# SpaceXAI Launches New Flagship LLM Grok 4.5

- Category: Flash (快讯)
- Published: 2026-07-09T11:56:00.000Z
- Source date: 2026-07-09
- HTML: https://www.88z.ai/en/news/flash-20260709-bb0b4304e5e529b1
- Markdown: https://www.88z.ai/en/news/flash-20260709-bb0b4304e5e529b1.md
- Other language: https://www.88z.ai/news/flash-20260709-bb0b4304e5e529b1.md

SpaceXAI has released its flagship model, Grok 4.5, featuring a 1.5T parameter MoE architecture and joint training with Cursor. It rivals Claude Opus in performance, achieving 80 TPS inference speed and surpassing Opus 4.8 in programming on DeepSWE Bench, with over 4x better token efficiency. It currently supports a 500k context window, expanding to 1 million next week, with a 2T parameter version coming later this month.

## At a glance

- Grok 4.5 uses a 1.5T parameter MoE architecture, reaches 80 TPS inference speed, and surpasses Opus 4.8 in coding.
- Supports 500k context window expanding to 1 million soon, with a 2T parameter version arriving later this month.

## The update

SpaceXAI has officially launched its new flagship large language model, Grok 4.5. Built on a mixture-of-experts architecture with 1.5 trillion parameters, the model was optimized through joint training with Cursor. In terms of performance, Grok 4.5 directly rivals Claude Opus, achieving an inference speed of 80 tokens per second. Regarding coding capabilities, it outperforms Opus 4.8 on the DeepSWE Bench benchmark, while also delivering more than four times greater token efficiency. For context support, Grok 4.5 currently handles a 500,000-token context window, with plans to upgrade to one million tokens next week. Additionally, SpaceXAI announced that a larger version with 2 trillion parameters will be released later this month to further address complex task requirements.

## Sources

- [量子位](https://mp.weixin.qq.com/s/cSNPCSgkPhXjL7CG9ELYFQ) (2026-07-09)

This article is an AI-assisted synthesis by 88Z AI based on the public sources below. Product claims and availability are attributed to their publishers. It does not represent independent testing or first-hand reporting.