The Perplexity Sonar API Retirement: What Changes

Perplexity is shutting down the Sonar chat completions API on September 27, 2026. Every Perplexity model on LLM Gateway runs on that endpoint today, so the Perplexity Sonar API retirement affects all three of them — but not in the same way, and not on the same date.

The short version: perplexity/sonar keeps its model id and keeps working, on a new transport, from September 25. perplexity/sonar-pro and perplexity/sonar-reasoning-pro have nowhere to go and stop being routable on September 27.

What Replaces the Sonar API

Perplexity's successor is the Agent API, a single POST /v1/agent endpoint where search is a tool the model can call rather than something baked into the model. That is a real behavioral difference: on the old Sonar endpoint every request searched, while on the Agent API the model decides.

Preserving Sonar's behavior therefore takes more than swapping a URL. For perplexity/sonar the gateway attaches a web_search tool and requires it, so every request still searches the web the way it always did. You do not send the tool yourself and you do not pay for a search you did not ask for — you get the grounded model you were already calling.

Nothing Changes in Your Requests

Keep calling the same endpoint with the same model id:

1curl https://api.llmgateway.io/v1/chat/completions \2  -H "Authorization: Bearer $LLM_GATEWAY_API_KEY" \3  -H "Content-Type: application/json" \4  -d '{5    "model": "perplexity/sonar",6    "messages": [{"role": "user", "content": "What shipped in space news this week?"}]7  }'

1curl https://api.llmgateway.io/v1/chat/completions \2  -H "Authorization: Bearer $LLM_GATEWAY_API_KEY" \3  -H "Content-Type: application/json" \4  -d '{5    "model": "perplexity/sonar",6    "messages": [{"role": "user", "content": "What shipped in space news this week?"}]7  }'

Everything that came back before still comes back:

• The answer, non-streaming and streaming alike.
• JSON schema output, unchanged.
• The top-level search_results array, with one entry per source and each source's url, title, snippet, date and last_updated. On streaming responses it arrives on the chunk that delivers the sources.
• citations, for clients that read Perplexity's original field.

One thing is new. Sources now also arrive as standard annotations on the assistant message — the same url_citation shape every other search-grounded model on the gateway emits:

1{2  "annotations": [3    {4      "type": "url_citation",5      "url": "https://example.com/article",6      "title": "Article Title",7      "date": "2026-09-16",8      "last_updated": "2026-09-17"9    }10  ]11}

1{2  "annotations": [3    {4      "type": "url_citation",5      "url": "https://example.com/article",6      "title": "Article Title",7      "date": "2026-09-16",8      "last_updated": "2026-09-17"9    }10  ]11}

Perplexity never populated that field on the old endpoint, so if you were special-casing Sonar in your citation handling, you can stop. The web search docs cover the normalized citation shape across providers.

It Gets Cheaper

Sonar's old pricing charged a flat fee on every single request on top of tokens, whether or not the answer needed much searching. The Agent API drops that fee and meters searches individually, billed off the search count Perplexity reports back. Output tokens are priced higher than before, input tokens considerably lower, and cached input is now priced separately.

Netted out, a grounded question measured about a third cheaper than on the old path in our own testing. Your mileage depends on how output-heavy your prompts are; current rates for the model are always on the models page. As before, the search charge is reported separately as cost_details.web_search_cost so you can see exactly what grounding cost you.

Why sonar-pro and sonar-reasoning-pro Are Not Being Migrated

Perplexity's Agent API rejects both model ids outright — they are not models there, and there is no drop-in successor. What the Agent API offers instead are presets that trade off search depth and cost.

We could have pointed the two retiring ids at a preset and called it a migration. We decided not to, for one reason: those presets run a different underlying model, with different search behavior, different citation style and different pricing. Silently swapping that in under an unchanged model id means your evaluation results move, your costs move, and nothing in your code says why. A model id is a contract. When we cannot honor it, an error you can see beats a substitution you cannot.

So from September 27, requests to perplexity/sonar-pro and perplexity/sonar-reasoning-pro return an error instead of a quietly different answer.

What to Do If You Use Them

Check first. Filter by model in Model Usage in your dashboard to see whether either id shows up in your recent traffic. Most projects using Perplexity for grounded answers are on perplexity/sonar and have nothing to do.

If you are affected, you have two paths:

1. Move to perplexity/sonar. A one-line change to the model string. You keep Perplexity's search index and citation behavior, and you get the new pricing. This is the closest thing to a like-for-like move.
2. Move to another search-grounded model. Several providers offer native web search through the same web_search tool, so the request shape you already use carries over. The models page filtered to web search lists them with current pricing. If you were on sonar-reasoning-pro specifically for the reasoning, pair a reasoning model with the web_search tool — see the reasoning docs.

Either way, routing means you can test a replacement against real traffic before you commit: send a slice of requests to the candidate model and compare the answers and the costs in your dashboard.

Deprecations Are Part of the Job

Providers retire endpoints. That is the reality of building on models you do not host, and it is a large part of what a gateway is for: absorbing the churn so a transport change upstream does not become a code change downstream. In this case one model id survives untouched because the gateway rewrote the request for you, and two do not because honesty was worth more than a silent fallback.

If you need help moving a workload off the retiring models, reply to the announcement email or reach us at contact@llmgateway.io — we are happy to look at your traces with you.

• Try LLM Gateway free
• Web search documentation
• How routing picks a provider