Perplexity's Sonar API is still answering on its retirement day, but the replacement it points to changed twice this month, so the migration should pin a model and a search setting, not a preset.

A single call to `https://api.perplexity.ai/chat/completions` with model `sonar`, using the engine's own key, returned HTTP 200, a normal answer, a citations list, and `"model":"sonar"` in the response. Perplexity's migration page still reads, verbatim, "Sonar will be supported until September 27, 2026," and still says nothing about what requests get after that. The engine's local `main` is still at `20b1172` (22 August), with `perplexity: "sonar"` on line 14 of `src/adapters.ts` and the chat completions URL on line 91; no local branch mentions perplexity, sonar or agent. The remote was not checked this run.

Two September entries matter. First, verbatim: "The Agent API `fast` preset now uses Fast Search (`search_type: "fast"`). `web_search` drops from $2.50 to $1.00 per 1,000 invocations, and search is about 800 ms faster." Second, the `low` and `medium` presets now use `openai/gpt-6-luna` and `high` uses `openai/gpt-6-sol`. Perplexity's own migration guide maps `sonar` to `fast`, `sonar-pro` to `low`.

The incident still carries one post, the 24 September launch note ("The rollout may take up to two weeks to complete"), and no completion message. Separately, Search Engine Journal (Roger Montti, 25 September) reported a short Google paper describing SAFE, an AI system for investigating "AI slop" and coordinated abuse, which the paper says is in "early deployment." Montti links it to the spam update; that link is his reading, not something Google has said.

BrightEdge says ChatGPT's share of AI referral traffic hit 95.1% in August 2026, with Gemini 2.4%, Perplexity 1.1% and Claude 0.8%, and that ChatGPT referrals rose 101% from January to August. The release gives no sample, no site count and no method beyond "BrightEdge data."