Grok 4.5 launches: Musk calls it an 'Opus-class' model
Elon Musk's xAI launched Grok 4.5 on 8 July and calls it an 'Opus-class' model. Independent tests place it among the best – but not quite at the very top.
On Tuesday 8 July, Elon Musk's AI company xAI launched a new version of its Grok language model. Grok 4.5 is the company's first model built specifically for programming and so-called agentic tasks, where the artificial intelligence works independently across many steps. Musk himself was not modest in his description: he called it an "Opus-class" model. But what does that actually mean – and does the claim hold up when independent tests weigh in?
For ordinary users this is more than just another technical update. It is about how quickly the most powerful AI models now swap places at the top, and about a price war that may decide who can afford to use them.
What's new in Grok 4.5
Grok 4.5 is sharpened for coding and agentic tasks. According to several technology sites, the model is built on a 1.5-trillion-parameter foundation referred to as "V9," and is said to have been trained partly on real working sessions from the coding tool Cursor. The model has a context window of around 500,000 tokens – simply put, how much text it can hold in mind at once – which lets it work through large codebases and long documents in one go. xAI made the model available to the public the day after launch.
From chatbot to "agent"
It is precisely the agentic tasks that make this generation of models interesting. Where an ordinary chatbot answers one question at a time, an agent model can break a larger task into many steps, use tools along the way and work through the problem almost on its own. In practice it might, for example, write a program, test it, spot its own mistakes and fix them over several rounds before delivering a finished answer. It is this ability to act, not just answer, that companies are now competing over – and it is exactly why Grok 4.5 is marketed first and foremost at programmers.
The model is made available through xAI's own channels and to developers via an application programming interface, so that other companies can build services on top of it. For Norwegian users and businesses, the razor-sharp race means, above all, more choice and lower prices. When the most powerful models cost fortunes, advanced artificial intelligence was in practice reserved for the very largest players. When a model that claims to belong in the top tier is offered at a fraction of the price, the threshold for adopting the technology drops – from customer service to software development – and more dare to experiment.
What does "Opus-class" mean?
The phrase refers to Claude Opus, the top model from the American company Anthropic, which has long been a yardstick for what the most expensive and most capable AI models can do. When Musk says "Opus-class," he is claiming that Grok 4.5 plays in the same division as the best its competitors have to offer.
"It is an Opus-class model, but faster, more token-efficient and lower cost," Musk wrote. In a follow-up he was more precise: "Grok 4.5 is roughly comparable to Opus 4.7, but much faster."
The nuance is worth noting. Being "roughly comparable to Opus 4.7" is not the same as beating the rival's very latest flagship. It is a confident but measured claim – and it comes from the seller himself.
What do the independent tests say?
Here the picture becomes more mixed. On the independent Artificial Analysis Intelligence Index, Grok 4.5 lands in fourth place with a score of 54 – ahead of every Gemini model and every open-weights model, but not at the very top. On some coding tests it performs strongly, with more than 83 percent on Terminal Bench 2.1 and around 65 percent on SWE Bench Pro. TechCrunch, which covered the launch, summed it up aptly: the results showed Grok was competitive with the best, but "just short of best-in-class."
The price war is the real story
Where Grok 4.5 stands out most clearly is on price. The model costs 2 dollars per million input tokens and 6 dollars per million output tokens – reportedly more than 60 percent cheaper than both Opus 4.8 and GPT-5.5. xAI also claims the model uses far fewer tokens to solve the same task. For companies running AI at scale, it is precisely the cost-versus-performance calculation that decides, not first place on a leaderboard. Cheap and "nearly best" can, in practice, beat expensive and best.
The race tightens
Grok 4.5 is a fresh reminder of how crowded the top has become. The most powerful models largely disappeared behind closed doors in 2026, and new releases now come so thick and fast that "best in the world" rarely lasts long. At the same time, each new generation demands enormous amounts of computing power – and therefore electricity – which makes the question of where the world's AI should be run from ever more pressing. The contrast with those who choose the opposite strategy, as Microsoft did when it open-sourced advanced speech AI, is becoming sharper too.
For the user the conclusion is simple: competition is fierce, prices are falling, and "class" claims should be read with a critical eye. The launch was covered by outlets including TechCrunch, and xAI's own presentation is available at x.ai.
