[AINews] SpaceXAI launches Grok 4.5, first Opus-class model post Cursor acquisition
Latent Space ● Covered by 5 sources
xAI just dropped Grok 4.5, calling it an 'Opus-class' model built specifically for coding and AI agents. It's not the smartest model out there, but it's way cheaper and faster than GPT-5.5 or Opus 4.8.
Based on reporting by Latent Space — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Elon Musk has a habit of overselling things, so it's worth pausing on what actually shipped here. Grok 4.5 arrived today from xAI, built jointly with Cursor, and it's the company's first model explicitly designed for coding and agentic work rather than general chat. Musk pitched it ahead of launch as "Opus-class, but faster, more token-efficient and lower cost," and for once the framing lines up with the numbers that followed.
The model is a genuine jump in scale: 1.5 trillion parameters, three times the size of Grok 4.3. That's a real architectural leap, not a fine-tune with a new name slapped on it. Cursor, which co-trained the model, called it their most powerful yet and the first one built for more than plain software engineering, then rolled it into their product with double usage limits for the first week.
On pricing, xAI is clearly playing a different game than Anthropic or OpenAI. Grok 4.5 costs $2 per million input tokens and $6 per million output, compared to $5/$30 for GPT-5.6 and $5/$25 for Opus 4.8. Cache hits get a 75% discount, though anything past 200k tokens costs double, and the context window actually shrank from Grok 4.3's 1M down to 500k — something Musk says will likely get reversed within a week.
Artificial Analysis put actual numbers behind the vibes: Grok 4.5 lands 4th on their Intelligence Index at a score of 54, a 16-point jump over Grok 4.3, trailing only Fable 5, GPT-5.5 and Opus 4.8. But the efficiency story is where it separates itself — it burns roughly 60% fewer output tokens than Opus 4.8 per task and uses a fraction of the total tokens Fable 5 or GPT-5.5 need in coding-agent benchmarks. Cost per task on the Coding Agent Index comes to $2.59, which is the number xAI clearly wants people repeating.
And this lands against a backdrop where benchmarks like SWE-Bench Pro are already being called saturated by OpenAI's own evals team, meaning raw leaderboard position is losing meaning fast. Grok 4.5 isn't trying to win that game outright. It's betting that being nearly as good for a third of the price is the more interesting pitch when GPT-5.6 ships tomorrow and the frontier race becomes less about who's smartest and more about who's affordable enough to actually run in production.
My take — AI-written commentary, not fact-checked reporting
I'll say the unpopular thing: xAI picking efficiency over topping leaderboards is the correct move, because the benchmark race stopped meaning much the moment evaluators themselves started calling their own tests saturated. What I actually want disclosed and nobody's asking for is training data provenance —
Read more about this at: Latent Space
