SpaceXAI released Grok 4.5, an Opus-level model trained specifically for coding and agents.

Built for real-world engineering. It handles large codebases, long-running tasks across multiple repositories, and works well with a variety of tools. Co-trained with Cursor. Uses half the tokens per task and runs faster than comparable models, at $2/M input and $6/M output.

Benchmarks

  • SWE Bench Pro: 64.7% (beats Opus 4.7's 64.3%)
  • Terminal Bench 2.1: 83.3%
  • DeepSWE 1.0: 62.0% (beats Opus 4.8)

Pricing: $2/M input, $6/M output. Half the tokens per task, higher tokens/sec compared to similar models.

Where to Use It

1. Cursor (all paid plans)

Select "Cursor Grok 4.5" from the model dropdown. Available on desktop, web, iOS, CLI, and SDK.

  • 50% off through July 14
  • First week: double usage

2. Grok Build CLI

One-line install:

curl -fsSL https://x.ai/cli/install.sh | bash

Requires SuperGrok plan. SpaceXAI reset all limits, so you can start free.

3. SuperGrok Free Trial (7 days)

Sign up here: https://grok.com/?referrer=website#subscribe

Personal Take

Tested it on frontend tasks. Feels stronger than Opus 4.8 there, faster, and the price gap shows.