An automation built against the wrong Perplexity API goes quiet, not down
An engineer wires Perplexity's Sonar model into a research agent in August 2026, tests it against a handful of real queries, ships it, and moves on to the next feature. Nobody touches that integration again for two months. Sometime after that, part of its asynchronous workload simply stops coming back, with no error, no alert, and no obvious reason in the logs.
Perplexity's own migration documentation explains what happened: "Sonar Chat Completions support ended on September 27, 2026." Synchronous and streaming calls to the old model names are being automatically rerouted to the replacement, but asynchronous requests "are not reformulated and are no longer supported." An automation that only ever called Sonar asynchronously did not get a deprecation warning. It got silence.
Most pricing guides for "perplexity api pricing" that still rank were written against that retired API, one table of flat per-model token rates. What replaced it is not one API with new prices. It is five separately priced products, and most of the cost surprises in a Perplexity-backed build now come from mixing those products up, not from the token rates themselves.
The five products, and what each one actually bills on
Perplexity's pricing page lists the current lineup, each billed on a different unit:
- Orchestrator API - per token, at each model's rate, with no per-request fees at all.
- Agent API - per token, at the model's rate, plus a separate fee for every tool call the model makes.
- Search API - per successful request, with no token charge, a flat fee regardless of how long the response is.
- Embeddings API - per token, priced per embedding model.
- Decisions API - per input token only, output tokens are free, plus a small fee if web grounding is turned on.
The Agent API is the one most automation builds actually use, since it is the only product in the list that can both reason over a prompt and call tools like web search in the same request. It is also the one where the bill has the most moving parts.
What the Agent API actually costs, worked out
Perplexity's model pricing reference lists perplexity/sonar, the direct successor to the old default Sonar model, at $0.25 per million input tokens and $2.50 per million output tokens. On top of that, every tool call the model makes inside a request is billed separately. Perplexity's pricing page lists the Agent API's web search tool at $0.0025 per standard call or $0.001 per Fast Search call, with other tools priced individually (fetch URL at $0.0005 per call, a sandbox session at $0.03 plus $0.0025 per search inside it).
As an illustration, not a quote from any single source: a research automation that runs 5,000 grounded queries a month, each averaging 400 input tokens and 600 output tokens, with one standard web search call per query, would spend roughly $0.50 on input tokens (2 million tokens at $0.25/million), $7.50 on output tokens (3 million tokens at $2.50/million), and $12.50 on search calls (5,000 calls at $0.0025 each), for about $20.50 a month total. Change the token counts or the search tier and the number moves, but the shape, token cost plus a per-call tool fee, stays the same.
The standalone Search API is not a cheaper way to do the same search
It is tempting to assume the separate Search API, since it skips the model entirely, is the cheap option for anything that just needs search results. The numbers on the same pricing page say otherwise for the standard tier.
The Agent API's bundled web search tool costs $0.0025 per call, which works out to $2.50 per 1,000 calls. The standalone Search API's standard tier costs $5.00 per 1,000 requests, exactly twice as much per request for what is nominally the same search product. The Fast Search tier is the one case where the two line up: $0.001 per call in the Agent API and $1.00 per 1,000 requests in the standalone Search API work out to the same $1.00 per 1,000.
So the standalone Search API is not a cheaper substitute for the Agent API's search tool. It is a different product for a different job, raw search results with no model reasoning attached, and for the standard tier specifically, it costs more per call than getting the same search bundled into a model request that also gives you an answer.
Embeddings and Decisions: the two products most guides skip entirely
Because most competitor pricing roundups only ever covered Sonar, and now only cover its replacement Agent API, two other priced products rarely get mentioned at all.
The Embeddings API prices four models by token: pplx-embed-v1-0.6b at $0.004 per million tokens, pplx-embed-v1-4b at $0.03 per million, and two context-aware variants at $0.008 and $0.05 per million respectively, all per the same pricing documentation.
The Decisions API is priced differently again: pplx-decider-v1.3 at $0.02 per million input tokens and a faster pplx-decider-flash-v1.3 at $0.01 per million, with output tokens free and a flat $0.001 added only if web grounding is enabled on that call. If a project needs classification or routing decisions rather than open-ended generation, this line item is cheaper than running the same decision through a general-purpose model on the Agent API, and it is worth checking before defaulting to whichever product the rest of the build already uses.
Neither of these two products shows up in the Sonar-era pricing tables still circulating online, for a simple reason: they did not exist under the old API at all. A guide written against the retired Sonar lineup has no row to update for either one, it is missing them entirely, which is a different and easier mistake to miss than a stale number sitting in an otherwise-current table.
What this means if Perplexity is one piece of a larger automation
Perplexity is rarely the only tool in a build. It usually sits inside an AI automation that also calls another model for drafting, writes results into a CRM, or triggers a follow-up action once a research step finishes. When that is the case, the Perplexity line item is one of several costs to size, not the whole bill, and mixing up the Agent API's bundled search fee with the standalone Search API's per-request fee is an easy way to double a cost estimate before a project even ships.
If the automation itself is the orchestration layer, calling Perplexity for grounded search and then handing the result to another step, that routing and error handling is the part our workflow automation team builds. A pricing change or a model swap on Perplexity's side then stays contained to one step instead of forcing a rebuild of the whole pipeline.
A quote priced off a cached table and a quote priced off the live docs will not match
Two proposals for the same Perplexity-backed build can land on different numbers without either one containing a math error, if one of them was priced against a blog's rate table and the other against the current pricing page. The gap isn't scope. It's which page the estimate came from, and only one of those pages still reflects a product that accepts requests.
That is the check worth doing before any number goes into a proposal: open the live pricing page, confirm whether the build needs the Agent API's bundled search tool or the standalone Search API's flat per-request fee, and price the token rate and the tool fee together rather than picking one. Our AI automation team runs that check against current docs on every build that touches Perplexity, and our workflow automation team handles the orchestration layer when Perplexity is one step inside a larger pipeline rather than the whole project.



