On June 28, 2026, xAI, the company led by Elon Musk, rolled out its next-generation large language model, Grok 4.5, as a private beta limited to internal teams at SpaceX and Tesla. Built on a newly designed foundation model called V9, it scales up to 1.5 trillion parameters, and xAI says internal evaluations put it close to, or possibly ahead of, Anthropic's top-tier Opus model. For now, the company is field-testing it inside its own group before any wider release.
1.5 Trillion Parameters on the New V9 Foundation
The headline figure for Grok 4.5 is its 1.5 trillion parameters. The previous generation, Grok 4.4, arrived in late May 2026 with around 1 trillion parameters, so this marks roughly a 50 percent increase, achieved in only about a month.
The model is built on a new foundation that xAI calls V9. It is designed separately from the smaller model currently running on X, which suggests xAI plans to roll out a series of derivative models from this base. Stacking up generations this quickly is hard for rivals to ignore.
Cursor Data as the Key to Stronger Coding
Another distinctive feature of Grok 4.5 is that it folds usage data from the code editor Cursor into supplemental training. Cursor is a popular AI-assisted editor among developers, and its vast record of real work contains countless examples of actual coding flows, debugging steps, and design decisions. Training on this kind of data is meant to raise the model's accuracy in coding and technical reasoning.
On top of that, xAI has layered on continuous tuning through reinforcement learning (a method that improves answers through trial and error) and updates to what it calls the Grok Build harness.
Why SpaceX and Tesla Come First
There is a clear intent behind deploying the model to the floors of SpaceX and Tesla before any public release. Both companies operate at the intersection of hardware design, software, and real-time decision-making. Tasks such as rocket trajectory calculations and vehicle software development are far harder than public benchmarks, and they come up every day.
These environments give xAI a testing ground that competitors cannot easily replicate. While many AI companies measure performance on public benchmarks and simulated tasks, xAI has the advantage of honing its model on real operational data from within its own group.
The "Near-Opus" Claim Is Still Unverified
On performance, Musk says internal evaluations show the model is close to, and perhaps beyond, Opus. That claim, however, rests entirely on evaluations conducted inside SpaceX and Tesla, and as of early July no independent third-party benchmark results have been released. Some developers who tried an early build describe it as feeling similar to Opus, but that too remains anecdotal. Judging its true ability will require published evaluation metrics or a system card.
What Comes Next: Monthly Models and Grok 5
xAI has laid out an aggressive roadmap for the second half of 2026. Through the end of the year, the company reportedly plans to ship a newly trained foundation model, built from scratch, every month. Further out, Grok 5, said to target somewhere between 6 trillion and 10 trillion parameters, is reportedly already in training on a large compute base called Colossus 2.
Whether monthly iteration translates directly into compounding gains is still an open question. Even so, the sheer speed of this development pace looks certain to intensify competition among AI companies.
Summary
Grok 4.5 is xAI's latest model, scaled to 1.5 trillion parameters on the new V9 foundation and sharpened for coding with Cursor data. It is being field-tested inside SpaceX and Tesla, while its claimed near-Opus performance still awaits independent verification. With plans for monthly model releases and the ambitious Grok 5 on the horizon, it is a development worth watching as the second half of 2026's AI race takes shape.
