xAI's Grok 4.5 enters private beta at SpaceX and Tesla

Elon Musk said on X on 28 June 2026 that xAI's Grok 4.5 had entered private beta inside SpaceX and Tesla — the first confirmed deployment of the model within Musk's own companies, ahead of any wider release. It is a characteristically Musk launch: a flagship claim delivered by a single post, with the model itself placed only in the hands of engineers who work for the same parent umbrella. The substance underneath the announcement is thinner than the headline suggests.

Strip away the framing and a small number of concrete facts remain. Grok 4.5 is in a closed beta at two Musk portfolio companies; it is not on the public API, not on Grok's consumer surface on X, and not available to outside developers. Musk claimed that early internal evaluations put the model close to — and possibly above — Anthropic's Claude Opus. That comparison is the load-bearing marketing claim, and it is precisely the part xAI has not substantiated.

The model sits on xAI's ninth-generation foundation architecture, V9, which the company says finished pre-training on 26 May 2026 at 1.5 trillion parameters. That is roughly three times the size of the v8-small model that currently serves Grok's production traffic on X and inside Tesla vehicles. According to reporting, xAI folded data from Cursor, the AI coding environment, into supplemental training — a signal that Grok 4.5 is being tuned towards technical reasoning and code rather than general chat. Scale and a coding tilt are plausible ingredients for a stronger model; neither is evidence of one.

The gap in this story is not the model — it is the evidence. xAI has published no benchmarks and no system card. The only people who have evaluated Grok 4.5 are SpaceX and Tesla engineers, who work for companies Musk controls, so the evaluation is neither independent nor blind. None of the public leaderboards the industry actually watches — LMArena, Artificial Analysis, SWE-bench — has scored it. Until one does, 'close to Claude Opus' is an assertion, not a result.

This matters beyond xAI's own credibility. A model that has not published a system card has not publicly documented its safety testing, its known failure modes, or its evaluation methodology — the very disclosures rivals such as Anthropic and OpenAI now treat as table stakes for a frontier release. The absence turns a competitive claim into an act of faith in the person making it.

Grok 4.5 is also a waypoint. xAI has said it intends to ship a new from-scratch foundation model every month through the end of 2026, with monthly V9 variants — 4.5, 4.6, 4.7 and on — feeding the training run for a larger Grok 5 on the Colossus 2 supercluster. On that cadence, 4.5 is less a destination than the first public marker of an aggressive release treadmill, and the real test is whether any of those versions arrives with numbers attached.

One further point deserves emphasis for anyone tracking the frontier seriously. A monthly cadence of from-scratch foundation models, if xAI truly sustains it, is either a genuine training-infrastructure advantage or a quiet redefinition of what 'foundation model' means down to something lighter. Both readings matter, and only published evaluations can tell them apart. Until Grok 4.5 — or 4.6, or 4.7 — appears on a leaderboard someone outside Musk's companies controls, the right posture is interested but unconvinced, and the burden sits with xAI to convert a launch post into a measured result.