NeuralStackxAI released Grok-3 on July 14 2026 with a 500000 token context window and built-in real-time web search. The model achieves 89 percent on the GPQA benchmark and introduces native agentic tool calling for code execution and API orchestration. Early API access opened the same day for enterprise customers at 20 dollars per million tokens.
Training completed on a 120000 H100 cluster over four months using a new mixture-of-experts architecture with 1.8 trillion total parameters. xAI stated the model reduces hallucination rates by 47 percent compared with Grok-2 on internal factuality tests.
Background context includes xAI's prior focus on maximum truth-seeking models and its ongoing buildout of a Memphis supercluster. The company previously raised 6 billion dollars in May 2026 to expand compute capacity.
Competitors including OpenAI and Anthropic have signaled similar agentic releases before year end. Grok-3's web access differentiates it from closed training cutoffs used by most rivals.
Why this matters
The launch accelerates the shift toward always-on agentic systems that can act across tools without human prompts. Enterprises gain a new option for research and coding workflows that require current information.
Longer term the release pressures other labs to match both context length and live data capabilities while raising questions about data provenance and attribution in real-time retrieval.
xAI plans to open source a 70 billion parameter distilled version by October 2026 to broaden developer adoption.