Perplexity strengthens its work agents and lowers their costs

Perplexity has introduced Teammates, AI agents designed for continuous work, and has revamped its entire Agent API. The challenge is twofold: to do better with lower costs, while targeting broader enterprise use cases.

Perplexity has unveiled Teammates, AI agents designed to work continuously in a collaborative space, and announced on November 5, 2026, a redesign of its Agent API with less costly effort levels per request. The service aims to compete with approaches from OpenAI and Anthropic, with agents capable of chaining tasks, tracking projects, and maintaining working memory. For businesses, the goal is clear: automate more without letting the bill rise too quickly.

Teammates, des collègues virtuels qui ne demandent pas de café

Teammates relies on Perplexity Computer, a system that deploys agents on the user's device to perform tasks. Perplexity specifies that Teammates goes further by supporting "ongoing, large-scale" work, meaning continuous and large-scale, with multiple roles assigned within the same space. Each agent has its own thread in the application and can also be followed in Slack.

Aravind Srinivas, Perplexity's CEO, sums up the ambition on X: "Over time, we envision every org having hundreds of teammates working for each person." The phrase clearly indicates where the product wants to go: towards hybrid teams, where humans lead while the agent executes, searches, and gathers evidence. One detail remains, however, that is less appealing to financial departments: Teammates requires the Max subscription at 200 dollars per month.

Un Agent API plus flexible, mais surtout moins cher

The Agent API update is based on named effort levels that group model choice, search limits, tool calls, and search behavior. In practice, a developer can request a "fast" mode without adjusting each parameter separately. Perplexity says it has lowered costs at all levels while maintaining or improving quality on several benchmarks.

The medium level illustrates this shift well. On BrowseComp and DSQA, its score reportedly rose from 55% to 77%, while its cost dropped by nearly 90%. At the other end of the spectrum, the xhigh level maintains a WANDR score of 76.5% with a configuration estimated at about 25% of its August 2026 cost. WANDR, as a reminder, is an evaluation set focused on information gathering tasks where the agent must find verifiable sources.

Pourquoi Perplexity change sa recette

The logic is simple: in AI agents, cost depends as much on the model as on the number of steps, tools used, and context retained. Perplexity has therefore reduced budgets, introduced remaining capacity notifications, and worked on context compaction, that is, condensing history into summaries to keep the essentials without overloading the agent's memory.

The group also introduced extraction, which consists of transmitting to the agent only relevant passages from a page, quoted verbatim. According to Perplexity, this approach reduces input tokens per request by about 50% and improves value for money in some cases. The same logic applies to search: queries can return up to 4,000 tokens of results, compared to 2,000 previously, and Fast Search is now priced at 1 dollar per 1,000 queries.

Une stratégie pensée pour les usages d’entreprise

Ultra, the new most ambitious level, targets broad search and data collection when completeness matters more than latency. Perplexity indicates it mobilizes three agents in parallel, then a fourth to consolidate results, remove duplicates, and draft the final response. The service reportedly reached 85.9% on WANDR, with a cost of 21.71 dollars per task, the best result among tested configurations.

This architecture shows that Perplexity no longer sells just a response engine, but a complete workflow, from clue spotting to synthesis. Developers also have dynamic, pinned, or frozen versions of configurations, with a clear difference between automatic updates and reproducibility. It's serious, precise, and a bit less glamorous than a robot that "thinks" on its own — but much more useful for an IT service that needs to keep control.

By the numbers

  • 200 dollars per month — Max subscription required for Teammates.
  • 77% — medium score on BrowseComp and DSQA after the update.
  • 4,000 tokens — maximum volume of results per search.
  • 21.71 dollars per task — announced cost for Ultra on WANDR.
  • November 5, 2026 — date of Perplexity's announcement.

To read