What Just Happened in AI
On February 25, 2026, Perplexity launched what it calls the most ambitious product in its three-year history: Perplexity Computer. The pitch is unlike anything else in the market — instead of a single AI model trying to do everything, Computer orchestrates 19 different AI models in parallel, dispatching each subtask to the model best suited for it. It can run autonomous workflows for hours, days, or even months, with minimal human intervention.
This is a different bet than ChatGPT, Claude, or Gemini. Those products are built around a single model (or model family) trying to be good at everything. Perplexity is saying: stop trying to make one model do it all. Use the right model for each subtask. Let an orchestrator decide which one.
If that bet pays off, it changes how we think about AI agents entirely. If it does not, it is an expensive experiment for the people paying 200 dollars a month to use it. Either way, it is one of the most interesting products in AI right now, and worth understanding in detail.
What Perplexity Computer Actually Is
Perplexity Computer is a cloud-based AI agent that runs in a secure isolated sandbox. You give it a project — research the top 50 SaaS competitors in your space, build a tax return, analyze a codebase, write a market report — and it executes the entire workflow autonomously. It breaks the project into subtasks, decides which AI model handles each one, runs them in parallel, spawns sub-agents when it gets stuck, and returns a finished deliverable.
Under the hood, Perplexity uses Claude Opus 4.6 as the core reasoning engine — the orchestrator that plans the work and routes the subtasks. The other 18 models handle specialized work:
- Gemini — deep research and creating sub-agents for parallel investigation
- Nano Banana — image generation
- Veo 3.1 — video production
- Grok — speed-optimized lightweight tasks
- GPT-5.2 — long-context recall and wide-ranging search
- Plus 13 more — covering coding, summarization, embeddings, structured data extraction, and other specialized capabilities
The orchestration is what makes it different. You do not pick the model. The orchestrator does — automatically, based on what each subtask actually needs. Perplexity calls this a "model-agnostic harness," and it means the model lineup can change as new models are released without you having to rewrite anything.
The Big Idea: Orchestration Over Monolith
For the past three years, the assumption in the AI industry has been that the path forward is bigger, smarter, more general models. GPT-5. Claude Opus 4.6. Gemini 2.x. The bet is that if you make one model good enough, it can do everything.
Perplexity is challenging that assumption. Their argument goes like this:
- Different models are genuinely better at different things. Claude is the best at reasoning and code. GPT is best at creative writing and structured tasks. Gemini is best at research with grounded sources. Grok is fastest. Each one has real strengths and real weaknesses.
- Until now, you had to pick one and live with its weaknesses for tasks it was not good at.
- What if instead of picking one, you used all of them — and let an orchestrator dispatch each task to whichever model handles it best?
That is the entire premise of Computer. It is not trying to build a smarter AI. It is trying to build a smarter system that uses multiple AIs intelligently.
Whether this produces meaningfully better results than just using a single top-tier model is the open question. Early reviews suggest the answer is "yes, for complex multi-step workflows." For simple Q&A or single-task work, you do not need orchestration — you just need a fast model.
How It Works in Practice
Here is what a real session looks like. Suppose you tell Computer:
"Research the top 25 competitors of [our product] in the US market. Build a comparison spreadsheet with pricing, target customers, key features, and their main differentiators. Add a column for which ones have raised funding in the last 12 months."
What happens next:
- The orchestrator (Opus 4.6) plans the work. It identifies subtasks: find competitors, gather data on each one, structure the data into a spreadsheet, search for funding news.
- It spawns parallel sub-agents. A Gemini sub-agent does the deep research on each competitor. A GPT-5.2 sub-agent handles the long-context synthesis. A coding sub-agent (Claude or another model) builds the actual spreadsheet output.
- It uses connected accounts. If you have Computer connected to your Gmail, GitHub, Slack, Notion, or other tools, it can pull supplemental information from those sources without asking.
- It accesses the live web. Web search and browsing are built in. No separate plugins.
- It writes code if needed. If a task requires data manipulation or generating a file, Computer can execute code in its sandbox.
- It only checks in when truly stuck. Unlike most AI tools that ask 10 clarifying questions before starting, Computer just goes. It asks for human input only when it cannot resolve something on its own.
- You get a finished deliverable. Not a draft. Not a summary. The actual spreadsheet, with the columns you asked for, populated with researched data.
The whole thing can take minutes for a simple task or hours for a complex one. Long-running workflows are explicitly supported — Perplexity says some tasks can run for days or even months.
Key Features
Persistent Memory
Computer retains context across sessions. It remembers your past work, your preferences, your files, and the findings from previous tasks. You do not have to re-explain context every time you start a new conversation. For Enterprise users, this extends to organization-wide memory shared across team members.
Hundreds of Connectors
Computer integrates with the tools you already use: Gmail, Outlook, GitHub, Linear, Slack, Notion, Snowflake, Databricks, Salesforce, SharePoint, HubSpot, Datadog, and many more. Enterprise admins can install custom connectors via the Model Context Protocol (MCP). The connectors give Computer the ability to read your real data without you having to manually copy-paste or upload files.
Web Access Built In
This is where Perplexity's search-engine roots show. Computer can browse the web, read articles, scrape data from public pages, and pull live information from the internet. No separate browsing plugin required.
Long-Running Tasks
Most AI tools time out after a few minutes. Computer is designed to run for hours, days, or months. The sandbox keeps state. The orchestrator keeps working. You can come back the next day to a finished result.
Sub-Agent Spawning
When Computer encounters an obstacle, it does not just give up or hallucinate a solution. It spawns a sub-agent specifically to solve that problem — find an API key, write a custom script, do additional research, whatever is needed. This is one of the things that makes it actually agentic instead of just "ChatGPT with more steps."
Sandbox Isolation
Every task runs in an isolated cloud sandbox. Your data does not leak between tasks. Perplexity also commits explicitly that task inputs, outputs, connector data, and sandbox contents are not used to train their models — a promise that matters for enterprise and privacy-conscious users.
Pricing and Access
Computer is not cheap. It is exclusive to Perplexity Max, which costs 200 USD per month (or 2,000 USD per year, saving roughly 17 percent). It is not available as a standalone purchase. To use Computer, you have to subscribe to Max.
What you get with Max:
- Access to Computer (the agent)
- 10,000 monthly credits
- Unlimited Pro searches
- Access to the underlying models (Opus 4.6, GPT-5.2, etc.)
- Sora 2 Pro video generation
- The Comet AI browser
- Unlimited Labs usage
The Credit System
This is where it gets tricky. Every task consumes credits based on complexity. Simple tasks cost very little. Complex tasks can burn through thousands.
Documented examples from users:
- Generating alt-text for an image: ~30 credits
- An extended coding session: thousands of credits
- Scanning a 280,000-line Python codebase: ~21,000 credits (over twice the monthly allowance)
- Building a single website page: 200 USD in additional credits over two days
- One 40-minute task: 15,000 credits (50 percent over monthly allowance)
The default spending cap is 200 USD per month — meaning the maximum you can pay (subscription plus one round of additional credits) is 400 USD before tasks are paused. Heavy users can raise this cap to 2,000 USD per month.
If you do a few research tasks per week, the included 10,000 credits will likely cover you and your total stays at 200 USD. If you run Computer as continuous infrastructure — daily complex projects, long autonomous workflows, code analysis — budget 300 to 500 USD per month realistically.
Enterprise Pricing
Enterprise Max access is 325 USD per seat per month (3,250 USD per year). It includes everything in Max plus enterprise-grade security: SOC 2 Type II certification, SSO/SAML authentication, SCIM provisioning, granular admin controls, and full audit logging.
The Product Family
Since the original launch, Perplexity has expanded Computer into a family of products:
Personal Computer (March 11, 2026)
Announced at Perplexity's inaugural Ask 2026 developer conference, Personal Computer is a version of Computer that runs on a dedicated Mac mini in your office or home. It gives the cloud-based AI agent persistent local access to your files, apps, and sessions — a 24/7 digital proxy that works on your behalf.
The clever part: the Mac mini runs continuously, providing local file/app access, while the AI processing still happens on Perplexity's cloud servers. Sensitive actions still require user approval, all actions are logged, and there is a kill switch. Initial access is via waitlist.
Computer for Enterprise (March 11, 2026)
Also announced at Ask 2026. Computer for Enterprise is the multi-model agent embedded directly into the tools enterprises already use. Employees can query @computer directly inside Slack channels and threads. Enterprise customers get business-grade connectors for Snowflake, Datadog, Salesforce, SharePoint, and HubSpot.
The most striking claim from Perplexity: in internal testing of more than 16,000 queries, Computer for Enterprise completed an estimated 3.25 years of human work in four weeks, saving roughly 1.6 million USD in labor costs. Take that with a grain of salt — it is a vendor-reported number, not independently verified — but the direction is clear: this is being positioned as a serious replacement for entry-level knowledge work.
Computer for Taxes (April 2026)
Perplexity extended Computer into U.S. federal tax preparation for the 2026 filing season. Computer for Taxes goes further than answering questions — it drafts the actual federal return on official IRS forms. It can also review professionally prepared returns for accuracy and compliance. Perplexity claims that during testing, on one attorney-prepared return, deductions under the 2025 No Tax on Overtime provisions were understated by 67 percent, leaving thousands of dollars unclaimed. Access is via "Navigate my taxes" inside Computer, and a Pro subscription (17 USD per month) is required.
Real Use Cases
Based on user reports and Perplexity's own documentation, here is what people are actually doing with Computer:
- Competitive research — "Research my top 50 competitors and build a comparison spreadsheet." This is the killer demo, and it works as advertised.
- Market research over time — Long-running workflows that monitor a market or topic for weeks and report back changes.
- Code review and code generation — Scanning entire codebases for issues, refactoring, generating documentation. Heavy on credits but powerful.
- Content production pipelines — Research, draft, fact-check, format, publish — the entire pipeline as one task.
- Tax preparation — As of the April 2026 update, drafting and reviewing federal tax returns.
- Data analysis — Plain-English queries against Snowflake or Databricks (Enterprise) without writing SQL.
- Document drafting and synthesis — Pulling from multiple sources, structuring the output, delivering a finished document.
What it is not good for: simple Q&A (use regular Perplexity Pro, much cheaper), conversational chat (use Claude or ChatGPT), real-time interactive coding (use Claude Code or Cursor), anything privacy-sensitive that you cannot trust to a cloud service.
How It Compares
vs. ChatGPT and Claude
ChatGPT and Claude are single-model products. You get one model (or one model family) and you use it for everything. They are excellent at conversation, coding, writing, and reasoning — but they each have weaknesses, and you cannot easily mix their strengths.
Computer is multi-model orchestration. You get whichever model is best for each subtask, automatically. The trade-off: you give up direct control and pay 10x the price.
Verdict: if you do simple tasks and like talking to one AI, ChatGPT or Claude at 20 USD/month is fine. If you do complex multi-step workflows that benefit from different model strengths, Computer is justifiable.
vs. OpenClaw
OpenClaw is the self-hosted, open-source alternative — a gateway you run on your own server that connects messaging channels to AI models you choose. It is free (other than your model API costs), private, and customizable. We covered the setup in detail in our OpenClaw guide.
Computer is the managed cloud service version. You give up control over hosting, models, and data location, but you get a finished product with no setup required.
Verdict: if you want full control, privacy, and customization, OpenClaw. If you want it to just work and you can afford the 200 USD/month, Computer.
vs. Microsoft Copilot and Salesforce Einstein
Microsoft and Salesforce are Computer's direct enterprise competitors. They have deeper integrations with their respective ecosystems but are locked into a single AI provider. Computer's pitch to enterprises is that 19 models beats one, regardless of how deeply integrated the one is.
Whether enterprises buy that argument is the open question of 2026.
vs. DIY Frameworks (LangChain, AutoGen, CrewAI)
Open-source agent frameworks let you build your own multi-model orchestration. They are free, infinitely flexible, and a giant time sink. You spend weeks wiring up what Computer gives you out of the box.
Verdict: if you have a developer team and a unique workflow, build your own. If you have a credit card and a deadline, pay for Computer.
Honest Critique
Computer is a real product with real strengths, but it is not perfect. The honest concerns:
1. It Is Expensive
200 USD per month is a lot. For most casual users, the regular Perplexity Pro at 20 USD/month does 90 percent of what they need. Computer is priced for professionals who can directly attribute revenue or saved time to its output — if you cannot, you are paying for capability you are not using.
2. The Credit System Is Unpredictable
Until you have used Computer for a few weeks, you have no way to budget. Perplexity has not published a per-task credit table, so you find out how expensive a task is by running it. Heavy users routinely blow through their monthly allowance and pay extra.
3. Cloud-Only
Everything runs on Perplexity's servers. Even Personal Computer (which adds local file access via a Mac mini) still does the AI processing in the cloud. If your data cannot leave your premises, this is a non-starter.
4. Vendor Lock-In
You depend on Perplexity for the orchestration logic, the connectors, the model lineup, and the entire abstraction. If they raise prices, change terms, or get acquired, you are stuck. With self-hosted alternatives like OpenClaw, you can switch providers at any time.
5. 19 Dependencies
Every model in the orchestration is a third-party dependency. If OpenAI, Anthropic, Google, or any other provider has an outage, parts of Computer degrade. The orchestration is impressive but it is also a more complex system with more failure modes.
6. The "Unverified Productivity Claims"
Perplexity's "3.25 years of work in 4 weeks" claim from Computer for Enterprise is great marketing but not independently audited. Take any vendor-reported productivity numbers with skepticism until you see real-world case studies from companies you know.
Should You Pay for It?
Here is the practical decision framework:
Yes, Subscribe to Computer If:
- You do high-volume research that takes hours of human time
- You need multi-step autonomous workflows
- You are replacing entry-level knowledge work (research assistant, junior analyst, content researcher)
- Your time is worth more than 200 USD/month and Computer can save you a few hours per month
- You work in tax prep, market research, competitive intelligence, or business strategy
No, Stick With Cheaper Tools If:
- You just want fast Q&A and search (Perplexity Pro at 20 USD covers it)
- You do mostly conversational AI work (ChatGPT or Claude at 20 USD covers it)
- You need self-hosted/private (use OpenClaw or build your own)
- You are price-sensitive and not generating revenue from AI work yet
- You need real-time interactive coding (use Claude Code or Cursor)
The math is simple: if Computer saves you two hours of human research time per month at 100 USD/hour, it pays for itself. If you cannot make that math work, do not subscribe.
The Bottom Line
Perplexity Computer is a genuine bet on a specific vision: orchestration is better than monolithic models. Whether that vision turns out to be correct is the most interesting question in AI right now.
If multi-model dispatching produces meaningfully better outcomes for complex workflows than just using a single top-tier model, Perplexity has built the future. If a sufficiently good monolithic model can match it, this whole approach is a detour.
What we know for sure: it works. Real users are running real workflows on it. Some are getting impressive results. It is not vaporware.
What we do not know yet: whether the orchestration approach scales beyond demos and produces durable advantages over just using GPT-5 or Claude 4.6 alone. The next 6 to 12 months will tell us.
For now: if you are doing serious agentic work and you can afford 200 USD per month, Computer is the most interesting tool to try. For everyone else, watch this space and use the cheaper alternatives until the value proposition becomes obvious.