Introducing deep research

• OpenAI launches deep research: agentic multi-step web research in 5–30 min.
  • Replaces hours of human analyst work with a single query.
  • Powered by optimized o3; reads PDFs, images, uploaded files.
• Humanity’s Last Exam: 26.6% vs o1’s 9.1% and Claude 3.5 Sonnet’s 4.3%.
• GAIA benchmark: 67.36% pass@1 — current state of the art.
• Target market: finance, science, policy, engineering pros + big-purchase consumers.
• Pricing: Pro 250/mo, Plus/Team 25/mo, Free 5/mo; o4-mini fallback after limit.
  • Lightweight o4-mini version for cost efficiency post-quota.
• Acknowledged: still hallucinates, weak confidence calibration, conflates rumors with sources.

Isa Fulford and Zhiqing Sun (OpenAI) · ** · Read on openai.com