According to Zhidongxi on August 29, last night, Musk's AI company xAI launched the programming-savvy reasoning model Grok Code Fast 1 , with the API call name grok-code-fast-1.
Musk announces programming model
Grok Code Fast 1 supports a 256KB context window length, focusing on fast response and affordability . The model invokes dozens of tools before the user even finishes reading the first paragraph of a thought trajectory. xAI also invests in fast cache optimization, with cache hit rates typically exceeding 90% .
The pricing for this model is: $0.2 (approximately RMB 1.43) per million input tokens, $1.5 (approximately RMB 10.70) per million output tokens, and $0.02 (approximately RMB 0.14) per million cached input tokens.
As shown in the official chart released by xAI, Grok Code Fast 1's performance is comparable to models like GPT-5 and Gemini 2.5 Pro , but it's priced several times lower . It's slightly more expensive than Qwen3-Coder , but it delivers a higher number of tokens per second.
Grok Code Fast 1 compared to GPT-5, Gemini 2.5 Pro, Qwen3-Coder and other models
Grok Code Fast 1 is reportedly particularly adept at TypeScript, Python, Java, Rust, C++, and Go. It can complete common programming tasks with minimal supervision, from building a project from scratch, providing insightful answers to codebase issues, to performing precise bug fixes.
For example, in less than a day, users can build the following combat simulator prototype using Grok Code Fast 1 on Cursor. Users can quickly iterate and precisely control it to achieve the desired effect.
For example, in less than a minute, users can complete the process from ideation to design model. The mobile responsive design is also excellent. The UI it generates looks good on both desktop and mobile devices without too much adjustment.
xAI built Grok Code Fast 1 from scratch, using a completely new model architecture. It constructed a pre-training corpus rich in programming-related content, and in the post-training phase, curated a high-quality dataset that reflects real-world pull requests and programming tasks. As a result, Grok Code Fast 1 mastered common tools like grep, the terminal, and file editing.
xAI takes a holistic approach to evaluating model performance, combining public benchmarks with real-world testing. Grok Code Fast 1 achieved a score of 70.8% on the full subset of SWE-Bench-Verified, using its own internal testing tools.
Grok Code Fast 1's overall ToyBench result currently ranks 5th, behind GPT-5, Claude Opus 4, Gemini 2.5 Pro, and DeepSeek Reasoner.
Currently, Grok Code Fast 1 is free for 7 days and includes access to GitHub Copilot, Cursor, Cline, Roo Code, Kilo Code, opencode and Windsurf.
This article comes from the WeChat public account "Zhidongxi" (ID: zhidxcom) , author: Li Shuiqing, editor: Xinyuan, and is authorized to be published by 36Kr.
