Grok 4.5 Released

• SpaceXAI released Grok 4.5 on 8 July 2026, calling it its 'smartest model built to excel at coding, agentic tasks, and knowledge work' and 'our strongest model ever', trained alongside Cursor; TechCrunch notes it was the first release since SpaceX's stock-market listing weeks earlier^[1][2][4]
• Cursor describes it as its first model trained with SpaceXAI: a new pretrained mixture-of-experts base, then trillions of tokens of Cursor data, with reinforcement learning on problems 'designed to be so hard that even frontier models fail at them'; Elon Musk puts the base at about 1.5 trillion parameters, roughly three times its predecessor, a figure the maker has not confirmed^[5][6]
• Musk called it 'an Opus-class model, but faster, more token-efficient and lower cost' and said SpaceXAI's internal view was that it is roughly comparable to Claude Opus 4.7 but much faster^[4]
• SpaceXAI's own charts put it near the top on engineering tasks: 64.7% on SWE Bench Pro (Claude Opus 4.8 69.2%, Fable max 80.4%, GPT 5.5 58.6%), 83.3% on Terminal Bench 2.1 (Fable 84.3%, GPT 5.5 83.4%), 53% on DeepSWE 1.1 (Fable 70%, GPT 5.5 67%) and 29.0% on SWE Marathon, the best listed^[1][3]
• Independent testing ranked it fourth overall: 54 on the Artificial Analysis Intelligence Index behind Claude Fable 5 (60), Claude Opus 4.8 (56) and GPT-5.5 (55), and 76 on the Coding Agent Index in Grok Build, level with GPT-5.5 in Codex and one point behind Fable 5 in Claude Code at roughly $2.50 per task against $5.07 and $11.80^[3][6]
• Its weak spot is truthfulness: on AA-Omniscience accuracy rose from 35% to 52% but the hallucination rate rose from 25% to 54%, and it is not yet offered in the EU, where the API console got it on 17 July^[2][3][6]

Notable features

• Priced at $2 per 1 million input tokens and $6 per 1 million output tokens, more than 60% below Opus 4.8 and GPT-5.5 list prices; requests over 200K input tokens cost double^[1][6]
• A 500K-token context window (down from Grok 4.3's 1M) with vision input and configurable reasoning of low, medium or high, high by default^[2][6]
• About 80 tokens per second, and SpaceXAI says it solves SWE Bench Pro tasks with 15,954 output tokens on average, about 4.2 times fewer than Opus 4.8 (max) at 67,020^[1]
• Trained across tens of thousands of NVIDIA GB300 GPUs, with reinforcement learning over hundreds of thousands of mostly multi-step software-engineering tasks in a highly asynchronous stack^[1]
• The default model in Grok Build, able to build multi-sheet Excel models and native PowerPoint and Word documents, with plugins for Word, PowerPoint, Excel and Outlook; free limited-time use in Grok Build and Cursor^[1]
• Available in Grok Build, in Cursor on all plans and from the API console as grok-4.5^[1][5]