Perplexity API Pricing (November 2026): Sonar Is Retired, Here's What Each Effort Level Costs

The short answer

As of November 8, 2026, there's no "Sonar pricing" anymore. Perplexity ended Sonar Chat Completions support on September 27, 2026. Its Agent API bills per token at the model's rate plus a fee per tool call, and you pick an effort level from fast to ultra. Perplexity's own numbers put a fast request at about $0.0014, a high request at $0.12 to $0.14, and an ultra task at $21.71. The Search API alone is $5 per 1,000 requests.

Search "Perplexity API pricing" and most of what comes back is a Sonar rate table from earlier this year. That table describes a product Perplexity has retired. If you're budgeting a research tool, a lead enricher or a support bot on those numbers, you're budgeting the wrong thing. And on November 5, Perplexity cut the price of every effort level again, so even the summer numbers are stale.

The four Perplexity APIs and how each one bills

From Perplexity's pricing page, read November 8, 2026:

Two details on the Search API that help the bill: one request can carry up to five queries and still counts as one request, and failed or rate-limited requests aren't charged. A request that returns no results is charged.

What each effort level costs, next to the Sonar model it replaced

This is the table nobody else lines up. Perplexity's "Agent API vs Sonar benchmarks" page lists a cost per request for every effort level and every retired Sonar model, measured on two public benchmarks. Read November 8, 2026:

Ultra isn't on that page. Perplexity's November 5 post gives one number for it: $21.71 per task on its WANDR research benchmark, running three agents in parallel and a fourth that merges what they found.

Two caveats before you budget on this. These are benchmark questions, not your questions, so a short lookup will cost less and a sprawling one more. And the page doesn't say whether its numbers were measured before or after the November 5 price cut. Treat them as the order of magnitude, then check your own bill.

What's inside each level

Each effort level is a bundle Perplexity manages: a model, a step budget, tools and a system prompt. From the effort docs, November 8, 2026:

• fast: OpenAI's gpt-6-luna, 1 step, Fast Search. Single facts and quick summaries.
• low: gpt-6-luna, up to 5 steps, web search plus one page fetch. Everyday research questions.
• medium: gpt-6-luna, up to 15 steps. Questions that chain several searches.
• high: gpt-6.1-terra, up to 40 steps and 40 tool calls. Deep research without running code.
• xhigh: Anthropic's Claude Opus 5.5, up to 300 steps, with a code sandbox and finance search. Open-ended agent work.
• ultra: Claude Opus 5.5 with several agents in parallel. No fixed configuration; you call it by name.

When you call a level by name, Perplexity can swap the model under you. Its docs say updates "target the same cost band," but if you need a fixed price, pin a version (for example effort="fast@2026-10-16.0").

Two billing details that surprise people

1. The fast level runs on the priority tier. Its configuration sets the service tier to priority, which Perplexity bills at 2x the listed token rate on models that support it. It's still the cheapest level by a wide margin, but if you copy fast's settings to build your own setup, drop the priority tier and you pay half the token rate.
2. You can't price a high run from the rate card. High runs on gpt-6.1-terra, and when we checked on November 8 that model had no row on the Agent API Models page. Perplexity's pricing page says completed responses report usage.cost.total_cost. Log that field from day one. It's the only reliable way to see what a high request costs.

Tool fees sit on top of tokens: $0.0025 per standard web search, $0.001 per Fast Search, $0.0005 per page fetch, $0.005 per people or finance search, and $0.03 per code sandbox session.

What 1,000 lookups a month costs at each level

Take a common small-business job: a tool that researches 1,000 inbound leads or prospect companies a month and writes a short brief on each. Using Perplexity's own per-request costs from the table above:

The spread from low to ultra is more than 1,800x. For a lead brief, low or medium is usually plenty. The expensive mistake isn't picking the wrong vendor, it's defaulting to the deepest level because the demo looked better. Run 20 real leads through low and 20 through high, compare the briefs, and only pay for the jump if your team can tell the difference.

If you're on the old Sonar Deep Research for this kind of job, the math moves in your favor: $0.57 to $0.61 a request there against $0.12 to $0.14 on high, Perplexity's stated replacement.

Perplexity vs building on Claude or OpenAI directly

The Agent API is a reseller layer with search built in. It runs OpenAI, Anthropic, Google and xAI models at each provider's own token rate (Claude Opus 5.5 is listed at $4 in and $20 out per million tokens, the same as Anthropic's API pricing), and adds Perplexity's search and tools. So the question isn't which is cheaper per token. It's whether you want to build the search part yourself.

If your tool needs fresh web facts with citations, Perplexity saves you the search plumbing. If it only needs to read your own documents, go direct: OpenAI's API pricing and Claude API vs OpenAI API cover that choice. And if you only want Perplexity for your team to ask questions in a browser, you don't need the API at all. That's a seat, covered in Perplexity pricing.

Where the money really goes is the wiring: which leads get researched, where the brief lands in your CRM, and who checks it before it reaches a customer. That's an AI implementation question, and it's worth getting right before the API bill is.

Today's micro-action

If anything in your stack still calls a Sonar model, open your logs and check whether those requests are being converted or failing. Perplexity says async Sonar requests are no longer supported at all. Then add usage.cost.total_cost to whatever you log per request, so next month's bill isn't a surprise.

Sources

Perplexity API docs, "Pricing," docs.perplexity.ai/docs/getting-started/pricing (read November 8, 2026); "Effort," docs.perplexity.ai/docs/agent-api/effort (read November 8, 2026); "Agent API Models," docs.perplexity.ai/docs/agent-api/models (read November 8, 2026); "Agent API vs Sonar benchmarks" and "Migrate from Sonar to the Agent API," docs.perplexity.ai/docs/agent-api/migrate-from-sonar (read November 8, 2026); Perplexity blog, "Improvements to every Agent API effort level," November 5, 2026 (read November 8, 2026). The 1,000-request monthly figures and the low-to-ultra spread are our arithmetic from Perplexity's published per-request costs.