SpaceXAI, the artificial intelligence (AI) division of SpaceX led by Tesla CEO Elon Musk, has released its cutting-edge AI model "Grok 4.5."
SpaceXAI unveiled "Grok 4.5," which excels at coding, reasoning, and AI agent tasks, on the 8th (local time). The model was co-developed with the AI coding corporations Cursor.
SpaceXAI said on its blog that "Grok 4.5 is the highest-performing model we have released so far." Until now, Grok had been assessed as lagging behind frontier (cutting-edge) AI models such as Anthropic's Claude, OpenAI's GPT, and Google's Gemini, but SpaceXAI claimed the newly released "Grok 4.5" has improved to a level comparable to those models.
According to the performance benchmarks SpaceXAI released the same day, "Grok 4.5" scored 83.3% on Terminal-Bench 2.1, which measures coding ability in a terminal environment, higher than Claude Opus 4.8 (78.9%) and similar to GPT-5.5 (83.4%).
On SWE-Bench Pro and SWE-Bench Multilingual, which measure coding performance in languages such as Python, Java, and C, it scored lower than Opus 4.8 but higher than GPT-5.5.
SpaceXAI also emphasized that "Grok 4.5" has double the "token efficiency" compared with major AI models. The move appears to reflect a trend in which corporations, facing soaring token expense, are pursuing "model optimization" by combining inexpensive models with costly high-performance models.
SpaceXAI set the price for "Grok 4.5" at $2 per 1 million input tokens and $6 per 1 million output tokens. That is far cheaper than Anthropic's "Opus 4.7," which costs $5 per 1 million input tokens and $25 per 1 million output tokens. The industry says that if the model's actual performance matches SpaceXAI's description, it could secure significant price competitiveness.
Musk said on X (formerly Twitter), a social media platform, the same day, "Based on internal evaluations, Grok 4.5 matches Anthropic's Opus 4.7 in performance but is much faster," adding, "We are developing with a focus on real-world usefulness rather than benchmarks." He added, "It also looks possible to more than double the current speed."
"Grok 4.5" was trained using tens of thousands of Nvidia GB300 graphics processing units (GPUs). SpaceXAI said it did more than just increase the volume of training data, boosting quality through rigorous validation such as deduplication and quality assessment.