To say Elon Musk's AI company has trained some of the most unhinged models on the internet would be an understatement. Grok’s sordid past includes cosplaying as “MechaHitler” and a foray into deepfake porn generation that briefly got the platform banned in some regions. As concerning as that might sound, the recently renamed Eloncorp known as SpaceXAI says Grok, now in version 4.5, has cleaned up its act, wiped its browser history, covered up the swastikas, and is ready to take on more serious endeavors such as tending to your legal quandaries, fiddling in Microsoft Excel, and generating code. “Today we’re launching Grok 4.5, SpaceXAI’s smartest model built to excel at coding, agentic tasks, and knowledge work,” the company wrote in a blog post. “The model is equally adept at office work, scoring number one on Harvey’s Legal Agent Benchmark.” If the company is to be believed, this incarnation of Grok is a whole lot less Van Wilder and more The Office. That is, the Microsoft Office. “Grok Build is capable of building complex Excel models that involve research from the web, multi-sheet formula use, and even leaves stickies or notes behind for future reference,” the company writes. Nothing like slipping in a passive aggressive sticky note to remind your boss you’re totally on board with the office AI mandate. And if Grok does go off the rails and starts fudging the numbers, perhaps it can help keep you out of jail when regulators come knocking – Harvey's Benchmark result aside, recall that Musk and Tesla ended up paying only $40 million to settle fraud charges with the SEC when Musk tweeted out "funding secured" over a supposed Tesla buyout offer that may never have existed. According to SpaceXAI, the model was trained on “tens of thousands" of Nvidia GB300 GPUs alongside Cursor, which it’s currently in the process of acquiring for $60 billion. A major emphasis with this training run was placed on quality rather than quantity. “Beyond raw token volume, we invested heavily in data filtering and curation: deduplication, quality scoring, and domain focused selection so that the data mixture stayed high-coverage and high-signal.” The model was further refined through reinforcement learning — the same technique originally used by OpenAI and DeepSeek to imbue their models with chain of thought “reasoning” capabilities — to teach the model hundreds of thousands of tasks. This has apparently helped cut down on the number of thinking tokens required to solve complex problems, which, along with faster serving speeds of up to 80 tokens a second, means higher-quality results with less delay. Or, at least that’s what SpaceXAI says. Independent benchmarks by Artificial Analysis show the model still isn’t as good as Anthropic’s Claude Fable, but roughly matches OpenAI’s GPT-5.5 and Claude Opus 4.8 and Sonnet 5. GPT-5.6 — which, much like Fable, set off alarm bells in Washington — is still in preview and hasn’t quite made it on the leaderboard just yet. Also, it's cheap. SpaceXAI is charging $2 per million input tokens and $6 for every million tokens generated. For comparison, GPT-5.5 will set you back $5/M input tokens, $0.50/M cached tokens, and $30/M output tokens. Grok 4.5 is available starting Wednesday in Grok Build, Cursor, and the SpaceXAI console to anyone who doesn’t call the European Union home. You fine folks will have to wait a little longer with rollout expected in mid-July. ®
