Perplexity Computer explained: 19+ frontier models, unified with files, tools, memory, and multi-agent workflows in a single system. A practical guide to why it's novel and how to use it.
Thanks so much for reading and commenting, Mike! π€ I'd love to hear your thoughts once they make it available to Pro users. Also, have you been using Kimi much? I haven't had the chance to test it yet.
Honestly, I use all of the models as of 2026. Gemini 3.1 Pro, GPT-5.2, Sonnet 4.6, Grok 4.1 (soon 4.20?), and Kimi K2.5 are all elite.
I like Kimi most when fact-checking long PDFs or documents. It's excellent for document retrieval and citation handling.
Kimi is also one of my favorites, because in spirit, it's open-weights/open-source.
I really wish I had the "Model Council" feature. It looks like Kimi would shine brightest as part of a team with lots of stuff to cross-reference. It would be very good at analyzing and fact-checking alongside other AIs.
I also kind of make my own "council" manually when fact-checking stuff, which Kimi also does well, lol.
Two things. The main appeal of openclaw was the possibility of sending messages in whatsapp, telegram or whatever and having the agent respond there as well.
I donβt think perplexity computer does this, right? At least from what I read in this article.
Secondly, itβs not a genuinely novel idea to use different models from different providers. This has been possible for some time now. Iβve been doing this for the better part of last year.
The hard part would be letting the system decide which model to use based on the task. And that is not solved properly yet and itβs quite tricky. So if we stick to creating agents and manually assigning which model is used by that agent plus a falllback is not a new idea.
Vladi, thank you for reading! That's the difference between Openclaw and Computer - you don't need to assign models yourself, it's done for you by the orchestrator layer. π€
Iβve seen other system do that as well and were very good at it. Either poor performance or costs were too high. I guess we will see how this one performs.
Yes. So Iβve been testing more since that article, and considering how many Claude fans are on Substack (myself included), this might be an unpopular opinion π but the performance is impressive. The same tasks are done faster, because Computer spawns not only agents and subagents, but also different models simultaneously. I havenβt seen anything quite like this yet.
The biggest barrier to widespread adoption is the price tbh. And the fact that so many people moved from ChatGPT to Claude in the past 6 months - and it's easy to be loyal to Claude, because it *is* a solid product.
I think perplexity hit the nail on the head with this model as more and more people will want a defacto interface without root controls inside their laptops.
I will always advocate that any user should have admin privileges and not the machine. You made me curious and I'll thinker with it this weekend. π
Hahaha, please do Markπ€ I was playing with it again last night, so now I can say more about how fast it feels. Since it uses all these agents simultaneously, itβs actually faster than β dare I say it? β Claude on its own. Iβm reaaaally curious to learn what you think!
Interesting read. I currently use Perplexity for research and to run scheduled tasks (I tried the later in ChatGPT and found it unreliable). I have Perplexity pro and definitely not taking full advantage of it. Will definitely keep an eye on this π
I love most of that branded callout box generator, it is an amazing idea, I was literally wondering if you are already using it, haha.
I'm most curious about the research engine piece though. You mentioned it runs seven search types in parallel and reads full source pages, not snippets. Would love to see it's actual output :)
And the watermark complaint, yes, I feel you! The same complaint I have about coding is, everytime Cursor or Claude push a git commit, they'll add a signature saying "co-written with Cursor / Claude .....". These guys care so much about their own identity!!!
Thank you, Jenny! Iβve been using Figma for the tables and callout boxes, so this will definitely be a time saver. When I have more time, Iβll probably turn it into a customizable app and open-source it on Substack π€
Bummer that they don't have it on the Pro plans, I'd love to try it out especially that I've not warmed up to OpenClaw yet. Saving this until it comes. Thanks for the brilliant breakdown, Karo!
Karo, you probably wrote about this before, but please point me in the right direction. I've done a little vibe coding that I usually run from my website. But I would like to think more about converting them into stand alone apps that can be monetized.
I am intrigued! Must go check this out.
Let me know what you think Karen! And thank you for reading π€
Hi Karo!
Thank you so much for this helpful and timely tutorial.
I'm also a huge fan of Perplexity Pro. I think it's the best value.
(I love how it lets me use Claude, Grok, ChatGPT, Gemini, and Kimi.)
In any case, I saw Perplexity Computer in my Perplexity Pro account lastnight. (Or was it the night before.)
I had no clue what it was! But, they let Pro users preview it for 24 hours or so.
But now, they have "clawed" it back behind a paywall, lol.
I appreciate your superb breakdown nevertheless! (I hope they let us back in soon. It looks very cool!)
Cordially,
Mike D
Thanks so much for reading and commenting, Mike! π€ I'd love to hear your thoughts once they make it available to Pro users. Also, have you been using Kimi much? I haven't had the chance to test it yet.
Hi Karo!
:)
Honestly, I use all of the models as of 2026. Gemini 3.1 Pro, GPT-5.2, Sonnet 4.6, Grok 4.1 (soon 4.20?), and Kimi K2.5 are all elite.
I like Kimi most when fact-checking long PDFs or documents. It's excellent for document retrieval and citation handling.
Kimi is also one of my favorites, because in spirit, it's open-weights/open-source.
I really wish I had the "Model Council" feature. It looks like Kimi would shine brightest as part of a team with lots of stuff to cross-reference. It would be very good at analyzing and fact-checking alongside other AIs.
I also kind of make my own "council" manually when fact-checking stuff, which Kimi also does well, lol.
Hope you have a great day, Karo.
Cordially,
Mike D
Sounds like a gamechanger
Two things. The main appeal of openclaw was the possibility of sending messages in whatsapp, telegram or whatever and having the agent respond there as well.
I donβt think perplexity computer does this, right? At least from what I read in this article.
Secondly, itβs not a genuinely novel idea to use different models from different providers. This has been possible for some time now. Iβve been doing this for the better part of last year.
The hard part would be letting the system decide which model to use based on the task. And that is not solved properly yet and itβs quite tricky. So if we stick to creating agents and manually assigning which model is used by that agent plus a falllback is not a new idea.
Vladi, thank you for reading! That's the difference between Openclaw and Computer - you don't need to assign models yourself, it's done for you by the orchestrator layer. π€
Iβve seen other system do that as well and were very good at it. Either poor performance or costs were too high. I guess we will see how this one performs.
Yes. So Iβve been testing more since that article, and considering how many Claude fans are on Substack (myself included), this might be an unpopular opinion π but the performance is impressive. The same tasks are done faster, because Computer spawns not only agents and subagents, but also different models simultaneously. I havenβt seen anything quite like this yet.
The biggest barrier to widespread adoption is the price tbh. And the fact that so many people moved from ChatGPT to Claude in the past 6 months - and it's easy to be loyal to Claude, because it *is* a solid product.
I think perplexity hit the nail on the head with this model as more and more people will want a defacto interface without root controls inside their laptops.
I will always advocate that any user should have admin privileges and not the machine. You made me curious and I'll thinker with it this weekend. π
Hahaha, please do Markπ€ I was playing with it again last night, so now I can say more about how fast it feels. Since it uses all these agents simultaneously, itβs actually faster than β dare I say it? β Claude on its own. Iβm reaaaally curious to learn what you think!
I'll drop a DM when I finish.
Fantastic!
Interesting read. I currently use Perplexity for research and to run scheduled tasks (I tried the later in ChatGPT and found it unreliable). I have Perplexity pro and definitely not taking full advantage of it. Will definitely keep an eye on this π
Great insights as always. Thanks for sharing
My pleasure, thank you for reading!
Youβve sold me on it. Iβve never dug deep into what Perplexity has to offer but might have to try this out.
Let me know what you think!
Karo, thanks for such a timely review!
I love most of that branded callout box generator, it is an amazing idea, I was literally wondering if you are already using it, haha.
I'm most curious about the research engine piece though. You mentioned it runs seven search types in parallel and reads full source pages, not snippets. Would love to see it's actual output :)
And the watermark complaint, yes, I feel you! The same complaint I have about coding is, everytime Cursor or Claude push a git commit, they'll add a signature saying "co-written with Cursor / Claude .....". These guys care so much about their own identity!!!
Thank you, Jenny! Iβve been using Figma for the tables and callout boxes, so this will definitely be a time saver. When I have more time, Iβll probably turn it into a customizable app and open-source it on Substack π€
Hi Karo- Loved this post. Thank you for sharing the comparison. Haven't noticed Computer in my Pro account yet, but definitely checking it out.
Thank you for reading Anuja! π€ It's not available for Pro users yet, unfortunately. But hopefully very soon!
Bummer that they don't have it on the Pro plans, I'd love to try it out especially that I've not warmed up to OpenClaw yet. Saving this until it comes. Thanks for the brilliant breakdown, Karo!
Thank you! I'm really curious to see what you think. Hopefully they'll open it up very soon! π€
Karo, you probably wrote about this before, but please point me in the right direction. I've done a little vibe coding that I usually run from my website. But I would like to think more about converting them into stand alone apps that can be monetized.
For example:
https://raftech.org/memorymatch
(To help memory of our senior patients and family)
https://RAFTECH.ORG/sys701prep
(Practice Exam Study for Security+)
I couldn't remember how to contact you in private, so hope this is okay. Thank you.
Richard
That's absolutely ok - I'll send you a DM as soon as I'm home. Thank you for reaching out Richard π€
Speedy insights! Thank you, Karo! βΊοΈ
My pleasure! Thank you so much for reading!
Thanks for this! Loving the multi modal orchestration.
Yes, it's fascinating. Thank you so much for reading Hodman π€
This could be a game-changer for people who want a single system to own projects instead of juggling separate AI tools.
Exactly!
Itβll be interesting to see how it stacks up against OpenClaw and Claude in practical workflows.
Thatβs true, thatβd be a great test to run! Thank you for reading Dennis π€