one year on
xAI releases Grok 3, claiming benchmark wins over GPT-4o
Elon Musk’s xAI unveils its latest flagship model, trained on roughly 200,000 GPUs, with reasoning modes and a new DeepSearch feature.
Elon Musk’s xAI late Monday released Grok 3, its latest flagship AI model, claiming it beats GPT-4o on benchmarks including AIME and GPQA. The model was trained at xAI’s Memphis data center using roughly 200,000 GPUs, representing about 10x the compute of Grok 2, according to Musk.
Grok 3 is a family of models including a faster mini variant and two reasoning models — Grok 3 Reasoning and Grok 3 mini Reasoning — that can fact-check themselves before answering. xAI claims Grok 3 Reasoning surpasses OpenAI’s o3-mini-high on several popular benchmarks, including the AIME 2025 math benchmark. A new Big Brain mode allocates extra computing for harder queries. The update also introduces DeepSearch, an agentic research tool that scans the web and X.
Access is tiered: Premium+ subscribers on X get first access, while a new SuperGrok plan at $30/month unlocks additional reasoning queries and unlimited image generation. Musk said a voice mode is coming within about a week, and the enterprise API will follow in weeks. He also reiterated xAI’s plan to open source the previous generation Grok 2 once Grok 3 is stable.
The launch follows months of delays; Grok 3 had been slated for 2024.
Musk claimed Grok 3 is 'a maximally truth-seeking AI, even if that truth is sometimes at odds with what is politically correct' and that it was developed with roughly 10 times more computing power than Grok 2.
view the original post →One year later — open only if you can handle spoilers
Grok 3 marked xAI's first credible claim to frontier capability, but independent benchmarks later showed gaps in reliability and safety alignment. The SuperGrok subscription tier failed to gain significant traction relative to competitors' offerings.