The Proof · AI tool
- Who it's for
- Founders, analysts, and heavy researchers who will use Deep Research, Model Council, the Computer agent, and the premium connectors most weeks. For everyone else, the $20 Pro tier runs the same core search.
- Real cost
- Free tier (gated after a few anonymous searches). Pro is $20 per month. Max is $200 per month. Both run the same core search; Max adds the frontier models, Model Council, far higher Deep Research and Computer limits, and the premium data connectors.
We bought Max, signed in, and drove every tool: nine models, Model Council, Deep Research, the Computer agent, the premium data connectors, Finance and Academic. The advanced tools are genuinely powerful. The everyday search still leans on content mills, and most people do not need the $200 tier.
What's good
- Model Council (Max-only) convenes GPT-5.5, Claude Opus 4.8, and Gemini 3.1 Pro, then lays out where they agree, where they disagree, and what each found alone. Genuinely useful for a high-stakes call.
- Deep Research ran 8 steps across multiple searches and produced a structured, multi-section cost report that reads like a paid consultant brief.
- The Computer agent planned a task, pulled 47 sources, and built an actual Excel spreadsheet of mini-PC specs in 1 minute 21 seconds.
- Academic mode swaps the content mills for real sources (arxiv, SSRN, Reuters) and returned a rigorous, correctly-caveated answer.
- Premium data connectors (PitchBook, CB Insights, Statista, Wiley, Midpage legal) come included from the Pro tier up, alongside your own Google Drive, Gmail, and Dropbox.
- Max unlocks the frontier models; Claude Opus 4.8 gave a sharp cost analysis and flagged that the comparison was not quality-equal.
Where it breaks
- Everyday search and Deep Research still lean heavily on SEO content mills; one Deep Research report cited a politics magazine for a claim about Anthropic billing.
- At $200 a month it is ten times Pro, and most people will never touch the tools that justify the gap.
- The breadth (Spaces, Skills, Workflows, Memory, Connectors, Computer) is a lot to learn before it pays you back.
- You still have to click through and verify. Cited is not the same as sourced.
How we tested #
Our first pass was the free, logged-out version: four queries, a false-premise trap, a stats trap, and a hard signup wall after three searches. It refused to fabricate and got the facts right, but it cited content mills and gated us fast.
So we bought Max, the $200-a-month top tier, signed in, and drove every tool in it: the full model picker, the four search modes, Model Council, Deep Research, the Computer agent, Spaces, Connectors, Workflows, Memory, and the Finance and Academic surfaces. Everything below is what those tools actually returned, with the receipts.
What Max actually unlocks #
Two things gate behind Max. First, the models. The picker carries nine: Best, Sonar 2, GPT-5.4, GPT-5.5, Gemini 3.1 Pro, Claude Sonnet 4.6, Claude Opus 4.8, Kimi K2.6, and Nemotron 3 Ultra. Two of them, GPT-5.5 and Claude Opus 4.8, are Max-only. Second, the modes. The search box switches between Search, Deep Research, Learn step by step, and Model Council, which is also Max-only.
The everyday answer is genuinely strong #
We asked Claude Opus 4.8 to compare a year of running Llama 3.3 70B locally on a Mac Studio against a frontier API, with the math. It did not hand-wave. It separated the two cost structures (local is a fixed cost with near-zero marginal cost; API is pure pay-per-use), priced a Mac Studio M3 Ultra with 192GB at about $5,800 plus roughly $95 a year of power, modelled light, moderate, and heavy API usage in a table, and then added the caveat that mattered: GPT-5.5 outperforms Llama 3.3 70B, so this was never a quality-equal trade. Fifteen sources, sound reasoning.
Model Council is the standout #
This is the feature that justifies the tier for a certain kind of user. Ask a contested question and Model Council convenes three frontier models, here GPT-5.5, Claude Opus 4.8, and Gemini 3.1 Pro, each reasoning independently, then synthesizes them. We asked whether a bootstrapped solo founder should build on a frontier API or self-host an open model.
The output is not one answer. It is a “Where Models Agree” table (all three converged on starting with an API and deferring self-hosting, with the GPU utilization trap as the key reason), a “Where Models Disagree” table, a “Unique Discoveries” list (Opus flagged that TGI entered maintenance mode in December 2025; Gemini flagged GPT-5.5 cached-input pricing at $0.50 per million), and a synthesis. It even noted that GPT-5.5 “naturally highlights its own model’s strengths,” a sharp piece of self-awareness most single answers never give you.
Deep Research is consultant-grade, with a catch #
We asked Deep Research for the real total cost of ownership of a small AI SaaS in 2026. It ran eight steps, fired off three rounds of searches (inference pricing, churn benchmarks, Stripe fees and hidden costs) with “Insights” passes between them, and wrote a structured report: an executive summary, four pillars (model inference, cloud hosting, Stripe fees, churn) each with their own pricing tables, and a synthesized cost-stack model at ~$500K ARR. It is the kind of brief you would otherwise pay a consultant for, and you can export it.
The catch is the sourcing. Deep Research pulled from dozens of low-authority SEO blogs (Groovy Web, Bananalabs, contracollective, zendevy, churntools) alongside the few solid ones, and at one point cited a politics magazine for a claim about Anthropic’s agent billing. The structure is excellent. The inputs need a human pass.
Computer is an agent that actually ships #
Computer is the agent. You describe a job and it plans a task list, researches, builds an artifact, and shares it. We asked for a spreadsheet comparing six mini PCs for local LLMs, with citations per row. It wrote a four-step plan, researched across 47 sources, and produced a real Excel file (Mac Studio M3 Ultra: 512GB, $9,499, ~14-16 tok/s; Mac mini M4 Pro: 64GB, $2,299, ~10-15 tok/s, and so on) in 1 minute 21 seconds, then offered follow-ups including a monthly price monitor. The same engine builds slide decks, websites, and reports.
Academic fixes the thing that was broken #
The source-quality problem has one clean fix: Academic mode. We asked whether peer-reviewed studies show AI coding assistants improve productivity. The sources shifted to arxiv, SSRN, Reuters, and InfoQ, and the answer was rigorous: the Microsoft and Accenture randomized trials (4,867 developers, +26% tasks), the METR slowdown (experienced developers took 19% longer while believing they were 20% faster), and a 37-study systematic review, laid out in a comparison table. That is the same evidence our own Study on AI coding speed is built on, sourced properly.
The other half of the data story is Connectors. From the Pro tier up, Perplexity includes premium research data that normally costs a fortune on its own (PitchBook, CB Insights, Statista, Wiley journals, and a US case-law library), and connects to your Google Drive, Gmail, and Dropbox so it can answer over, and act on, your own files. Worth being precise: this is a Pro feature, not a Max one, which is part of why Pro is enough for most people.
Finance is a terminal in the chat #
The Finance surface is a Bloomberg-lite: live futures and the VIX, real-time quotes, market-news summaries, a screener, earnings, congressional-trade tracking, a watchlist, and a portfolio you can sync through Plaid. You can ask plain-English questions about any of it.
There is more we drove and will not belabor: Spaces (custom collections with their own instructions and files), Workflows (pre-built agent recipes for financial models, clinical briefs, store optimization), Memory (a persistent, structured store of what you have worked on), Skills, and the Health and Patents surfaces.
Where it still leaks #
The one finding from the free-tier review survives Max intact: the everyday search and Deep Research over-trust the open web. Perplexity sells “the most trusted sources,” but in practice the default sourcing skews toward SEO content farms, and the deeper the research, the more of them it rakes in. Academic mode and the premium connectors are the antidote, but they are opt-in. If you run a default query and act on the headline number, you are trusting a blog Perplexity found, not a source it vetted. Cited is not the same as sourced.
Real cost #
The free tier exists but gates after a few anonymous searches. Pro is $20 a month ($17 when billed annually) and already includes the Computer agent, the premium data connectors, and most of the models, which is the core of what most people need. Max is $200 a month ($167 annually), ten times the monthly Pro price, and the extra is narrow but real: the frontier reasoning models (GPT-5.5 and Claude Opus 4.8), Model Council, and much higher limits for running Deep Research and the agent at scale. The gap is not the quality of an ordinary answer, or even the premium data. It is the top models, the council, and the ceiling on heavy use.
The verdict #
Situational, and that is not a knock. Perplexity Max is the deepest toolset in consumer AI right now: Model Council, Deep Research, and the Computer agent each did real, useful work in our tests, and the included premium data is a genuine edge for anyone who needs it. If you are a founder, analyst, or researcher who will run deep reports, convene the model council on hard calls, and let the agent build things most weeks, the $200 pays for itself.
If you are not, you will spend $200 to use a $20 product. Buy Pro, turn on Academic mode for anything that matters, and keep clicking through to the sources. The tools are extraordinary. Most people just do not need this many of them.