No public access. No independent benchmarks. Musk’s word.
That’s the current state of Grok 4.5. The announcement came June 28, reportedly timed to Musk’s 55th birthday, with multiple sources confirming the private beta rollout at SpaceX and Tesla. Musk stated the model is built on a 1.5-trillion-parameter V9 foundation model, that figure traces to Musk’s own announcement and hasn’t been independently confirmed. The architecture is self-reported.
Supplemental training incorporated data from the Cursor IDE platform, according to Musk. Cursor, the AI coding tool, was acquired by SpaceX, the acquisition is covered separately in TJS’s markets-pillar coverage. What that training data actually optimizes in practice is editorial inference; Musk’s stated claim is that Cursor data was incorporated. Don’t expect a detailed technical paper to show up alongside this: no arXiv preprint has been identified for Grok 4.5.
The performance claim is the part nobody should run with yet. Musk claimed early evaluations show Grok 4.5 performing close to or potentially exceeding Anthropic’s Claude Opus. No independent benchmark results have been published. The sources corroborating this claim, Economic Times, Jang, are both paraphrasing Musk’s own statement, not citing a third-party evaluation. That’s a vendor claim traced to its original source, which is the vendor. Treat it accordingly.
Disputed Claim
Self-reported benchmarks. Read carefully.
The cadence announcement is the more structurally interesting claim. Musk stated that SpaceX plans to release a completely new trained-from-scratch model monthly through the remainder of 2026. That’s an aggressive schedule, training a frontier-scale model from scratch monthly would require sustained compute and engineering capacity that xAI has been building toward. It’s Musk’s stated plan, not a demonstrated track record at that frequency. Watch whether the next release arrives on schedule before treating the cadence as a given.
What’s not in this announcement: pricing, public API access, context window specifications, or inference cost data. Grok 4.5 is currently an internal tool at SpaceX and Tesla. For developers evaluating xAI’s competitive position, this is a signal, not a product they can use today. The existing public tier remains SuperGrok Heavy, Grok 4.5 isn’t in that stack yet.
The part nobody mentions: private beta testing at Musk’s own companies isn’t the same as independent evaluation at scale. SpaceX and Tesla are specific deployment environments with particular workloads. Performance in those contexts doesn’t generalize to arbitrary enterprise use cases until independent testing validates it.
What to Watch
xAI is betting on velocity, releasing a new foundation model monthly, iterating inside its own companies first, then presumably pushing to broader availability. Whether that cadence produces compounding capability gains or compounding technical debt will become clear over the next quarter. For now, watch whether Grok 4.5 exits private beta with third-party benchmark data attached.
If it does, that changes the conversation. If it doesn’t, the pattern of vendor-claimed performance without independent validation continues.