--- title: "Perplexity Sonar API pricing in 2026: the rate card after September 27" category: "pricing" url: https://keirolabs.cloud/blog/perplexity-sonar-api-pricing-2026 --- ## The 60-second version Sonar Chat Completions support ended on September 27, 2026, and most of the pricing pages on the web still have not noticed ([migration guide](https://docs.perplexity.ai/docs/agent-api/migrate-from-sonar/overview)). We pulled the full current rate card from Perplexity's own documentation on October 12, 2026. The short read: - **What actually ended.** Synchronous and streaming Sonar requests keep working, gradually reformulated as Agent API requests behind the scenes. Asynchronous Sonar is gone; the docs point it at background mode. The recommended path is the Agent API, whose presets map to the Sonar models: Sonar and Sonar Pro map to `fast`, Sonar Reasoning Pro to `low`, Sonar Deep Research to `high` ([docs](https://docs.perplexity.ai/docs/agent-api/migrate-from-sonar/overview)). - **The current request fees, which is what most pages get wrong.** Sonar bills $5, $8 or $12 per 1,000 requests depending on the search context size. Sonar Pro bills $6, $10 or $14. If you read $5/$8/$12 quoted for Sonar Pro, you are reading a page last verified in 2025 ([models](https://docs.perplexity.ai/docs/agent-api/models)). - **What a finished answer costs.** A plain Sonar answer at low context runs about $0.007: a $0.005 request fee plus roughly $0.002 of tokens at $1 per million in and out. A Sonar Pro answer at medium context runs about $0.03. The request fee is the bill; the tokens are the tip. - **The flat option.** The Search API is $5 per 1,000 requests, up to 5 queries per request count as one billing unit, failed requests are not billed, and there are no token charges at all. - **Rate limits moved to usage tiers.** The Agent API meters 1 to 33 queries per second on a rolling one-second window, on tiers that advance with lifetime spend and never downgrade. The Search API runs a flat 50 query units per second with a 50-unit burst, independent of tier ([docs](https://docs.perplexity.ai/docs/admin/rate-limits-usage-tiers)). ## What ended on September 27 For two years "Sonar" meant one thing: a Chat Completions endpoint that took a prompt, searched the web, and answered with citations, billed per token plus a request fee. On September 27, 2026, Perplexity ended support for that endpoint's shape ([migration guide](https://docs.perplexity.ai/docs/agent-api/migrate-from-sonar/overview)). The transition is gentler than the announcement sounds. Existing synchronous and streaming requests keep working, and Perplexity is reformulating them as Agent API requests gradually. What died outright is async Sonar: fire-and-forget requests with a callback are no longer supported, and the docs point that use case at the Agent API's background mode. The Agent API is not just a rename. It gives the request a model, a preset, and a stack of tools the old endpoint never had, which is why the pricing story now has more moving parts than a token rate card. The presets are the bridge: pick the preset that matches the Sonar model you were using and the behavior carries over. The mapping, from Perplexity's own migration table: Sonar and Sonar Pro map to the `fast` preset, Sonar Reasoning Pro to `low`, Sonar Deep Research to `high`. One model quietly left the lineup: plain Sonar Reasoning. The current card lists Sonar, Sonar Pro, Sonar Reasoning Pro and Sonar Deep Research. If your code calls `sonar-reasoning`, it is calling a model that no longer appears in the pricing documentation, and the migration guide's reasoning preset is the replacement. ## The current rate card Every figure below is from Perplexity's models documentation, verified October 12, 2026 ([models](https://docs.perplexity.ai/docs/agent-api/models)). | Model | Input | Output | Citation tokens | Reasoning tokens | Request fee per 1,000 (low / medium / high context) | |---|---|---|---|---|---| | **Sonar** | $1 / M | $1 / M | not charged | not charged | $5 / $8 / $12 | | **Sonar Pro** | $3 / M | $15 / M | not charged | not charged | $6 / $10 / $14 | | **Sonar Reasoning Pro** | $2 / M | $8 / M | not charged | $3 / M | $6 / $10 / $14 | | **Sonar Deep Research** | $2 / M | $8 / M | $2 / M | $3 / M | none; $5 per 1,000 search queries instead | The request fee scales with the search context size you configure per request, and low is the default. That makes the fee the dominant line on short answers and a rounding error on long ones, which is the shape of the whole card: a one-line factual answer and a 3,000-word research brief pay the same fee at the same context size. REQUEST FEE PER 1,000 REQUESTS, BY SEARCH CONTEXT SIZE (OCT 2026) Sonar Pro (high) $14 Sonar (high) $12 Sonar Pro (medium) $10 Sonar (medium) $8 Sonar Pro (low) $6 Sonar (low, default) $5 Search API (flat) $5 Sonar Reasoning Pro shares the $6/$10/$14 ladder. Sonar Deep Research has no request fee: it bills $5 per 1,000 search queries plus $3/M reasoning and $2/M citation tokens. Token rates shown are per 1M. Source: docs.perplexity.ai/docs/agent-api/models, 2026-10-12. ## What a Sonar answer costs The token meters matter less than you think. Here is the arithmetic on three typical answers, using published rates. Volumes are illustrative; the fees and the per-million rates are not. **A quick factual answer, Sonar, low context.** Say 1,500 input tokens and 500 output tokens. Tokens: 1,500 at $1 per million is $0.0015, plus 500 at $1 per million is $0.0005, so $0.002. Fee: $0.005. Total: about **$0.007 per answer**. **The same answer on Sonar Pro, medium context.** Fee: $0.010. Tokens: 3,000 input at $3 per million is $0.009, and 1,000 output at $15 per million is $0.015. Total: about **$0.034 per answer**. Sonar Pro's output rate is 15x its input rate, so the bill scales with how much the model says, not how much you ask. **A Sonar Reasoning Pro answer, low context.** Fee: $0.006. Tokens: $2 per million in, $8 per million out, and reasoning tokens at $3 per million. A 2,000-token question with 20,000 reasoning tokens and a 1,500-token answer runs about **$0.08**, and the reasoning meter is the one that moves. Reasoning work is where token costs quietly double. **A Sonar Deep Research run.** No request fee. The meters: $5 per 1,000 search queries, $3 per million reasoning tokens, $2 per million citation tokens, $2 in and $8 out per million. A run that fires 10 searches on a 10,000-token prompt and produces 20,000 reasoning tokens, a 10,000-token report and 8,000 tokens of citations costs roughly $0.05 + $0.06 + $0.02 + $0.08 + $0.016, about **$0.22 per run**. Deep Research is priced like a job, not a chat, because that is what it is. The pattern across all four: on short answers the request fee is 60 to 75% of the bill. Per-token pricing is how Sonar markets; per-request fees are how Sonar invoices. ## The Agent API bill, decoded The Agent API charges the model you picked at its own published token rates, plus the tools the agent used. Perplexity hosts third-party models at each provider's list price ([models](https://docs.perplexity.ai/docs/agent-api/models)), a sample as of October 2026: | Model | Input per M | Output per M | |---|---|---| | claude-opus-5-5 | $4 | $20 | | claude-sonnet-5-5 | $2 | $10 | | gpt-6-luna | $0.10 rising to $0.20 above 272k | $0.50 to $0.75 | | gemini-3.5-flash | $1.50 | $9 | | grok-4.20 | $1.25 to $2.50 | $2.50 to $5.00 | | perplexity/glm-5.3 | $1.40 | $4.40 | | perplexity/kimi-k3 | $3 | $15 | | perplexity/nemotron-3-ultra | $0.25 | $2.50 | Several of those are tiered: the input rate steps up past a context threshold, around 200k tokens by default, with smaller thresholds on some models. Cache reads come discounted, and on the Sonar models the request fee applies on top of whatever model is running. Then the tools. This is the part of the bill the old Sonar never had, and it is metered per invocation ([web_search docs](https://docs.perplexity.ai/docs/agent-api/tools/web-search)): - **web_search: $2.50 per 1,000 invocations**, or **$1.00 per 1,000 with search_type "fast"**. The docs say it plainly: "Fast calls cost $1.00 per 1,000 invocations instead of $2.50, plus model tokens." - **finance_search: $0.005 per invocation**, so $5 per 1,000, the same flat rate as the Search API. - **fetch_url, image_search and people_search are priced per invocation** too; check the tools pages before you budget, because the rates differ per tool. - **sandbox** bills per 20-minute container session, plus per search query run inside it. An agent that runs a gpt-6-luna turn with three standard web searches adds $0.0075 of tool calls to the token bill. Trivial once, real at a million runs. ## Search API: the flat option If the Agent API's meter stack sounds like a lot, the Search API is the opposite: one number. **$5 per 1,000 requests.** Up to 5 queries in one request count as one billing unit. Invalid requests, rate-limited requests and upstream failures are not billed. A successful response with zero results is billed, because the search did run. And there are no token charges, because the Search API does not use tokens; it returns search results and you do the synthesis. The catch is the work that lands on you. The Search API gives you results, not an answer: no synthesis, no citations formatted for your model, no reasoning. It replaces the retrieval layer of Sonar at roughly half the cost of a single low-context Sonar request fee, and leaves the writing to your model and your prompt. A cheaper fast search type exists on both products for when freshness matters more than depth. ## Rate limits moved to usage tiers The old Sonar had one rate limit story. The 2026 stack has two, and they are independent. **The Agent API meters by usage tier.** Tier 0 allows 1 query per second, tier 1 allows 3, tier 2 allows 8, tier 3 allows 17, and tiers 4 and 5 allow 33, enforced on a rolling one-second window. Tiers advance with cumulative lifetime API spend and never downgrade ([docs](https://docs.perplexity.ai/docs/admin/rate-limits-usage-tiers)). Your rate limit is a function of your invoice history, which is either a loyalty program or a leash, depending on your build. **The Search API ignores the tiers.** It runs a flat 50 query units per second with a 50-unit burst on every account. A multi-query request consumes one unit per query inside it, so a 5-query request spends 5 units of the 50. And here is the trap worth bolding: **the billing unit is not the rate unit.** A 5-query Search API request bills as one request, $0.005, and rate-limits as five units. Teams that size their concurrency off the invoice discover the difference at the 429. For the full limits of every provider in one table, we published that separately in [Search API rate limits in 2026](/blog/search-api-rate-limits-2026). ## Sonar versus a search-plus-extract pipeline The honest question is not "what does Sonar cost" but "what does a cited answer cost", because there are two ways to build one. **Buy it finished.** Sonar at low context: about $0.007 per answer. Sonar Pro at medium: about $0.034. Perplexity's answer quality leads the independent agentic search indexes, 80 on the September 2026 [Artificial Analysis](https://artificialanalysis.ai/agents/search-api) run where the closest competitor in that table, Keirolabs, scored 78 ([our own comparison](/blog/web-search-apis-for-agents-2026)), so part of what you pay is a genuinely better retrieval and synthesis loop, plus one bill, zero pipeline code, and citations in the response shape your model expects. **Build it from parts.** A search call, an extraction call, your own model doing the synthesis. On Keirolabs' published rates, the /answer endpoint (search plus synthesis, 5 credits) runs $0.012 per answer on the Essential plan and $0.004 on Startup, and the raw pipeline of 0.1-credit search plus 3-credit extraction runs about $0.0025 per page on Startup ([pricing](https://keirolabs.cloud/pricing)). Against Sonar's $0.007, that is 3x cheaper at the finished-answer level and about 3x cheaper again if you own the synthesis. We sell one side of this comparison, so weight the tradeoffs yourself: the pipeline buys you control over retrieval, domains, freshness and the synthesis model, and it bills per component so the cost is visible line by line. Sonar buys you the finished answer and the quality scores that come with it. At low volume, the difference is coffee money; at a million answers a month, it is the difference between $7,000 and $2,500. ## The stale-data tour Because Sonar's pricing changed twice this year and the retired names still rank, the web is full of confidently wrong numbers. Here is what we found while fact-checking this post, so you can check whatever page you are reading against the same list. **Retired model names still quoted.** Sonar Small, Sonar Large and Sonar Huge Online disappeared from the lineup long before this year, and pages still list their per-token rates as current options. **The old Sonar Pro request fee.** Sonar Pro's request fee is now $6/$10/$14 per 1,000 requests by context size. The $5/$8/$12 ladder quoted across most comparison pages belongs to plain Sonar and to 2025. Any page applying $5/$8/$12 to Pro is describing a card that has not existed for months. **Plain sonar-reasoning.** It is gone from the current lineup; only Sonar Reasoning Pro remains. Pages that price both are one revision behind. **And the newest failure mode: answer engines themselves.** While fact-checking this post we asked a popular answer API for Perplexity's rates and got Sonar Pro at $5/$5 per million with request fees of $1/$2/$3. No published Perplexity document has ever matched those numbers. The retrieval layer for pricing research is now as stale as the pages it reads, which is the strongest argument we know for citing the vendor's own rate card and checking its date. The fix is procedural, not clever: any pricing number you plan to budget against should trace to the vendor's own documentation with a date. This post's date is October 12, 2026, and every figure above links to the page it came from. ## The boring takeaway Sonar did not die on September 27; it moved into the Agent API and the pricing got more granular. Budget it in this order: the request fee first, because on short answers it is most of the bill; the tool invocations second, because agents multiply them quietly; and the tokens last, because at Sonar's $1/$1 they barely register until Deep Research starts metering reasoning. If you want finished, cited answers and one invoice, Sonar through the Agent API is still the benchmark, and its quality scores justify the premium. If you want the retrieval layer alone, the Search API at $5 per 1,000 with no token charges is one of the flattest prices in the category. And if you own a synthesis model and want the parts priced line by line, a search-plus-extract pipeline undercuts both. Every number on this page was verified against Perplexity's own documentation on October 12, 2026, and carries its source link. When the card changes, and it changed twice this year, this page gets rechecked. Send corrections with a link and they get applied.