Perplexity's Model Council runs one prompt through up to eight AIs at once

Perplexity's Model Council now lets Pro subscribers run one prompt through up to eight AI models in parallel, see the answers side by side, and let a separate "judge" model stitch them into a single response. The feature, which launched on Perplexity's web app in February 2026 for the company's top-tier Max subscribers, expanded to its Computer platform on Tuesday and dropped the price of admission to $20 per month.

That is the surface change. The deeper one is what the feature trades: instead of betting on one model's answer, a subscriber bets on the arbitration between several.

The mechanism, as Perplexity describes it, works in two passes. When a user submits a question, the system fans it out to between two and eight models chosen by the user from a mix of frontier providers and open-weight options. Each model returns its own answer independently. A separate review model then reads the parallel outputs and produces a synthesis report that flags where the models agree, where they split, and what each one contributed that the others missed. The user sees both the raw answers and the synthesis.

This is a deliberate departure from how most AI assistants still work. Single-model products like ChatGPT, Claude, or Gemini pick one engine and return one answer, leaving the user to trust the pick or switch tools. Model Council is closer to a panel: the reader, not the assistant, decides which voices get a seat.

The pitch, in Perplexity's own words, is a "more accurate, higher-confidence answer" drawn from running multiple frontier models at once. The Register, covering the Computer launch this week, framed the same move as "tokenmaxxing": a deliberate choice to spend more compute per question in exchange for more signal.

Both readings point at the same trade-off, and the cost is real. Every additional model in the panel is another paid API call's worth of tokens, and Perplexity is metering Model Council usage on top of the $20 Pro subscription. A two-model run is cheap. A six- or eight-model run on a long, citation-heavy query can climb fast, and Lumienai's read of the announcement is that complex tasks will run up the bill.

That is the part the feature name does not carry. Council is not a smarter single model; it is a metered aggregation. A user who fires an eight-model panel at every prompt is paying for eight reasoning passes, and the synthesis step is a ninth.

There is also a quality cost that is harder to meter. Averaging several answers can produce a vague, consensus-shaped reply that is worse than one sharp one. A panel that splits three-to-two on a contested point surfaces the disagreement, which is the point of the product. A panel that splits eight-to-zero on a straightforward factual question has spent eight times the compute to confirm what any single model could have answered. The product puts that adjudication in the user's hands, but it does not do the adjudication for them.

The original web version, launched in February, was fixed: Max subscribers got three models with no choice of which three. The Computer version is the first time users can pick the panel themselves, and the first time anyone paying $20 a month, not just Max subscribers, can use the feature at all. Both moves point the product at a different reader: someone who is already willing to read multiple AI answers and would rather pick the panel than inherit it.

Perplexity has not published benchmarks on whether Model Council answers are measurably more accurate than single-model answers on the same prompt. The company's marketing frames the feature in confidence terms, and the synthesis report is the closest thing to an accuracy claim built into the product. Independent benchmarking would need to pin down which model acts as the judge, which the company has not disclosed, and which prompts reward aggregation in the first place.

The product's promise is that a reader who is not sure which AI to trust for a given question can see several answers side by side instead of picking a model and hoping. The product's cost is that the same reader is now paying for every seat at the table, including the one that decides whose seat mattered most.