Here’s a strange one for you. On June 28, Elon Musk announced a new model, gave it real specs, made a bold comparison to one of the top models on the market, and… basically nobody outside of two companies can touch it. Let’s go over what Grok 4.5 is, and why the claims around it deserve a bit of skepticism for now.

What Musk Said

The announcement came straight from a post on X. Musk said Grok 4.5 is built on xAI’s new V9 foundation model, which has around 1.5 trillion parameters. For context, that’s roughly three times the size of the v8-small model that currently handles regular Grok traffic on X, so this is a real jump in scale, not a small tweak.

On top of that base model, xAI added extra training using data from Cursor, the AI coding assistant that SpaceX agreed to acquire earlier this year. The idea is to make Grok noticeably better at coding and technical reasoning by learning from real developer workflows.

Musk’s exact words were that early evaluations show performance “close to, perhaps exceeding” Claude Opus. He also mentioned that reinforcement learning is ongoing through something called the Grok Build harness, and that both the model and the harness are improving daily.

Where Is It Running?

Right now, Grok 4.5 is only in private beta inside two companies: SpaceX and Tesla. There’s no public access, no API, and no announced release date. The model is being tested internally on real engineering work before it goes anywhere near a wider audience.

This approach makes sense from a testing standpoint. SpaceX and Tesla generate huge amounts of technical documentation, code, and engineering data, so it’s a demanding real-world environment to stress-test a model before a broader release.

The Part Worth Being Careful About

Here’s where it’s worth slowing down a bit. Every detail we have about Grok 4.5 comes from one source: Musk’s own post. There’s no independent benchmark, no system card, and no third party has run any tests on it. That “close to, perhaps exceeding Opus” line is xAI grading its own homework, and there’s precedent for that kind of self-reported claim not matching up with outside testing once a model launches.

It’s also worth pointing out that xAI hasn’t submitted its current public model, Grok 4.3, to major third-party benchmarks like LMSYS Arena or Artificial Analysis either. So there isn’t much of a track record here to compare against, which makes the Opus comparison even harder to evaluate on its own.

None of this means the claim is false. It just means there’s no way to check it yet, and the honest answer right now is that we don’t know how good Grok 4.5 is.

The Bigger Plan Behind This

The part of the announcement that might matter more than Grok 4.5 itself is the roadmap Musk mentioned alongside it. He said xAI plans to release entirely new foundation models, trained completely from scratch, every single month for the rest of 2026.

That’s an unusual claim. Training a foundation model from scratch isn’t like pushing a small update. It means running the entire pipeline again at full scale: pretraining, fine-tuning, reinforcement learning, all of it. Labs usually take months between major from-scratch releases, not weeks. If xAI can pull off a monthly cadence, that says a lot about the amount of compute running through their Colossus data centers.

Should You Care About Grok 4.5 Right Now?

Not that much yet, and that’s fine. There’s nothing to test, nothing to compare, and nothing you can build with today. What’s worth keeping an eye on is whether xAI delivers a public release in the coming weeks, and whether the “close to Opus” claim holds up once people outside of two companies get their hands on it.

Until we get more info about this model, this one goes in the “interesting announcement, wait and see” pile.