The acquisition of the developer tool Cursor by SpaceXAI (formerly xAI) and the associated close cooperation are bearing their first fruits: With Grok 4.5, a new flagship AI model has been introduced, which is expected to score particularly well in the areas of coding, agentic tasks, and knowledge work. At the same time, it is priced in the affordable range for flagship models. However, EU users will still have to be patient. Presumably due to legal requirements, the release for the EU region is postponed to mid-July, as SpaceXAI announces.
Grok 4.5 was trained on tens of thousands of Nvidia GB300 GPUs, according to the manufacturer, with a particular focus on data filtering, deduplication, and quality assessment. During training, it was primarily prepared for multi-stage software engineering tasks with automated and model-based evaluation. The model runs at around 80 tokens per second – comparable to “Flash” models from other providers – and is said to offer double the token efficiency for the same tasks compared to leading models.
Significantly cheaper than other top models
Grok 4.5 is not only said to be capable of handling difficult and long-running tasks, but also to be more broadly positioned, for example, for work in the legal or financial sectors. Furthermore, its capabilities in cybersecurity are said to have been further developed. The latter could accommodate the US government's desire, which recently slowed down the new top AI models from OpenAI and Anthropic due to security concerns.
