Perplexity API

What Perplexity actually meters

Perplexity meters tokens and requests separately, and the docs state the formula outright: total cost per query equals token costs plus a request fee. That fee applies to Sonar, Sonar Pro and Sonar Reasoning Pro, and not to Sonar Deep Research.

The fee tracks search_context_size, set inside web_search_options. Low is the default and the cheapest, high buys maximum search depth. Moving Sonar from low to high more than doubles the fee, from $5 to $12 per 1,000, without touching the token rate.

That split matters. Perplexity's published Sonar example uses 9 input and 411 output tokens at low context: tokens cost $0.000009 and $0.000411, the request fee costs $0.005. So 92% of that line is search, not the model. Budget from token counts alone and you'll be out by an order of magnitude.

Sonar Deep Research inverts the shape, dropping the request fee and adding three meters instead. Perplexity's worked example for one call totals $0.816123, of which $0.581841 is reasoning tokens. Every response carries a usage.cost object splitting request_cost from token costs, so you reconcile per call.

How credits work

Perplexity runs prepaid credits, not postpaid invoicing. You buy credits in the API console, Stripe takes the payment, and calls draw the balance down. Auto reload adds credits when the balance falls below a threshold you set, and the docs recommend enabling it.

Cumulative purchases also drive your usage tier, which sets your rate limits. Tier 0 starts at $0 and Tier 5 at $5,000 of lifetime purchases, with $50, $250, $500 and $1,000 in between. Tiers count lifetime spend rather than current balance and never downgrade: on the Agent API that's 1 query per second at Tier 0 against 33 at Tier 5, and on Sonar 50 requests a minute against 4,000.

Credit expiry, rollover, pooling across projects and deduction order are all not documented, and neither is any refund path for unused balance. Enterprises can buy credits through AWS Marketplace.

What happens when you hit the limit

You get blocked, not billed. Perplexity's docs are blunt about it: run out of credits and your API keys are blocked until you top the balance up. Calls then return a 401, which the FAQ lists alongside invalid and deleted keys as a credential failure rather than a billing one. There's no grace allowance and no arrears billing, which makes auto reload worth checking before any launch.

Rate limiting behaves differently. Exceed your tier's QPS or per-minute limit and you get a 429 with a Retry-After header, and Router requests rejected with a 429 aren't billed.