NeuralStackxAI launched Grok-3 on July 20 2026 with an 8 million token context window and built-in real-time web browsing capabilities. The model achieved 92.4 percent on MMLU and 89.1 percent on GPQA Diamond while running at 1.2 times the inference speed of its predecessor. xAI also introduced a new mixture-of-experts architecture with 1.8 trillion total parameters.
Enterprise customers can access Grok-3 through the xAI API starting July 27 2026 at $15 per million input tokens and $60 per million output tokens. Early benchmarks show a 35 percent improvement in multi-step reasoning over Grok-2 on the DROP dataset. The release includes native support for code execution and image analysis within the same context window.
Background context includes xAI's prior focus on Grok-2 which launched in December 2025 with a 2 million token limit. The company raised 6 billion dollars in Series C funding in March 2026 to scale training infrastructure. Grok-3 training completed on a cluster of 100,000 H100 GPUs over 92 days.
Competitors including OpenAI and Anthropic have announced similar long-context updates scheduled for Q4 2026. xAI's integration of live web data differentiates the model for research and news summarization use cases. The company plans to open source a 70 billion parameter version by September 2026.
Why this matters
Grok-3 raises the bar for context length and real-time capabilities which could accelerate adoption in legal document analysis and scientific research. The pricing and performance metrics pressure other labs to match both scale and cost efficiency. Industry analysts expect this release to influence enterprise procurement decisions through the end of 2026.
Longer context windows reduce the need for retrieval-augmented generation pipelines in many workflows. xAI's rapid iteration cycle signals increased competition in the frontier model space. Developers should evaluate Grok-3 for applications requiring extended conversation history or live data feeds.
Future updates will likely include multimodal video understanding and expanded agent capabilities by early 2027. Organizations evaluating frontier models should benchmark Grok-3 against Claude 4 and GPT-5 equivalents on their specific datasets.