NeuralStackxAI released Grok 3 on July 19 2026 with an 8 million token context window and integrated real-time web search plus code execution. The model achieves 94.2 percent on MMLU 89.7 percent on GPQA and 78.4 percent on SWE-Bench. API access opened immediately for enterprise customers at 15 dollars per million input tokens.
Training completed on a 120 thousand H100 cluster over 92 days using a new mixture-of-experts architecture with 1.8 trillion total parameters. xAI published the full system card and red-team results on the same day. Early benchmarks show 40 percent fewer hallucinations than Grok 2 on factual retrieval tasks.
Background context includes xAI raising 6 billion dollars in March 2026 to build Memphis supercluster phase two. The company previously open-sourced Grok 1 weights in 2024. Grok 3 training data cutoff is set at May 15 2026.
Elon Musk stated the model will power future X platform features and Tesla robotaxi planning. xAI also released a distilled 70 billion parameter version for on-device use.
Why this matters
The release intensifies competition in the frontier model race and pressures OpenAI and Anthropic on pricing and context length. Developers gain immediate access to long-context reasoning without stitching multiple calls.
Enterprises can now run multi-hour document analysis and agent workflows in a single session. Regulators are watching xAI closely after the company skipped the voluntary safety commitments signed by other labs in 2025.
Analysts expect Grok 3 to capture 12 percent of the inference market by year end. xAI plans a Grok 3.5 update with video input before December 2026.