Details
- Perplexity announces GPT 5.6 Terra and Luna as new models integrated into Perplexity Computer.
- Terra becomes the default model for all Computer subagents, indicating its role in complex, goal-oriented task execution.
- Luna is designated as the primary model for scheduled automations, targeting recurring workflows where speed and cost efficiency are critical.
- Terra is also available as an orchestrator model in Computer, meaning it can coordinate and manage multiple subagents across a workflow.
- Perplexity states that Terra is built for complex, goal-oriented work, making it especially suitable as a subagent handling multi-step research or operational tasks.
- On the WANDR benchmark, an in-house wide-and-deep research agent evaluation, Terra reportedly scores 11 points above Anthropic Claude 3.5 Sonnet, suggesting a notable performance gain in professional research-style workloads.
- WANDR is designed to test both wide retrieval across many sources and deep extraction of granular facts, so Terra’s uplift on this benchmark implies stronger search-plus-reasoning capabilities.
- The company also highlights that Terra delivers order-of-magnitude cost reductions versus Sonnet on WANDR-style tasks, positioning it as a more affordable option for intensive research and agentic workflows.
- Luna is positioned to handle recurring workflows, such as scheduled automations in Perplexity Computer, where lower latency and cost per run are more important than maximum benchmark scores.
- Together, Terra and Luna expand Perplexity Computer’s agent model lineup, separating heavy, complex subagent work (Terra) from fast, routine automation (Luna).
Impact
By splitting roles between Terra and Luna, Perplexity is formalizing a two-tier agent strategy: Terra for high-complexity, research-grade workloads and Luna for high-frequency automations. Terra’s reported WANDR advantage over Sonnet, combined with significant cost reductions, could pressure rival frontier models to optimize not just raw accuracy but cost-performance in agentic research settings. This move also deepens Perplexity Computer’s positioning as an orchestration layer where specialized models are allocated to different parts of a workflow, narrowing the gap with ecosystems built around Claude, GPT-4o, and similar multi-model agent stacks.
