Grok Code Fast 1

Launched August 28, 2025, Grok Code Fast 1 is a from-scratch reasoning model built specifically for agentic coding workflows. xAI's design goal was to eliminate the latency tax that reasoning loops and tool-calling impose on coding agents: it sustains roughly 92 tokens/second while supporting native tool use (grep, terminal, file editing) across TypeScript, Python, Java, Rust, C++, and Go. It was pre-trained on a programming-rich corpus, post-trained on real-world pull requests, and refined with launch partners under the codename 'sonic'. Aggressive pricing — $0.20/1M input and $1.50/1M output — made it one of the cheapest capable coding models of 2025.\n\nLimitations: 10K max completion tokens caps very long generations; it is a speed-optimized model, not a max-capability reasoner — hard architectural problems still favor Grok 4 or o3-class models. Closed weights, xAI API only.

Community-reported: ~80% LiveCodeBench, ~70.8% SWE-bench