xAI’s Grok 4.5 entered private beta on June 28, 2026, deployed inside SpaceX and Tesla with no public API, no external access, and no independent benchmarks. Every performance claim traces back to a single source: Elon Musk, posting on X.

The model is built on V9, xAI’s ninth-generation foundation architecture, and weighs in at 1.5 trillion parameters. That’s a three-fold jump over the v8-small model currently serving Grok traffic on X and inside Tesla vehicles, and it sits above Grok 4.3, the version paying developers hit through the API today. Musk has described 4.3 as having “many fundamental flaws.” Pre-training on 4.5 wrapped May 26.

Musk’s pitch is that Grok 4.5 approaches or exceeds Anthropic’s Claude Opus. There’s currently no way to check this. The Artificial Analysis Intelligence Index, an independently run composite across nine evaluations, presently has Claude Fable 5 and Claude Opus 4.8 at the top. Grok 4.5 isn’t on it and can’t be, because nobody outside two Musk-owned companies can run it.

The gap between claim and evidence is the story. Frontier labs have historically staged capability reveals against public benchmarks, however gameable, because narrative velocity depends on legibility to outside evaluators. xAI is doing the opposite: internal deployment first, marketing claims second, verification indefinitely postponed.

The corporate architecture makes this workable. The SpaceX–xAI merger closed in February 2026 at a combined $1.25 trillion valuation. In June, SpaceX signed a $60 billion all-stock acquisition of Cursor, the developer-workflow platform with over a million active users, expected to close in Q3. Cursor’s data was added to 4.5 via supplemental training rather than from the start. “not quite as good as having it in initial training,” one xAI engineer conceded on X.

The fix is already running. A 2-trillion-parameter model incorporating Cursor data from initial pre-training is expected to complete in late July, with public launch targeted for August. xAI has committed to shipping entirely new foundation models, trained from scratch, every month through the end of 2026, running on the Colossus supercluster in Memphis, currently over 200,000 NVIDIA GPUs and pointed at a million.

Cadence is the product. Verification is optional.

Sources