By ByteBot
Perplexity’s Sonar API died on September 27. Not deprecated with a grace period — dead. Requests to sonar-pro and sonar-reasoning-pro are now returning errors. Twenty-four hours later, Perplexity shipped Profiles, Skills, and Managed Connectors for the Agent API. That timing is not a coincidence. The message is clear: the old search API is gone, and what replaced it is something significantly more ambitious.
What Broke and Who Is Affected
If your app called sonar-pro or sonar-reasoning-pro directly, those endpoints are now failing. There is no soft migration path — the models are gone. Base sonar fared better: proxy gateways like LLM Gateway silently rerouted it to the Agent API on September 25, keeping responses consistent for most users.
The deeper issue is architectural. The Agent API is not a rename of Sonar. The entire call shape changed:
# Before — Sonar (now dead for sonar-pro)
completion = client.chat.completions.create(
model="sonar-pro",
messages=[{"role": "user", "content": "Latest AI agent frameworks?"}]
)
print(completion.choices[0].message.content)
# After — Agent API
response = client.responses.create(
preset="fast",
input="Latest AI agent frameworks?"
)
print(response.output_text)The messages array is gone. The choices[0].message.content accessor is gone. If your integration calls Perplexity directly rather than through a gateway, you are rewriting, not patching.
The Fastest Migration Path: Presets
For teams that just need search-grounded answers, presets are the lowest-friction path. Each preset bundles a model, system prompt, tool configuration, reasoning effort, and token budget under a single name. You pick the preset, pass your query, get an answer.
| Old Model | Preset | Use Case |
|---|---|---|
| sonar | fast | Quick lookups |
| sonar-pro | fast | Research questions |
| sonar-reasoning-pro | low | Multi-step reasoning |
| sonar-deep-research | high or xhigh | Expert analysis |
One immediate upgrade: unlike Sonar, the Agent API lets you specify any model from OpenAI, Anthropic, Google, or xAI at direct provider rates. You are no longer locked to Perplexity-branded models.
Profiles: Configure Once, Deploy Everywhere
Presets are Perplexity-managed. Profiles are yours. A Profile saves your agent’s complete configuration — model selection, system instructions, tools, Skills, managed connectors, and step budget — under a single versioned ID. Create it once in the API Portal; reference it by ID in every API call.
response = client.responses.create(
profile={
"type": "custom",
"id": "profile_8Qw3x7tJm2N6pR4",
},
input="Why did paid conversions drop last week?"
)Profiles are immutable. Every edit creates a new version. You can pin to a specific version with "version": "3" or always take the latest with "version": "latest". This matters for production: a team can update an agent’s behavior without touching application code, and a rollback is a one-field change.
Managed Connectors: Credentials Without the Plumbing
The part of this launch that deserves more attention than it is getting: Managed Connectors. Instead of passing API tokens on every request or running your own tool-execution server, you store credentials in the API Portal once and attach the connector to a Profile. The agent accesses live systems — GitHub, Slack, Google Drive, Datadog, Vercel, Stripe, Supabase, Microsoft 365 — without any credential passing in application code.
The practical result is striking. An incident responder Profile might bundle a Datadog connector, a GitHub connector, and a Slack connector along with an Incident Response Skill. When something breaks at 2 AM, your system calls the API with an incident ID and a repository name. The agent pulls telemetry from Datadog, clones the repo, runs git diff against the deployment tag, and posts a diagnostic summary to Slack. Your application code contains none of that logic — just the profile ID and the incident context.
This is the design choice that separates Perplexity’s approach from OpenAI’s Agents API, where you own the execution layer. Here, Perplexity hosts it.
The Bigger Picture
Perplexity spent two years being described as a search engine that talks back. This launch suggests a different roadmap. Sonar was a search API bolted onto a consumer product. The Agent API with Profiles, Skills, and Connectors is a managed agent runtime — infrastructure for teams building systems that need persistent configuration, live data connections, and composable capabilities.
That puts Perplexity in a crowded lane: OpenAI’s Agents API, Anthropic’s MCP ecosystem, LangGraph. But Perplexity’s angle — managed connectors with server-side credential storage, plus a transparent per-tool pricing model ($0.0025 per web search, $0.03 per sandbox session) — is a genuine differentiator from the “bring your own tools” approach most competitors require.
What to Do Now
- Check your integration immediately. If you call Perplexity directly and used
sonar-proorsonar-reasoning-pro, you have live failures. Switch topreset="fast"as a temporary fix. - Read the official migration guide. The Perplexity Sonar migration docs map every old model to a preset equivalent and show the full code translation.
- Evaluate Profiles if you have any complexity. If your current integration sets a system prompt or uses specific tools, moving that config into a versioned Profile now will save significant headaches as the platform evolves.
