Grok 4.5 Enters Private Beta at SpaceX and Tesla, Skipping the Public API

Elon Musk's xAI placed its 1.5-trillion-parameter Grok 4.5 into internal-only testing at SpaceX and Tesla on June 28, promising monthly frontier-model releases through year-end while keeping the flagship out of the public API.

Grok 4.5 Enters Private Beta at SpaceX and Tesla, Skipping the Public API

xAI placed its newest large language model, Grok 4.5, into private beta at SpaceX and Tesla on June 28, 2026, according to a post by Elon Musk. Built on the company's 1.5-trillion-parameter V9 foundation model, it will not reach public developers on any published timeline — but xAI also disclosed a roadmap it says will produce entirely new foundation models, trained from scratch, every month through the end of 2026.

A V9 foundation three times the size of the production API

Grok 4.5 runs on the V9 architecture, which completed pre-training on May 26 at 1.5 trillion parameters — roughly triple the 500-billion-parameter v8-small model that currently serves the public xAI API. V9 was optimized for NVIDIA Blackwell GPUs on the Colossus cluster in Memphis, and Cursor's developer-workflow data was folded into supplemental post-pre-training rather than the base run. Musk's team acknowledged that supplemental inclusion is "not quite as good" as building the data into initial training — a limitation the successor 2-trillion-parameter model in the pipeline is designed to fix.

The Opus claim is self-reported

Musk said internal evals show performance "close to, perhaps exceeding" Anthropic's Claude Opus. That claim came with three caveats. The evaluations are run by SpaceX and Tesla engineers inside the same corporate parent that owns xAI. Grok 4.5 has not been submitted to Artificial Analysis, LMSYS Arena or SWE-bench Verified. And the "Opus" comparison does not specify a version — Anthropic's Claude Sonnet 5 and Opus 4.8 currently sit at the top of the Artificial Analysis composite where Grok models do not appear.

Grok logo

Why SpaceX and Tesla go first

Both companies now sit under the same SpaceX corporate umbrella that acquired Cursor for $60 billion and folded in xAI earlier this year. Routing Grok 4.5 through those workloads first gives xAI a training signal — aerospace documentation, vehicle-manufacturing pipelines, production code — that public benchmarks cannot replicate. It also keeps the top model out of competitors' hands while the paid xAI API continues to run on Grok 4.3, a model Musk himself has described as carrying "many fundamental flaws."

Monthly foundation models — if the compute holds

The most consequential piece of Musk's announcement was not Grok 4.5 itself, but the roadmap around it: xAI plans to ship entirely new foundation models, trained from scratch, every month through year-end. No frontier lab has publicly sustained that cadence. For developers, it collapses the evaluation window every 30 days — a target that argues for model-agnostic architectures over xAI-specific integrations, at least until Grok 4.5 shows up on an independent leaderboard.

Reporting based on coverage from Tech Times, TechCrunch and CBS News.

Category: Machine Learning

Tags: AI Models AI Startups AI Physical AI AI Infrastructure

Related Articles