Perplexity’s Multiagent lets Fable 5.2 run Gemini, GPT and Opus in parallel
Aravind Srinivas says Multiagent runs inside Computer and Teammates, switched on through the Effort level setting, with Fable 5.2 directing rival models in parallel.

Perplexity CEO Aravind Srinivas introduced Multiagent on X on November 17. It works in Computer and Teammates and lets Fable 5.2 direct Gemini 4 Carbon, GPT 6.1 Sol and Opus 5.5 as parallel agent streams.
One orchestrator model sending work to three rivals
Srinivas introduced the feature with the words “Introducing Multiagent on Perplexity.” According to his post, it can be used with both Computer and Teammates, and it is switched on through the Effort level setting. His example is a setup in which Fable 5.2 acts as the orchestrator and coordinates Gemini 4 Carbon, GPT 6.1 Sol and Opus 5.5 at the same time, in separate parallel streams, on complex tasks.

The text of the post available to us is cut off after the words about complex tasks, and the link to the original returned an error when we tried to read it. The remainder of the sentence, any attached video details and any wider documentation are therefore not confirmed here. Perplexity’s help pages or blog may carry more detail. At the time of writing, no separate announcement from the company was among the sources we could check.
Fable 5.2 reached Computer and Teammates only last week. As we reported, Perplexity added Fable 5.2 to Computer and Teammates on November 9, with Srinivas claiming it beats Gemini Argon on the company’s own WANDR research benchmark. Multiagent now gives that model a supervisory role over models from three other developers.

Which plans can use it is still unclear
The post does not say which subscription tiers get Multiagent, what it costs in usage, or whether the parallel streams draw more credits than a single-model run. That matters because the surrounding products are gated. Teammates arrived in the Mac app on November 14 for Max subscribers only, and our earlier coverage cited the Max plan at $200 a month. Whether Multiagent is limited to that tier, or to the web, has not been stated in the material we could read.
Likewise, the post does not say how the Effort level toggle behaves, such as whether a higher setting always triggers multiple models or leaves the choice to the orchestrator. Nor does it explain how results from different models are merged, or how disagreements between them are handled.

What the move says about the multi-model race
The angle here is competition. Perplexity has pitched Computer as a system that spreads work across many models since its launch on February 25, 2026. Third-party coverage at the time described it as running work across 19 models in parallel, with users able to assign particular models to particular subtasks. Multiagent looks like a more explicit version of that idea, where the user opts in and the lead model is named.
The model lineup is also notable. Gemini 4 Carbon opened to all developers on November 10 at introductory rates of $2.00 per million input tokens and $10.00 per million output tokens, according to the Gemini API changelog, and those rates end on November 30. Our report on Gemini 4 Carbon pricing notes the changelog does not say what comes afterward. Because Perplexity pays model providers per use, a pricing change on any of the four models could change what a Multiagent run costs the company, though Perplexity has not said so.
For rivals, the pattern is different. OpenAI, Google and Anthropic each sell their own agents built mainly on their own models. Perplexity, which trains none of the three models named in the post, is positioning itself as the neutral layer that picks and combines them. The post gives no evidence yet that the combination beats any one of those models alone.
Claims that still need independent testing
Srinivas offered no benchmark scores, latency figures or cost comparisons in the post we could read. Earlier claims around Teammates, such as the rise in usage he described on November 9, were also the company’s own and were not independently verified. Parallel agents often finish sooner but consume more compute, so speed and price are the two things to look for once users start reporting results.
In the past days Srinivas has promoted Teammates running on Fable 5.2 for long jobs, including a browser game built over several hours. Those are demonstrations chosen by the company. A fair test of Multiagent would be a published, repeatable task run with and without the toggle, with the cost and time recorded.
Sign Up for the Daily AI Pulse
One email a day. All the stories that matter.
By signing up, you agree to our Terms of Use and acknowledge the data practices in our Privacy Policy. You may unsubscribe at any time.
Alex CarterAI News & Big Tech CorrespondentAlex Carter writes for Aiholics, keeping readers updated on the fast-paced world of AI and Big Tech.



