Particle.news

SpaceXAI Releases Grok 4.5, an Engineering‑Focused Large Language Model

Designed for multi-step coding and agent workflows, the release signals a push to cut per-task compute costs, raise token throughput, lower customer pricing

Overview

  • SpaceXAI released Grok 4.5 Thursday and made the model available on the SpaceXAI console, Grok Build and Cursor with introductory free usage.
  • The company says Grok 4.5 was jointly trained with Cursor on tens of thousands of NVIDIA GB300 GPUs and fine-tuned with reinforcement learning over hundreds of thousands of tasks to improve multi-step software work.
  • SpaceXAI reports the model uses roughly half the tokens per task of peer models and can output about 80 tokens per second, but those performance and efficiency claims are vendor-reported and lack independent benchmarks.
  • The company published aggressive pricing of $2 per million input tokens and $6 per million output tokens and said the model is not yet available in the EU, where access is expected around mid-July.
  • Grok 4.5 is positioned to challenge Anthropic and OpenAI on programming and agent-style workflows, and observers should watch for independent tests, the timing of EU rollout, and the still-unfinished $60 billion Cursor acquisition that underpins the collaboration.