Claude vs Perplexity

Claude and Perplexity get cross-shopped because both start at $20/mo and both answer questions typed into a box. That framing expired in 2026. Perplexity now ships a first-party MCP server, which means the real decision is no longer which product to subscribe to — it is whether you buy Perplexity as a second seat or install it as a tool inside Claude. The two compose in a way ChatGPT and Perplexity do not, and the pricing math changes once you notice.

What changed since the last version of this page

Perplexity became callable. The @perplexity-ai/mcp-server package exposes four tools against Perplexity’s Search and Agent APIs — perplexity_search, perplexity_ask, perplexity_research, and perplexity_reason. Claude, Claude Code, or any MCP client can run cited web research without a human opening perplexity.ai. It needs an API key from console.perplexity.ai and bills separately from any Pro seat.

Perplexity’s ladder stretched in both directions. Max landed at $200/mo with Model Council — one query fanned across several frontier models and the answers synthesized — and Enterprise Max at $325/seat/mo. At the other end, the Comet browser dropped its paywall in March 2026 and now runs free on Mac, Windows, iOS, and Android.

Claude’s long-context story got more precise, not bigger. A 1M-token window is the default on Claude 4.6 and later models, with no beta header and no long-context surcharge — a 900k-token request bills at the same per-token rate as a 9k one. But that is an API property. claude.com/pricing lists a 200k window on the individual plans. The previous version of this page implied a 1M window inside the chat product; that was wrong and is corrected here.

Claude grew a second agent surface. Cowork — a file-system-level agent that works across your files and connected apps rather than your repo — shipped on desktop in January 2026 and reached web and mobile in July for Max subscribers.

The asymmetry is the whole page. Perplexity is a server. Claude is a client. On the ChatGPT comparison that produces a standoff; here it produces a stack.

Where Claude wins

• Documents that do not fit anywhere else. The 1M-token window at standard pricing, plus up to 600 images or PDF pages in a single request, covers a deal room of contracts or a quarter of call transcripts. Perplexity is built to reach out to the web, not to hold your corpus.
• It is the thing that orchestrates. Claude Code, Cowork, Skills, an MCP client, and the Agent SDK are a runtime you program. Perplexity’s agentic story is a browser and an API — good ones, but you build around them rather than inside them.
• Deliverables, not answers. Redlines, first drafts, working code, structured extraction into a schema someone downstream consumes. Perplexity’s synthesis is tuned to resolve a question, which is a different output.
• It can call the other one. There is no configuration in which Perplexity runs Claude Code.

Where Perplexity wins

• Citations are the shape of the product, not a mode it enters. For a memo headed to a GC, a regulator, or a board, that decides whether the work survives being forwarded.
• Spaces. A research project with pinned sources persists across sessions and hands off to a colleague. A Claude Project is close, but the Space is built around sources rather than instructions.
• A free browser with an assistant on the page. Comet costs nothing on four platforms. Nothing in the Claude lineup occupies that surface.
• Retrieval you buy by the request. Sonar sells web research as metered capacity. When the consumer of that research is an agent rather than a person, seats are the wrong unit and Perplexity is the one selling the right one.
• Internal Knowledge Search. On the Enterprise tiers your own documents sit beside the live web in one answer, under a no-training guarantee.

Pricing reality

Entry is a genuine tie and the ladders diverge immediately.

• Individual — Claude Pro $20/mo, or $17/mo billed annually at $200 up front. Perplexity Pro $20/mo, $200/yr. A wash on both billing modes.
• The missing middle — Claude Max starts at $100/mo for 5× Pro usage and reaches $200/mo for 20×. Perplexity jumps straight from $20 to $200 with nothing between. If your usage is three times a Pro seat, only one of these two sells you a tier for it.
• Team seats — Claude Team standard is $25/seat/mo, $20 billed annually; premium seats are $125/mo, $100 annually. Perplexity Enterprise Pro is $40/seat/mo ($400/seat/yr). Roughly 2× in Claude’s favor at the standard rung, and what the Perplexity premium buys is Internal Knowledge Search and the training guarantee.
• Top of the ladder — Perplexity Enterprise Max is $325/seat/mo ($3,250/seat/yr), a fixed and large number. Claude Enterprise is $20/seat plus all model usage billed at API rates — a lower floor and a ceiling set by how hard your team works it. These are different billing shapes, and a per-seat comparison between them is not meaningful without a usage estimate.
• Metered — Sonar runs about $1 per million tokens at the base tier, $3 in / $15 out per million for Sonar Pro, plus roughly $5-12 per 1,000 web-retrieval requests. Claude’s API is $5/$25 per million for Opus 5 and Opus 4.8, $2/$10 for Sonnet 5 through August 31 2026 and $3/$15 after, and $1/$5 for Haiku 4.5, with web search at $10 per 1,000 searches, web fetch free beyond tokens, and the Batch API halving both. Sonar Pro and Sonnet land on the same $3/$15; the difference is which side charges you for retrieval.

Two individual seats is $40/mo. The cheaper configuration most people miss is one Claude Pro seat at $20 plus a metered Perplexity key, which puts cited retrieval inside the tool already doing the work.

Implementation effort

Claude as a chat product needs nothing. Claude as an agent — MCP servers, Skills, service accounts, a review of what it is allowed to write to — is a real project measured in days, and the line worth forecasting is the token bill rather than the seat count.

Perplexity as an app needs nothing either. Perplexity as an MCP server inside Claude is an API key and one config entry: under an hour, plus a metered bill that did not exist before. Enterprise Internal Knowledge Search is the only piece here that is a genuine connector project.

Watch-outs

• Paying twice for one job. Running the Perplexity MCP server inside Claude while keeping Perplexity Pro seats bills a seat and a meter for the same research. Guard: assign each person one surface, and read the API console spend at the end of the first month before renewing seats.
• Metered retrieval inside a loop. The web-retrieval fee bills per call, not per session, so an agent that retries is billable while it fails. Guard: set a spend cap on the API key before it goes anywhere unattended.
• Treating a 1M window as 1M of accuracy. Recall degrades as a context window fills — Anthropic’s own documentation names this and recommends compaction over stuffing. Guard: curate what goes in, and pass source documents rather than whole archives.
• Citation theater. Both products will return a source list containing a page that does not support the claim it is attached to. Guard: click through every citation carrying a number before the work leaves your hands.

Verdict

• Pick Claude if the output is a deliverable — a draft, a redline, an extraction, a script — or if the input is longer than anyone wants to read.
• Pick Perplexity if the output is an answer whose provenance will be questioned, or if a non-technical researcher needs Spaces and Comet rather than a chat window.
• Choosing one, with one budget line: Claude Pro at $20/mo, then add Perplexity through its MCP server on a metered key. It buys cited retrieval inside the tool doing the work and skips the second seat entirely. Reverse it only if the person spends most of the day researching and none of it building.
• Two seats at $40/mo when the researcher and the builder are different people.
• Neither, if the answers live in your own systems rather than on the open web — that is an enterprise retrieval problem, and the shortlist is Gemini Enterprise and Glean. If open-web search is the entire job, best AI search engines is the right page. If the incumbent you are actually displacing is OpenAI’s, read ChatGPT vs Perplexity instead.