Key points
- The Agent API now includes an ultra-preview effort level for broad research tasks.
- Developers can require approval for MCP and connector tool calls before execution.
- Older Anthropic and Google models are no longer listed in the Agent API docs.
Perplexity has updated its Agent API documentation to introduce an ultra-preview effort level, add approval mechanisms for tool calls and remove several older models from its published lists.confirmed These changes appear on the company’s API changelog page.Perplexity is adding new controls for developers to manage how AI tools run and how much work they do.The Agent API now exposes require_approval for MCP calls and introduces ultra-preview effort with relaxed constraints.
The facts
Web search and fetch_url operations see max_tokens rise to 4,000 from 2,000 across low, medium and high efforts.reported The high effort level allows up to 40 steps and 40 tool calls. The xhigh effort level permits 300 steps, with context compaction starting at 150,000 tokens.
The facts
Perplexity states that step budget reminders are now active for low, medium, high and xhigh effort levels as the model works. The fast effort level runs a single step and does not receive these reminders. The ultra-preview level is described as suitable for tasks where coverage matters most, though it can take a long time to run.
What changes, what does not
The parameter previously known as preset is now called effort. Perplexity notes that preset remains accepted as a deprecated alias and responses return both fields. The undocumented ultra effort name is removed and now returns an error. The wide-research alias now selects xhigh. The max_steps parameter accepts values up to 500, up from 100.
What builders can do with it
Developers can integrate approval workflows for MCP and connector tools by setting require_approval. Calls needing approval return an mcp_approval_request item and do not run until an mcp_approval_response is sent. Users can filter approvals to skip read-only tools or review every call. Integrations expecting fixed breakdowns in credit usage data must update to read the returned categories list, as the Credit Source breakdown is currently omitted.
Why it matters
- Applications using MCP tools can now enforce human-in-the-loop controls before external actions execute.
- Builders relying on the deprecated preset parameter will see it accepted but should migrate to effort for clarity.
- Systems parsing credit usage analytics must handle missing Credit Source breakdowns without assuming zero usage.
- Developers using frozen configurations must copy updated values to match public API fields for effort levels.
Topics: API, Perplexity, MCP, developer tools, documentation
Updated
Then / now / next
The API used preset for effort levels and lacked granular MCP approval controls.
The API uses effort, adds ultra-preview and allows developers to review tool calls before execution.
Developers should monitor for further documentation updates regarding model availability and pricing.
Story timeline
- Then The API used preset for effort levels and lacked granular MCP approval controls.
- 02 Nov 2026 Perplexity API Changelog page updated:
- Now The API uses effort, adds ultra-preview and allows developers to review tool calls before execution.
- Next Developers should monitor for further documentation updates regarding model availability and pricing.
The strongest doubt
Counterpoint from the sources
The source does not state whether the removed models are still technically callable via the API or only hidden from documentation and pricing calculators.
For readers
Developers
- Update integrations to use effort instead of preset and handle new approval workflows.
- Adjust analytics code to read returned categories lists for credit usage breakdowns.
Business
- Implement approval gates for MCP tools to maintain control over automated actions.
- Review cost implications of increased token limits and step budgets in new effort levels.
Policy
- Monitor how approval mechanisms for external tool calls affect data privacy and audit trails.
- Assess the impact of removing older models from documentation on long-term system stability.
Try it
Set require_approval to always for an MCP tool call.
Good result: The call returns an mcp_approval_request item instead of executing immediately.
Ask the article
Local keyword search of this article; no generated answers.Which of these is confirmed?
Check the answer
Perplexity has updated its Agent API documentation to introduce an ultra-preview effort level, add approval mechanisms for tool calls and remove several older models from its published lists.
The matched primary source and announcement wording support the confirmed label.
Did this article give you the evidence you needed?
Your choice stays on this device. No public tally.Questions readers ask
Does the preset parameter still work in the API?
Yes, preset is still accepted as a deprecated alias and responses return both fields.
Which models are no longer listed in the documentation?
Several older Anthropic and Google models, including claude-opus-4-8 and gemini-3-flash-preview, are removed from docs.
How do I approve an MCP tool call?
Send an mcp_approval_response with the previous_response_id to approve or deny the pending request.
Connected stories
Reading receipt
- The Agent API now includes an ultra-preview effort level for broad research tasks.
- Developers can require approval for MCP and connector tool calls before execution.
- Older Anthropic and Google models are no longer listed in the Agent API docs.