共有:

PERPLEXITY / MODEL COUNCIL

Asking one question to several AIs at once.

Perplexity's "Model Council" has expanded to its cloud platform, Perplexity Computer. Users can now pick 2 to 8 models — from OpenAI, Anthropic, and Google's frontier labs to open-weight models like GLM and Kimi — to run in parallel, surfacing where they agree and where they diverge.

AI Navigate Editorial2026.07.305 min read

GPT Claude Gemini GLM Model Council Comparison
Multiple opinions, collapsed into one synthesized view.
01
BACKGROUND

The quiet history of "ask several AIs at once"

A feature that launched quietly, then expanded over five months.

Model Council, which Perplexity began rolling out in 2026, lets users fire one question at multiple AI models simultaneously and see the answers laid side by side. According to Perplexity's own blog post, the feature originally ran three different models per query; after each model generated an independent answer, Perplexity's built-in synthesizer analyzed the outputs and organized them into a structured comparison table. As Storyboard18 reported, the design rests on the premise that different models excel at different tasks — summarization, reasoning, creativity, math, coding, handling of sources, and so on. What made it distinctive wasn't just listing several answers, but surfacing exactly where they disagreed.

It's easy to confuse this with the "19-model parallel execution" feature that Perplexity Computer had already introduced on May 2, 2026. That feature can run up to 19 models at once, but it simply lays out raw outputs without a synthesized view highlighting agreement or disagreement. Model Council has always been a separate, comparison-and-synthesis-focused feature.

On July 28, 2026, The Register reported that Model Council, five months after its launch, had expanded to the cloud-based Perplexity Computer platform as well. According to the article, users can now freely choose between 2 and 8 models, with options spanning frontier-lab models from OpenAI, Anthropic, and Google as well as open-weight models like GLM and Kimi. This marks a clear pivot: from a single vendor handing down "the one correct answer," to a mechanism for cross-checking a group of models shaped by different philosophies and training data.

02
BY THE NUMBERS

The expansion, in numbers

Five months after launch, the feature quietly grew.

3
Models run simultaneously at launch
2-8
Models selectable after the Perplexity Computer expansion
5 months
Time from launch to the Perplexity Computer expansion
19
For reference: Perplexity Computer's existing parallel-run cap (no synthesized view)
03
HOW IT WORKS

Four steps to use it

01

Enter your question

Type the question you want compared, just like a normal search.

02

Pick 2 to 8 models

Options include frontier models from OpenAI, Anthropic, and Google, as well as open-weight models like GLM and Kimi.

03

Each model answers independently

Each selected model reasons and generates its answer on its own, without seeing the others' responses.

04

The synthesizer builds a unified view

The built-in synthesis engine analyzes all the answers and organizes agreements and disagreements into a structured comparison table. You can still open any individual model's full answer.

04
WHO BENEFITS

Who it helps, and how

The more you tend to take one AI's answer at face value, the more this gives you to work with.

Engineers: measure each model's strengths and blind spots

In coding and technical reasoning, different models tend to be strong in different languages and weak with different frameworks. Send the same implementation question to several models, and the comparison view instantly shows which one's suggestion is more likely to hide a bug or a stale API call. It works as a lightweight cross-review before you commit to any single model's advice.

Business: a second opinion before you decide

For questions without a single right answer — market research, competitive analysis — the way models disagree is itself information. A conclusion every model reaches is more trustworthy; a point where they split marks where more digging is needed. It gives internal proposals and approvals a new metric: cross-model consensus.

PMs: more sparring partners for specs and priorities

Questions about requirements or prioritization often get different framings depending on which AI you ask. Send the same product problem to 2 to 8 models, and the synthesizer's agreement/disagreement summary can surface a point your team overlooked — or a judgment call that was quietly shaped by one model's particular habits.

05
COMPARE

A single AI vs. Model Council

Before: asking one AIModel Council: comparing several
Only one line of reasoning; errors are easy to missMultiple lines of reasoning can be cross-checked against each other
Hard to notice where a model's strengths and weaknesses don't fit the taskDifferences in summarization, reasoning, coding, etc. show up directly as agreement or disagreement
Reading other models' answers means asking againThe unified view and each individual answer sit on the same screen
06
NEXT STEPS

What to do next

Expect more high-stakes questions to get routed to more than one model.

The Model Council approach is likely to spread to other AI products over time. Here are three things readers can do right now.

  • For questions tied to important decisions, don't stop at one AI's answer — cross-check it through a tool that can run several models side by side.
  • In a unified view, pay more attention to where models disagree than where they agree, and dig into why (differing training data, differing strengths, and so on).
  • Always check each model's cited sources, rather than trusting the synthesizer's summary alone.

Still, this isn't unqualified good news. As The Register's headline wryly put it, "tokenmaxxing" is exactly what happens when 2 to 8 models run in parallel on every query — compute cost and token consumption scale right along with it. And agreement across multiple models doesn't guarantee factual correctness: models with similar training data and similar biases can converge confidently on the same mistake. Model Council visualizes where opinions diverge; it does not guarantee which opinion is right.