Details
- Perplexity evaluated OpenAI’s GPT-6 Astra on its WANDR wide-and-deep research benchmark.
- GPT-6 Astra achieved a soft F1 score of 0.682 on WANDR, the highest score among all models Perplexity has tested so far.
- The model’s average cost in this setup was $11.98 per task, indicating relatively high-effort runs on the benchmark.
- GPT-6 Astra scored 13.5% higher than Fable 5.1 while being 6.1% cheaper per task, showing a quality and cost advantage.
- It also scored 27.0% higher than Opus 5, at a 3.3% increase in cost per task, trading slightly higher spend for a large quality gain.
- These comparisons suggest that GPT-6 Astra delivers better WANDR performance than key rival models at comparable or lower cost levels on this benchmark.
- The announcement builds on Perplexity’s prior work on WANDR as an open benchmark for research agents, using it to profile new frontier models.
- WANDR focuses on wide discovery and deep, evidence-backed research tasks, making the results relevant for high-stakes research and analysis workflows.
Impact
By publishing GPT-6 Astra’s WANDR results alongside cost data, Perplexity signals that the latest OpenAI frontier model is highly competitive for complex research-agent workloads while remaining cost-manageable. The strong relative gains over Fable 5.1 and Opus 5 are likely to push rivals to optimize both quality and pricing for agent-style tasks and may accelerate enterprise adoption of Astra-class models in research-heavy domains.
