Agent API: effort level updates and the ultra-preview effort level

From the source

These effort level changes were not listed in earlier changelog entries. Each line names the current version. Earlier versions stay available as pinned versions . fast ( fast@2026-10-16.0 ): falls back to openai/gpt-5.6-luna when openai/gpt-6-luna is unavailable. low ( low@2026-10-24.0 ): max output tokens rise from 32,768 to 128,000. web_search max_tokens and max_tokens_per_page rise from 2,000 to 4,000. fetch_url content extraction uses openai/gpt-6-luna . medium ( medium@2026-10-22.1 ): web_search max_tokens and max_tokens_per_page rise from 2,000 to 4,000. fetch_url content extraction uses openai/gpt-6-luna . high ( high@2026-10-28.0 ): uses openai/gpt-6.1-terra instead of openai/gpt-6-sol .

Max steps rise from 15 to 40, and a run makes at most 40 tool calls. web_search max_tokens and max_tokens_per_page rise from 2,000 to 4,000. fetch_url content extraction uses openai/gpt-6-luna . xhigh ( xhigh@2026-10-29.0 ): stays on anthropic/claude-opus-5-5 , with openai/gpt-6.1-sol as the fallback model.

Max steps rise from 100 to 300, and context compaction starts at 150,000 tokens. xhigh sets no system prompt, so its frozen configuration no longer sends instructions . …

docs.perplexity.ai