xAI has released Grok 4.5, a model trained on tens of thousands of Nvidia GB300 GPUs. The model is designed for coding, agentic tasks, and knowledge work. Benchmark results show mixed performance, with Grok 4.5 scoring 83.3% on Terminal Bench 2.1, nearly matching GPT 5.5's 83.4% and trailing Fable 5 by just one point. However, the model lags behind on other benchmarks, such as DeepSWE 1.1, where it scores 53% compared to GPT 5.5's 67% and Fable 5's 70%. On SWE Bench Pro, Grok 4.5 scores 64.7%, beating Opus 4.8 in some configurations but falling short of Fable 5's 80.4%. xAI claims it used heavy data filtering, deduplication, and domain-specific selection during training to maintain data quality. The reinforcement learning stage covered hundreds of thousands of tasks, mostly from software engineering, with automated scoring. xAI built the training infrastructure for asynchronous learning, allowing agentic runs to stretch over many hours while training continued in parallel.
Grok 4.5 costs $2 per million input tokens and $6 per million output tokens, significantly lower than Opus 4.8's $5 input and $25 output, Fable 5's $10 input and $50 output, and GPT-5.5 and GPT-5.6's $5 input and $30 output. xAI also states that Grok 4.5 uses 4.2 times fewer tokens than Opus 4.8 on SWE Bench Pro tasks and delivers results at 80 tokens per second. The pricing strategy mirrors that of Chinese vendors like Zhipu and DeepSeek, aiming to win on cost rather than performance. Grok 4.5 is available now through Grok Build, Cursor, and the xAI console, with plugins for Word, PowerPoint, and Excel. The model is not yet available in the EU, with xAI targeting a mid-July launch. xAI trained Grok 4.5 alongside the code editor Cursor, which SpaceX acquired in mid-June for $60 billion in stock.
Source: thedecoder
