en.Wedoany.com Reported - US AI company xAI's next-generation large language model, Grok 4.5, has entered internal private testing at SpaceX and Tesla. On June 28 local time, Elon Musk disclosed that Grok 4.5 is built on a V9 base model with 1.5 trillion parameters and incorporates Cursor-related data in supplementary training; early evaluations indicate that the model's performance is close to, and may even surpass, Anthropic's flagship model Claude Opus. Currently, Grok 4.5 is still being continuously optimized through reinforcement learning, and the accompanying Grok Build test benchmark is also being refined.
The uniqueness of this private test lies in its deployment directly within the high-complexity engineering environments of SpaceX and Tesla. SpaceX involves rockets, satellites, Starlink networks, manufacturing engineering, and mission management; Tesla involves automotive R&D, factory production, autonomous driving, energy systems, and robotics. Placing the new model in these real-world engineering settings means xAI must not only evaluate its performance in general Q&A, code generation, and reasoning benchmarks but also observe its ability to handle engineering documents, R&D tasks, automated workflows, and complex business collaboration.
Grok 4.5's use of a 1.5 trillion parameter V9 base model indicates that xAI continues to advance along the large-scale base model route. Parameter scale alone does not equate to final capability, but large base models provide higher capacity for reasoning, programming, knowledge integration, and multi-task generalization. Factors that truly impact product performance include training data quality, post-training strategies, reinforcement learning methods, tool invocation capabilities, context processing ability, and online inference system efficiency. Musk's emphasis that reinforcement learning is still significantly improving the model suggests that Grok 4.5 has not yet reached its final release state.
The inclusion of Cursor data in supplementary training is the most industry-relevant aspect of this news. Cursor is one of the more frequently used AI programming tools among developers, and related data may help the model better understand real development workflows, code contexts, debugging paths, and engineering collaboration methods. Competition among large models has shifted from "being able to write code" to "being able to participate in software engineering." A strong programming model needs to understand project structures, function dependencies, test feedback, error logs, and multi-turn modification intents. If Grok 4.5 undergoes supplementary training on such data, it may enhance its code generation and engineering task processing capabilities.
Benchmarking against Claude Opus also indicates that xAI is positioning Grok 4.5 in the frontier model competition. Claude Opus has long been regarded as one of the stronger models for high-end text reasoning, code analysis, and complex task processing. Musk's phrasing of "close to, and may even surpass" remains an early internal assessment and does not equate to confirmed superiority in public third-party benchmarks. For external developers and enterprise clients, Grok 4.5's true competitiveness will require more comprehensive public evaluations, API performance, long-context tasks, programming tasks, and multi-turn agent task results.
The refinement of the Grok Build test benchmark is also noteworthy. Frontier large models are no longer evaluated solely through traditional exam questions and single-turn Q&A; more model companies are building internal benchmarks for real-world tasks. If Grok Build targets software construction, product generation, engineering execution, or agent development scenarios, it could become an important tool for xAI to measure a model's practical utility. Whether a model can stably decompose steps, invoke tools, write code, detect errors, and continuously improve in complex tasks will determine its ability to enter enterprise production workflows.
Musk also revealed that SpaceX will release a completely new model trained from scratch every month for the remainder of this year. If this pace is realized, it would mean that xAI and Musk's engineering ecosystem are attempting higher-frequency base model iterations. Unlike post-training or minor version updates, training a new model from scratch requires substantial computing power, data, training engineering, and evaluation system support. Releasing a new model monthly is highly challenging and will test xAI's engineering capabilities in training clusters, data pipelines, model architectures, and release processes.
Grok 4.5's private testing within SpaceX and Tesla could also influence AI application methods within Musk's ecosystem. Tesla can test the model's capabilities in engineering design, manufacturing optimization, after-sales service, internal software development, and robotics R&D; SpaceX can use the model in mission documentation, satellite networks, engineering simulations, and complex workflow coordination. If private testing yields stable results, Grok 4.5 may later be more deeply embedded into the R&D and operational systems of Musk's enterprises, rather than serving solely as a chatbot for general users.
This also reflects that competition in frontier AI models is shifting toward "model capability + real-world scenarios + engineering closure." OpenAI, Anthropic, Google, Meta, and xAI are all vying for stronger models, but whoever can embed models into real organizations to drive productivity gains is more likely to achieve long-term commercial value. Grok 4.5's choice to first conduct private testing at SpaceX and Tesla essentially stress-tests the model within complex engineering enterprises to verify its ability to enter high-value production scenarios.
Key points to watch moving forward focus on three aspects: first, when Grok 4.5 will be opened to external users or developers; second, whether its public evaluations can support the early claim of being "close to or surpassing Opus"; and third, whether the internal private testing at SpaceX and Tesla can translate into reusable enterprise-level AI capabilities. As reinforcement learning and the Grok Build benchmark continue to advance, whether Grok 4.5 can evolve from an internal test model into a major competitor in the frontier AI market will be the most important observation point for xAI in the next phase.
This article is compiled by Wedoany. All AI citations must indicate the source as "Wedoany". If there is any infringement or other issues, please notify us promptly, and we will modify or delete it accordingly. Email: news@wedoany.com
