Perplexity Sonar Chat Completions Is Past End of Support: What Changed in the Request, and Sonar Pro's 13/20
By Promptster Team · 2026-11-14
Does your Perplexity Sonar integration still work? Probably, for now, but it is no longer running what it used to. Perplexity's migration guide says: "Sonar Chat Completions support ended on September 27, 2026." Synchronous and streaming chat-completions requests "are being reformulated as Agent API requests, rolling out gradually by model"; asynchronous Sonar requests "are not reformulated and are no longer supported" (Perplexity, migrate from Sonar). If you call /chat/completions with sonar-pro, your request is being turned into something else on Perplexity's side.
That something else is what we measured. On our November battery, sonar-pro passed 13/20, the second-lowest score of any runnable configuration, and five of its seven misses were a wrong number inside an otherwise well-formed answer. The battery's calls on 2026-10-23 went through the old chat-completions URL (the battery code predates our migration), and Perplexity now reformulates those server-side as an Agent API effort: fast request running openai/gpt-6-luna. So the 13/20 is the post-migration behaviour either way. It does not measure the old Sonar Pro model.
What changed in the request
Promptster moved to the Agent API in PR #69 (merged 2026-10-23, live in production). The request and response differ in five places:
| Sonar Chat Completions | Agent API | |
|---|---|---|
| Endpoint | POST /chat/completions |
POST https://api.perplexity.ai/v1/agent |
| Prompt | messages array |
input, plus instructions for the system prompt |
| Model | model: "sonar-pro" |
model for perplexity/sonar; otherwise an effort level |
| Output cap | max_tokens |
max_output_tokens (what Promptster sends) |
| Answer text | choices[0].message.content |
the output_text parts of message items in a typed output array |
The pplx- key is the same, and usage.cost.total_cost still reports what the call cost, including web search fees.
The model row matters most. The guide maps every Sonar id to an effort level, and the effort level chooses the underlying model and tools on Perplexity's side. Of the four, only sonar still exists as an Agent API model (perplexity/sonar), so Promptster keeps it on that model. Promptster keeps all four user-facing ids and maps them in PERPLEXITY_AGENT_MODEL_IDS (shared/perplexity-agent.ts):
| Promptster id | Agent API request | Model on 2026-10-23 (per shared/perplexity-agent.ts) |
|---|---|---|
sonar |
model: "perplexity/sonar" + web_search tool |
perplexity/sonar |
sonar-pro |
effort: "fast" |
openai/gpt-6-luna |
sonar-reasoning-pro |
effort: "low" |
openai/gpt-6-luna |
sonar-deep-research |
effort: "high" |
openai/gpt-6-sol |
Keeping the ids means saved tests, schedules, presets and API callers that name sonar-pro keep working without a data migration. The cost of that choice is in our decision log: three of those ids "now run whatever model Perplexity assigns to the effort level". Results will differ from historical Sonar runs, and the response's model field names the model that answered.
Sonar Pro on the battery: 13/20
Four script-graded tasks, five runs each, max_tokens 4,000:
| Task | Passed | Diagnostic | Median latency | Spend |
|---|---|---|---|---|
json-order: strict JSON with a per-item discount |
2/5 | 2/5 | 2.8 s | $0.0208 |
stock-floor: ten-step round-down arithmetic |
5/5 | 5/5 | 3.3 s | $0.0227 |
seven-word-taglines: exactly 7 words per line |
3/5 | 5/5 | 2.9 s | $0.0210 |
launch-notes: final budget from superseded notes |
3/5 | 3/5 | 2.8 s | $0.0206 |
| Total | 13/20 | $0.0851 |
Twenty-five of the 30 runnable configurations in the battery went 20/20. Only Groq's Qwen 3.8 27B scored lower than Sonar Pro (6/20, partly because of an on-demand tier cap of 1,000 output tokens per minute).
What the failures look like
Wrong totals in valid JSON. Every json-order response parsed, had po: null, the right SKUs and 4 units. Three had the wrong total (correct: 168, i.e. 3 × $20 plus $120 less 10%):
{"po":null,"skus":["SKU-A07","SKU-B12"],"units":4,"total":426}
The other two wrong totals were 414 and 294. A JSON schema validator passes all five. Only a value check catches them.
Right date, wrong budget. On launch-notes, two of five replies were Nov 21, $38500 and Nov 21, $38250, in exactly the requested format. The correct budget is $37,500: $40,000, minus $5,000, plus half the cut back. Neither wrong figure matches an obvious misreading such as counting the unapproved $1,200.
One word too many. The two failed tagline runs broke only the word count; every other rule passed (the diagnostic is 5/5). Run 1's third line was "sip boldly refill endlessly and enjoy every journey", eight words.
Clean arithmetic chain. Every stock-floor run walked the ten steps correctly and ended Answer: 587. The misses are on the tasks where the arithmetic is short and implicit, not on the one that asks the model to show its working.
For comparison, GPT-6 Luna called directly through OpenAI's API passed 20/20 on the same battery for $0.0019. We did not test what the fast effort level changes about the request (tools, search context, reasoning budget), so we are not drawing a conclusion from that gap. It does show that the effort-level route and a direct model call are not interchangeable.
What it cost
cost_usd is Perplexity's own usage.cost.total_cost per call. Every call in the run files carries request_cost: 0.0025, so the flat per-request fee was $0.05 of the $0.0851 total, more than all the tokens combined. Prompts of 68–145 tokens were also billed $0.0013–$0.0016 of input cost each.
Those figures come from the old URL's billing. The fallback rates in shared/pricing.ts changed with PR #69, for when Perplexity does not report a cost: sonar $0.25 / $2.50 per million plus $0.0025 per request, and sonar-pro $0.20 / $1.00 plus $0.001 per request (the comment reads "openai/gpt-6-luna at priority (2x $0.10/$0.50) + one Fast Search"). Our decision log also notes that effort levels run their own tools, so sonar-pro pays a search fee even on prompts that need no search, like all four of these.
What to do
- If you call Perplexity directly: move to
POST /v1/agentnow rather than waiting for the reformulation to reach your model. Renamemessagestoinput(system prompt toinstructions),max_tokenstomax_output_tokens, and read text from theoutputarray. Move async jobs first, since they are not reformulated at all. - If you use Promptster: nothing to change. The ids are the same, and the response's
modelfield shows which model answered. - Re-baseline anything you measured on Sonar before October. Our August Sonar review and the search-grounded field guide describe the chat-completions Sonar models. Their findings on grounding were about those models, not about what
sonar-promaps to today. - Don't send self-contained arithmetic or extraction to
sonar-pro. These prompts needed no web search, paid for one anyway, and got wrong numbers on 5 of 15 short-answer calls. Use it for questions whose answer is on the web, and check numbers in code.
Method
Four script-graded tasks (no LLM judge), five runs per task, max_tokens 4,000, temperature 0.7, run 2026-10-23 through Promptster's own provider handlers, at git 54b0093 (before PR #69, so the calls used the chat-completions URL). cost_usd uses Perplexity's reported usage.cost.total_cost; the billed estimate in summary.json is $0.0731. Latency is wall-clock from one machine, not a Perplexity SLA. Failures were recorded, not retried. For the wider field on this battery, see Benchmarking open-source vs. closed-source models.
Run this comparison yourself — no account, no API key
- Run the strict JSON order task, where Sonar Pro returned totals of 426, 414 and 294 (correct: 168)
To send the same prompt to sonar-pro and a direct model side by side, that's Promptster.
All figures from content/blog/data/q4b-2026-battery/runs/perplexity__sonar-pro/ (one raw file per call, Perplexity's usage object verbatim) and summary.json, run 2026-10-23. Request mapping from shared/perplexity-agent.ts and the Perplexity entry in docs/DECISIONS.md. End-of-support date and reformulation quoted from Perplexity's migration overview, read 2026-10-23.