Grok 4.5 Enters Private Beta at SpaceX and Tesla: 1.5T Params, No Benchmarks, Monthly Releases
On June 28, 2026, Elon Musk confirmed that Grok 4.5 — built on xAI’s V9 foundation at 1.5 trillion parameters — entered private beta at SpaceX and Tesla. No external access was granted. No benchmark submissions have been made to Chatbot Arena, Artificial Analysis, SWE-bench Verified, or Terminal-Bench. The announcement came with a claim that internal evaluations show performance “close to, perhaps exceeding” Anthropic’s Claude Opus, which currently leads Chatbot Arena at 1512 ELO and scores 88.6% on SWE-bench Verified. Neither comparison can be independently checked.
The launch is three times the scale of the v8-small model currently powering public Grok traffic on X. That model — 500 billion parameters — has been in production while V9 training ran. Grok 4.4, released in late May, scaled to 1 trillion parameters. Grok 4.5 takes the V9 foundation to 1.5 trillion.
The Cursor Data Caveat
The 1.5 trillion parameter count is not the most consequential detail in Musk’s announcement. That distinction belongs to a technical disclosure buried in the post: Cursor developer-workflow data was added in supplemental post-pretraining rather than from the start of initial pretraining. An xAI engineer explicitly noted this is “not quite as good as having it in initial training.”
The 2-trillion-parameter run currently in progress — the model being positioned as the next Grok after 4.5 — is specifically designed to incorporate Cursor data from the beginning of pretraining. That means Grok 4.5, despite being the largest xAI model to date, carries a structural limitation in coding performance that its immediate successor is engineered to overcome. The model now in private beta may already be the bridge, not the destination.
The only available coding benchmark for Grok involves the earlier grok-code-fast-1 model on SWE-bench Verified: 70.8%, vendor-reported via xAI’s internal harness and not independently verified. For comparison, Claude Code running on Opus 4.7 posts 87.6% on the same test. No Terminal-Bench 2.1 submission exists for any Grok 4.x model.
Monthly From-Scratch Releases
The larger strategic announcement embedded in Musk’s post: xAI plans to release entirely new foundation models — not fine-tuned variants or updated checkpoints, but models trained completely from scratch — every month through the end of 2026, using SpaceX infrastructure.
If this cadence materialises, xAI will have shipped six independently trained foundation models by year-end. No competitor operates at that pace. The compute and infrastructure requirement is extreme: training a 1.5-trillion-parameter model from scratch is a months-long process at frontier scale. SpaceX’s Colossus cluster — 770,000 GPUs across the facility — is the resource base underwriting this claim.
Practical implication for developers on the xAI API: any production commitment to the current API model may need reassessment monthly. Each new model will arrive with a performance claim and, based on Grok 4.5’s launch pattern, potentially no accessible model for independent testing.
Model Progression
| Model | Params | Status | Notes |
|---|---|---|---|
| Grok v8-small | 500B | Live production | Current X/API default |
| Grok 4.4 | 1.0T | Released late May | Public |
| Grok 4.5 | 1.5T | Private beta | SpaceX, Tesla only |
| Grok 5 (next) | ~6T | In training | 2T intermediate also running |
xAI has not announced a public release date for Grok 4.5. A projection for the V9-Medium model entering public access was mid-June 2026 — that window has passed without a launch. No updated timeline has been provided. The model was originally behind schedule before this private beta deployment.