Perplexity swapped its orchestrator model and the cost-performance chart is brutal

Perplexity announced the update on X on July 10, slotting Grok 4.5 as the orchestrator inside Computer, the planning and coordination brain for its agentic browsing and task tool, available on Pro and Max plans. The move comes as labs race to find combinations, not just single models, that balance accuracy with cost, as Sonnet 4.6's pricing reset shows.

The company tested five different orchestrator setups on WANDR, its internal benchmark, and published a cost-performance chart. Grok 4.5 topped the curve with a score of roughly 0.328 at $4.76 per trial. Claude Opus 4.8, in a high-thinking configuration, came in at 0.254 at $9.46 per trial, nearly twice the price for worse accuracy, a gap that echoes the cost-efficiency arguments emerging across agent benchmarks, as LeanStral 1.5's theorem about AI pricing and formal verification suggests.

The rest of the field filled the middle: GLM 5.2 with an advisor model hit 0.297 at $4.67, GPT-5.6 Sol in medium-effort mode scored 0.289 at $2.64, plain GLM 5.2 scored 0.207 at $1.74, and GPT-5.6 Terra, the cheapest option tested, scored 0.149 at just $0.40 per trial. The spread shows that cheaper models can still deliver for simpler tasks, but the orchestrator role demands a different kind of reliability, a pattern that holds for coding agents too, per best practices for AI coding agents.

Perplexity did not publish the full WANDR methodology alongside the announcement, so the benchmark's scoring criteria and task set remain unverified outside the company's own testing. The lack of third-party validation is a recurring tension in agentic benchmarks, as research on sandbox benchmarks hiding real agent failures highlights.