On June 28, 2026, Elon Musk stated on X that Grok 4.5 had entered internal testing by SpaceX and Tesla. According to disclosures, this version is based on a 1.5 trillion-parameter V9 base model and includes Cursor-related data in supplementary training; early evaluations described it as close to, or possibly even surpassing, Opus. A more accurate assessment at this point is that Grok 4.5 has entered real-world enterprise validation but does not yet equate to official release or independent evaluation.
Private measurement locations are more noteworthy than the parameter numbers
SpaceX and Tesla each have engineering R&D, manufacturing, vehicle software, and extensive internal knowledge processes. If the model is tested at these two companies, it will not only assess chat performance but may also include code generation, long task execution, internal data retrieval, and tool calls. For xAI, such an environment can quickly expose issues in permissions, stability, and complex workflows, and also accumulate feedback for subsequent productization.
Adding Cursor data also points to programming ability, but "what data is added, what scope of authorization is, and what proportion of data is currently not publicly disclosed." Parameter quantities are only size information and cannot directly prove answer quality, reasoning efficiency, or usage cost.
"Approaching Opus" can only be considered a developer's judgment for now
Musk did not specify which version of Opus refers to here, nor did he disclose the review set, sample size, or test conditions. A model's superiority in internal tasks does not mean it is stronger in general Q&A, code repository modifications, or long-context tasks. Especially during the private testing phase, system prompts, toolchains, and inference budgets all significantly affect results.
Therefore, it is not appropriate to migrate models based on this at this stage. What enterprise developers really need to wait for is whether the API is open, context length and price, rate limits, tool call performance, and reproducible third-party reviews.
What signals does this move send?
Grok's competitive focus is shifting from single model releases to "model plus enterprise scenarios plus execution framework." Musk also mentioned the ongoing improvement of the Grok Build toolchain and said SpaceX plans to train new models monthly this year. High-frequency iterations can shorten feedback cycles but also bring issues of version stability and migration costs. If Grok 4.5 is officially launched, to determine whether it's worth using, first look at the actual task success rate, not just the number of parameters or the developer's horizontal evaluation.
Source of information
Elon Musk's original message posted on X; Investing.com report on June 28, 2026.
