Perplexity API pricing: Sonar retires September 27, 2026, replaced by five Agent API presets

Perplexity retires Sonar Chat Completions on September 27, 2026. Its replacement, the Agent API, is cheaper for three of the four migration paths and nearly three times more expensive for the fourth: Sonar Deep Research to the high preset, at $0.0825 against $0.029 for a matching call.

Read the preset row before setting a production default.

Rows are sorted by estimated cost per call, descending, the reverse of the fast-to-xhigh reading order. Rates shown are the short-context tier, under 272,000 tokens. The estimate uses each preset's own documented example token count and excludes tool calls. perplexity/sonar, Perplexity's own model, sits outside the preset system entirely: $0.25 input and $2.50 output per million tokens, $0.0625 cached, no request fee.

Basis and method

Figures in this table and article come from Perplexity's own documentation and help center, collected on September 25, 2026: the pricing page, the Sonar migration overview, the Agent API model catalog, the Sonar-to-preset benchmark comparison, and the API payment and billing help article. No figure here was produced by testing. The estimated cost per call multiplies each preset's documented example token count by its documented per-token price, then adds one web_search call at $0.0025 where a tool call is assumed; the Search API's effective per-query cost divides the posted price by the number of queries a batched request carries. Full methodology: https://llmwaves.com/methodology.

This article ranks the four migration paths by cost change against the Sonar model each one replaces, not by sticker price alone, the same against-predecessor lens the API pricing comparison hub applies across every vendor it covers.

The replacement for Sonar has a name, a price, and a deadline

Sonar Chat Completions stops taking new load on September 27, 2026, a date Perplexity's own migration guide states without qualification.

The guide also states the only official mapping from a Sonar model to an Agent API preset: Sonar becomes fast, Sonar Pro becomes low, Sonar Reasoning Pro becomes medium, and Sonar Deep Research becomes high.

A fifth preset, xhigh, has no Sonar predecessor at all. perplexity/sonar keeps running as its own model, billed at $0.25 input and $2.50 output per million tokens with no request fee, a rate that undercuts every one of its own retiring siblings.

The request fee itself disappears. Sonar charged $5, $8 or $12 per 1,000 calls depending on search context size, on top of token price.

Every preset instead pays only for the tools it actually calls, per Perplexity's own pricing page: $0.0025 per web_search, $0.0005 per fetch_url, $0.005 per people_search or finance_search, and $0.03 per sandbox session.

A call that used to carry a flat fee whether it searched once or ten times pays only for the searches it runs.

The presets do not run Perplexity's own model

Migrating from Sonar to a preset is not a move from one Perplexity model to a newer one.

Perplexity's own Agent API model catalog names the model behind each preset directly. The tool layer, the search behavior, and the citation style are Perplexity's.

The language model doing the reasoning is not.

The Luna tier: fast, low and medium

Fast, low and medium all route to one model, openai/gpt-5.6-luna. Luna prices input at $0.20 per million tokens and output at $1.20, the cheapest tier on the catalog. Three presets, one bill.

The Sol tier: high and xhigh

High and xhigh switch entirely to openai/gpt-5.6-sol. Sol prices the same pair at $5.00 and $30.00, a 25-fold jump on input alone.

A developer moving from medium to high is not buying more Perplexity. They are buying a different, far costlier third-party model wrapped in the same tool set, and the preset name gives no hint of that switch.

Three presets cost less. One costs nearly three times more

Perplexity's own benchmark comparison, published alongside the migration guide, reports the same asymmetry from the quality side: sonar-pro to low costs roughly 24% to 36% more per request on two named evaluations and 17% less on a third; sonar-reasoning-pro to medium costs 1.9 to 2.7 times more per request; sonar to fast ranges from 11% cheaper to 48% costlier depending on the evaluation.

Scores rise in every mapped pair. Whether that is "more quality per dollar," Perplexity's own phrase for the launch, depends entirely on which pair a workload happens to sit on.

No Sonar model ID survives the switch, and rate limits move too

No model ID named sonar-pro, sonar-reasoning-pro or sonar-deep-research exists on the Agent API.

Only perplexity/sonar and the preset system remain. A request built against the old alias does not fail over to a preset on its own; the substitution is a code change, not a routing default, so a developer who ships the endpoint swap without also swapping the model field inherits a 404, not a preset.

An open request for continued sonar-pro naming, filed on Perplexity's community forum on September 21, 2026, had no staff reply as of September 25, 2026.

The Search API is a separate product, and batching cuts its price to a third

The Search API is priced and billed separately from both Sonar and the Agent API: $5.00 per 1,000 requests on the standard tier, $1.00 per 1,000 requests on the Fast tier, which caps results at 20 per query.

A standard-tier request can carry up to five queries at once, billed as a single request but counted as five units against the rate limit.

Sending five queries per request buys Fast-tier pricing without leaving the standard tier, and without its result cap.

What else changes: credits, embeddings, and team seats

A Perplexity Pro subscription has not included a bundled API credit since February 2026; Perplexity's own API payment and billing help article states plainly that a subscription is not required to purchase credits or use the API.

Unused credits can still be refunded within 14 days of purchase, provided they have not been spent, and an account can enable automatic top-up when its balance drops below $2.

Two embedding models sit outside every other price on this page: pplx-embed-v1-0.6b at $0.004 per million tokens and pplx-embed-v1-4b at $0.03.

A Router API that would pick a preset automatically remains in private preview. None of this touches Enterprise, the seat-based subscription for the consumer product, which starts near $40 per seat per month and is billed on a wholly separate schedule from any figure above it.

Not every reader gets a clean answer

Two rows carry no rank at all. Latency and non-English performance both come back insufficient data, not because either is bad, but because Perplexity has not published a number for either one.

The two self-hosting rows fail for the same single reason: an open-weight release that does not exist.

Is the Perplexity API free?

No. A small number of trial credits may appear on a new account, but ongoing use bills per token and per tool call with no free tier, and a Perplexity Pro subscription carries no bundled credit.

Does a Perplexity Pro subscription include API credit?

Not since February 2026. The API payment help article states a subscription is not required to buy credits or use the API. It treats the two as separate purchases.

Is a five-query Search API request billed once or per query?

Once, at the standard tier's per-request price. It still counts as five units against the rate limit, so the discount is real but the throughput cost is not waived.

What replaces Sonar Pro on the Agent API?

The low preset. It runs openai/gpt-5.6-luna, not a Perplexity model, and costs roughly 85% less per call than Sonar Pro did at matching token counts.

One of the prices on this page stops existing on September 27, 2026

Every figure above was current on September 25, 2026. Two days later, the row for Sonar, Sonar Pro, Sonar Reasoning Pro and Sonar Deep Research stops applying to new calls, and the preset row is what remains.

Check the current rate card before that date changes what this page describes