Perplexity Sonar API Cost Calculator (2026)

Every Sonar API call adds two streams together. There is no platform fee, no minimum spend, no per-seat charge — but you absolutely cannot model your bill with the OpenAI-style token-only formula. The right shape:

``` cost = (input_tokens / 1,000,000) × input_price_per_M + (output_tokens / 1,000,000) × output_price_per_M + (requests / 1,000) × per_request_fee_per_1K ```

On Sonar (the entry tier), a 1,000-in / 500-out call with a medium-depth search at the $8/1K request rate bills as: $0.001 (input) + $0.0005 (output) + $0.008 (one search) = $0.0095 per call. The per-request fee is 84% of the total. On Sonar Pro the token side is bigger, but the per-request fee is still the line item that surprises teams who modeled the bill in a spreadsheet copied from an OpenAI calculator.

Sonar Reasoning Pro folds the search fee into the token price — useful when you want a predictable per-call cost or when the workload runs many shallow searches that would otherwise stack request fees. Sonar Deep Research adds a third stream: citation and reasoning tokens, billed separately and often dwarfing the prompt/answer tokens themselves on a long-form research run.

On a standard OpenAI or Anthropic call, the model reads your prompt and writes an answer. That's it. No external lookup, no infrastructure beyond GPU inference. Token pricing covers everything because nothing else happens.

On Sonar, every call kicks off a live search pipeline before the LLM even sees the prompt. Perplexity's index runs a query against the open web, fetches and ranks results, extracts the relevant passages, attaches citations, and only then hands a grounded context to the LLM. That entire pipeline costs real money to run — proxy bandwidth, parser compute, ranking models, citation extraction, freshness checks. None of that fits inside a token rate.

The per-request fee is the line item that pays for it. Sonar's $5-12 per 1,000 requests works out to roughly $0.005-$0.012 per search — competitive with standalone search APIs like Tavily (~$0.005/search) or Brave Search API (~$0.005/search), bundled with the inference instead of stitched together by you.

Practical implication: search depth is a cost lever. Sonar exposes low / medium / high search depth on most tiers. Low depth runs fewer parallel fetches and bills at the bottom of the per-request range. High depth runs more concurrent fetches, ranks more candidates, and bills at the top. For autocomplete-style 'what is X' calls, low depth almost always suffices. For research-grade synthesis, the high-depth premium pays for itself in answer quality. Pick consciously per call — defaulting every request to high depth roughly doubles your per-request bill with no quality lift on shallow queries.

Take a representative call — a 1,000-token prompt that returns a 500-token grounded answer with citations, roughly equivalent to a 750-word brief in and a 375-word reply out. With one medium-depth search at $8/1K requests, the per-call cost lands as:

Sonar: (1000 / 1,000,000) × $1.00 + (500 / 1,000,000) × $1.00 + (1 / 1000) × $8.00 = $0.001 + $0.0005 + $0.008 = **$0.0095 per call**.

Sonar Pro: 0.001 × $3.00 + 0.0005 × $15.00 + (1/1000) × $10.00 = $0.003 + $0.0075 + $0.010 = **$0.0205 per call**.

Sonar Reasoning Pro: 0.001 × $2.00 + 0.0005 × $8.00 + $0 (included) = $0.002 + $0.004 = **$0.006 per call**.

Sonar Deep Research: 0.001 × $2.00 + 0.0005 × $8.00 + (1/1000) × $5.00 + citation/reasoning tokens (often 5,000-15,000 extra output tokens) ≈ $0.002 + $0.004 + $0.005 + (10,000 / 1,000,000) × $8.00 ≈ **$0.091 per call**.

Notice the shape: Sonar Reasoning Pro is the cheapest per call because the search fee is included, even though its token rate is higher than plain Sonar. Sonar Deep Research is roughly 10x the cost of plain Sonar — that gap is almost entirely the citation and reasoning tokens, not the base price.

Multiply the per-call numbers by 100,000. This is a realistic mid-size workload — a research assistant feature in a SaaS app, a daily competitive-intelligence loop, a customer-facing 'cited answer' button:

Sonar: **$950/month**. Sonar Pro: **$2,050/month**. Sonar Reasoning Pro: **$600/month**. Sonar Deep Research: **$9,100/month**.

On 100k Sonar calls, the per-request fee alone is $800 — 84% of the total bill. Cutting search depth from medium ($8/1K) to low ($5/1K) drops the per-request line from $800 to $500, a 32% reduction with no model change. For workloads where most queries are short factual lookups, this is the highest-EV lever available on the entire Sonar billing surface.

On Sonar Pro, the same depth shift drops the per-request line from $1,000 to $600, and the total from $2,050 to $1,650 (20% off). Pro's higher token rate dilutes the per-request percentage, but the savings are still real money at scale.

Scale to 1M Sonar Pro calls per month — a serious production deployment, e.g., 33k active users × ~33 grounded answers per user per month:

Base bill: $3,000 input + $7,500 output + $10,000 per-request = **$20,500/month**. That's $0.0205 per call at scale.

Search-depth optimization (mix of 60% low-depth + 30% medium + 10% high): per-request weighted to $7.20/1K instead of $10/1K — drops the request line from $10,000 to $7,200, saving $2,800/month (14% off the total).

The biggest practical lever on Sonar Pro at scale is not the model tier choice — it's deciding which queries actually need the open-web crawl at all. If 40% of your traffic is 'recall' (the answer is in your own docs or in the LLM's pretraining), route those queries to a non-Sonar endpoint and only spend the per-request fee where freshness or citation matters. A hybrid stack — Sonar Pro for fresh/cited queries, GPT-5.4-mini for everything else — typically cuts a Sonar-heavy bill 30-50% with no user-facing quality loss.

Sonar Deep Research is built for long-form autonomous research — the kind of multi-hop synthesis where the answer is a 2,000-word report with 30+ citations. Take a workload of 10,000 Deep Research runs per month (e.g., a market-research SaaS that runs one per customer per day):

Per-run shape: ~3,000 input tokens (prompt + tools), ~2,000 output tokens (the actual report), ~12,000 citation + reasoning tokens (the chain-of-thought + cited passages the model traversed to produce the report), one Deep Research search at $5/1K.

Per-run cost: 0.003 × $2.00 + 0.002 × $8.00 + 0.012 × $8.00 + (1/1000) × $5.00 = $0.006 + $0.016 + $0.096 + $0.005 = **$0.123 per run**. At 10k runs/month: **$1,230/month**.

Citation + reasoning tokens are 78% of the bill. This is the line item that distinguishes Deep Research from every other tier — and it is the line item teams most often miss when modeling the cost. If your prompts encourage exhaustive chain-of-thought ('think step by step, cite every claim'), reasoning tokens balloon. If your prompts cap the report length and constrain the citation count, they stay bounded. Build cost-aware Deep Research prompts with our AI prompt generator — it caps output and citation count by default.

Sonar Deep Research is the most expensive Sonar tier on a per-call basis, and the temptation is to reach for it on every research-shaped query. Don't. The $5/1K search fee plus citation token meter is justified on a narrow set of workloads where the alternative is paying a human analyst.

**Justified**: long-form market scans (competitive landscape across 20+ companies), regulatory diligence (cite the source for every rule), academic-style literature reviews, autonomous research agents that need to produce a self-contained report, due-diligence workflows where every claim has to be traceable to a URL. On these, Deep Research's $0.10-0.20 per run replaces 30-90 minutes of human reading.

**Not justified**: 'what does X mean' lookups (use Sonar at $0.0095), simple cited Q&A (use Sonar Pro at $0.02), product-spec lookups where freshness matters but exhaustive citation does not (Sonar Reasoning Pro at $0.006). On these, Deep Research is paying 10x for capability you do not need.

The decision rule we use: if the answer has to survive being read by a skeptical reviewer who will check every citation, use Deep Research. If the answer is consumed in-line by a user who wants a quick grounded reply, use Sonar or Sonar Pro. Routing the right query to the right tier is worth more than any prompt-engineering optimization on either side.

The obvious alternative architecture is to stitch a general LLM (OpenAI, Anthropic) together with a standalone search API (Tavily, Brave, Serper). The math:

**GPT-5.5 + Tavily** for a 1,000-in / 500-out grounded answer: 0.001 × $5.00 + 0.0005 × $30.00 + Tavily search at ~$0.005 + retrieved-context tokens (typically 3,000-5,000 extra input tokens from the fetched passages) × $5/1M ≈ $0.005 + $0.015 + $0.005 + $0.020 = **$0.045 per call**.

**Sonar Pro** same shape: **$0.0205 per call** — roughly 55% cheaper.

Sonar wins on price for grounded-answer workloads because the pipeline is bundled. You don't pay LLM tokens to re-read the fetched search results — Perplexity does the extraction server-side and only bills you for the prompt and answer. The DIY stack pays for the same context twice (Tavily for the fetch, OpenAI for the read).

**When the stitched stack wins**: when you need precise control over which search index you use (e.g., domain-restricted retrieval, internal knowledge bases joined with web search), when you want to swap LLMs without rebuilding the search layer, or when your LLM-side work doesn't need search on every call. Sonar's per-request fee is unconditional — even calls that didn't need fresh data pay it. A stitched stack lets you call search only when you need it.

The hybrid that wins for most teams: Sonar Pro for queries that need fresh + cited answers, GPT-5.4-mini (calculator) for recall queries that don't, routed by a cheap classifier on the way in. That stack typically lands 40-60% below an all-Sonar bill at the same user-facing quality.

Perplexity runs two completely separate billing relationships. The **Sonar API** (priced per-token + per-request in the table above, accessed via docs.perplexity.ai and api.perplexity.ai) is for developers building applications. The **consumer Perplexity Pro subscription** ($20/month, accessed via perplexity.ai in a browser or the Perplexity app) is for end users chatting in the Perplexity UI.

What overlaps: the underlying models. Perplexity Pro consumers and Sonar API callers can both reach Sonar Pro, Sonar Reasoning Pro, and Sonar Deep Research under the hood. The user experience and the billing are entirely different.

What does NOT overlap: a $20/month Perplexity Pro subscription does **not** give you free unlimited API calls. It includes a $5/month API credit allowance — enough for roughly 250 plain Sonar calls or 100 Sonar Pro calls — as a sampler. If you're building anything past prototype, set up API billing independently and treat the Pro credit as a perk, not your budget.

Going the other direction: a paid Sonar API key does not include consumer Pro perks (unlimited Pro searches in the UI, Spaces, file uploads at consumer-Pro limits). Those are separate. If you want both — and most builders do, because the consumer UI is the fastest way to QA your own prompts — pay for both. The total is $20 consumer Pro + whatever your API bill is, and the two are billed to the same Perplexity account but tracked independently. The full breakdown of the consumer side lives in our Perplexity Pro cost guide.

**Mistake 1: defaulting every call to high search depth.** Sonar lets you pick low / medium / high search depth per call, and most workloads do not need high. Setting depth based on query shape (low for factual lookups, medium for general Q&A, high for synthesis) cuts the per-request bill 30-50% with no quality loss on shallow queries.

**Mistake 2: paying for search on queries that don't need it.** If 40% of your traffic is 'recall' (the LLM already knows the answer from pretraining), routing those queries to Sonar is paying $0.005-$0.012 per call for a search you didn't need. Add a cheap pre-classifier (gpt-5.4-nano is $0.000825 per classification call) that routes recall to a non-Sonar endpoint and reserves Sonar for queries that genuinely need fresh data.

**Mistake 3: not capping output on Sonar Pro or Deep Research.** Sonar Pro's output is $15/1M and Deep Research's citation+reasoning tokens bill at $8/1M. A 2,000-token answer that should have been 500 tokens costs 4x. Set max_tokens and citation count caps explicitly.

**Mistake 4: running Deep Research on Sonar-shaped queries.** Deep Research is roughly 10x the per-call cost of plain Sonar. Reserve it for genuinely long-form research where the citation chain is the deliverable.

**Mistake 5: ignoring the Perplexity Pro $5 API credit.** Even a $20/month Pro subscription you already pay for personal use throws in $5 of API credit — enough to run hundreds of prototype Sonar calls. Use it for dev/QA before you set up production billing.

Every price in this guide comes from Perplexity's live Sonar pricing page at docs.perplexity.ai/docs/pricing, fetched on 2026-06-20 and verified against the API console at api.perplexity.ai and against the citation/reasoning token billing notes in the Sonar developer changelog. When a number could not be verified against the official page (notably, granular pricing for legacy Sonar Online and Sonar Chat which were superseded in 2025), it was omitted.

Perplexity has revised Sonar pricing three times in the 12 months ending June 2026 — the June 2026 revision was the one that removed citation tokens from the bill for plain Sonar and Sonar Pro. Treat any pricing reference older than ~6 months with suspicion, including this one if you're reading it more than two quarters past the dateModified above.

**How to verify before you budget**: open docs.perplexity.ai/docs/pricing in an incognito window, copy the numbers for your target tiers into a spreadsheet, compare against this guide. If they match, this guide is current. If they don't, trust the live page. The per-request fee in particular is the most-revised number on the entire page — expect it to drift quarterly.

**Why the per-request fee is a range, not a single number**: Perplexity exposes search depth as a per-call parameter, and each depth maps to a different per-request charge. The ranges in our table ($5-12/1K for Sonar, $6-14/1K for Sonar Pro) span the published depth tiers — your bill will land somewhere in the range depending on your call mix. Most production teams run a mix and average to roughly the midpoint of the range.

**Reproducible methodology**: every row in the table has a citation on the live page; every worked example uses those rows; every FAQ answer reflects them. If you find a discrepancy with docs.perplexity.ai, treat the live page as canonical. The companion verified-prices document for this calculator series lives at /tmp/wave-a-prices-2026-06-20.md in our repo, with citation URLs for every figure.