What is Grok 4.5? xAI's new model, explained

If you've seen headlines saying Elon Musk's new Grok 4.5 might beat one of the best AI models around, here's the calmer version. Yes, xAI has a new model. No, you can't use it. And the 'might beat Claude' part is Musk's claim, not something anyone outside his companies has actually checked.

On 28 June 2026, Musk posted on X that Grok 4.5 had entered a private beta at two of his companies, SpaceX and Tesla. 'Private beta' means a small, invite-only test — in this case limited to engineers at those firms. It's the first time Grok 4.5 has been used anywhere that we know of, and it comes before any public launch.

Grok 4.5 runs on xAI's ninth-generation design, called V9, which finished its main training on 26 May 2026 with 1.5 trillion parameters. Parameters are roughly the number of internal dials a model can tune; more can mean more capacity, though not automatically better answers. That's about three times the size of the smaller Grok that currently replies on X and in Tesla cars. xAI also reportedly trained it on data from Cursor, a popular AI coding tool, which hints Grok 4.5 is aimed at writing and reasoning about code.

Here's the part worth remembering. Musk said early tests put Grok 4.5 close to, or above, Anthropic's Claude Opus — a top model. But xAI hasn't released any benchmark scores or a 'system card' (the document labs publish explaining how a model was tested and where it struggles). The only people who've tried it work for SpaceX and Tesla, which Musk controls. And none of the independent scoreboards the industry trusts — like LMArena or SWE-bench — has tested it. So treat 'beats Claude Opus' as a claim to check later, not a fact.

One more thing to expect: xAI says it plans to put out a brand-new model roughly every month for the rest of 2026, with these V9 versions (4.5, 4.6, 4.7 and so on) building towards a bigger model called Grok 5. So if you're waiting to actually try Grok 4.5, the more useful question may be which version reaches the public first — and whether it arrives with real numbers attached.

Let's put the 'is it actually good?' question in perspective, because it's the one that matters to you. Benchmarks exist precisely so buyers don't have to take a company's word for it — they're the independent report card. When a lab skips them, it's asking you to trust the marketing instead of the marks, and the sensible response is patience, not excitement. That's not cynicism about Grok specifically; it's the same standard you'd apply to any product that says 'trust us, it's the best' without showing the test results. The good news is that this usually resolves itself quickly. Once a model reaches the public or an outside evaluator, independent scores tend to appear within days, and the picture gets a lot clearer. So the practical advice is simple: if you're curious about Grok 4.5, bookmark it and check back when it's actually available and someone neutral has tested it. If it's as good as Musk says, the numbers will show it and you'll have lost nothing by waiting. And if the numbers never show up at all, that itself is your answer — genuinely great models tend to want to be measured.