On July 8, 2026, xAI (SpaceXAI) unveiled Grok 4.5 , a new model purpose-built for coding, agentic tasks, and knowledge work. It was trained on NVIDIA's rack-scale GB300 NVL72 system, using tens of thousands of GB300 GPUs.
July 8, 2026 · xAI
Grok 4.5 arrives — a coding model trained on NVIDIA's GB300 NVL72
Purpose-built for coding, agentic tasks, and knowledge work, and honed with Cursor on real engineering data. It's the first major model to cite the rack-scale GB300 NVL72 as its training environment — while claiming roughly 2× the token efficiency of leading rivals.
~2×
Token efficiency vs leaders
$2 / $6
Per M input / output tokens
Leaner output: 4.2× fewer tokens per task
Average output tokens on SWE-Bench Pro tasks — shorter is cheaper & faster.
Benchmarks: competitive with the field
Terminal-Bench 2.1 Grok 4.5 · 83.3%
GPT-5.5 xhigh 83.4% · Opus 4.8 max 78.9%
SWE-Bench Pro resolve Grok 4.5 · 64.7%
Opus 4.8 max 69.2%
DeepSWE 1.0 Grok 4.5 · 62.0%
The training rack · GB300 NVL72
Up to 50× the AI-factory output of the Hopper generation — in a single liquid-cooled rack.
72 + 36
Blackwell Ultra GPUs + Grace CPUs
−30%
Peak grid demand (power smoothing)
WHAT DEVELOPERS PRAISE
Practical for coding & agentic tasks
Strong speed-to-efficiency balance
Cursor integration & low pricing
One-prompt end-to-end app generation
WHAT'S STILL OPEN
Custom C/C++ GB300 stack not fully deployed
Further speedups expected with optimization
Independent large-scale evals still limited
Unavailable in the EU until mid-July
Continue reading The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.
Already purchased? Sign in ✓ Signed in — this article isn’t included in your current plan.Unlocking the full article…