by 4.5 points while cutting cost per query by about 22%, and it cut the sample page by about 95% of its tokens. The xhigh figure works out to about a 75% cost reduction at an unchanged WANDR score.
The post also says not every test helped: loading every skill up front raised cost by 13% on BrowseComp and DSQA. It describes an automated loop that tests new models as they appear; “A candidate replaces the current model only when it wins on quality at the same or lower cost.” One candidate scored 13.3 points above the baseline on BrowseComp but below it on Perplexity’s internal enterprise, legal and finance tasks, the post says.
As we read the post, most of the savings came from limiting what the agent carries and how long it runs: step and tool-call budgets, compaction and page extraction, rather than from a single cheaper model.
As we read the step-budget figures, more steps still buy more score: with gpt-6-luna, Perplexity reports about 9x the cost for about 2x the score from 1 to 15 steps.
As we read the model-selection section, Perplexity says it swaps a model into a level only when it wins on quality at the same or lower cost, and that one candidate scored higher on BrowseComp but lower on its internal tasks.
Perplexity, in its 5 November blog post
“every effort level now costs less per query, with research quality that is higher or unchanged”
Three Implications
FOR DEVELOPERS ON THE AGENT API Requests that call a level by name get the new configuration automatically; pinning a version keeps an older one, the post says.
FOR COST PER QUERY At the high level, Perplexity reports 88.0% at $0.073 per query with extraction, against 83.5% at $0.094 without.
FOR READING THE FIGURES They are Perplexity’s own, on public benchmarks and on WANDR, its own benchmark, and the ultra cost is per task, not per query.
The framework says: “Inference is a revenue engine: continuous, distributed across millions of applications and agents, incurring marginal cost with every query.”
As we read it, Perplexity’s post is about that marginal cost: what each research query costs and what score it buys. The framework does not discuss Perplexity or its Agent API.
What Is Not Established
The scores are Perplexity’s own, on public benchmarks and on WANDR, its own benchmark; they have not been independently checked. The post gives the mediumcost cut as a percentage, not in dollars, and it does not give per-query prices for every level. The ultracost is per WANDR task and is not comparable with per-query costs on other benchmarks. The post does not say how many developers use the Agent API.
This Week In The AI Supercycle
Anthropic’s Reality Check, The Capital Squeeze, & The New World Order
This week’s issue: for two years the AI build-out was priced as a growthstory. This week it started being priced as a credit story.
Perplexity says its Agent API effort levels now cost less with equal or better scores: medium rose from 55% to 77% on BrowseComp and DSQA while its cost fell almost 90%, and a new ultra level scores 85.9% on WANDR at $21.71 per task. By our arithmetic, the medium change is roughly 12 to 14 times as much benchmark score per dollar.
A note on sourcing. The scores, costs and configuration details are from Perplexity’s blog post of 5 November 2026, shared on X by chief executive Aravind Srinivas; they are Perplexity’s own figures on public benchmarks and on WANDR, its own benchmark. The score-per-dollar, cost-reduction and token figures are our arithmetic.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
