xAI released Grok 3 with a reasoning mode and a Deep Search feature, trained on roughly ten times the compute of Grok 2.
Grok's research mode is tied to live platform data, which is a genuine edge for trigger-event prospecting — but vendor-run benchmarks are marketing, and should be treated that way.
xAI released Grok 3 on 17 February 2025. The release added a reasoning mode, branded Think, and a web research mode branded Deep Search. Elon Musk stated the model was trained with about ten times the compute used for Grok 2.
xAI said the model learns from its own mistakes using synthetic training data. The comparative benchmark claims made at launch came from xAI engineers during a live stream, not from independent evaluation, and the models named as beaten were GPT-4o, Gemini, DeepSeek V3 and Claude.
Two minutes, once a week. What changed in AI, and what to run because of it.