xAI has released Grok 4.5, a new AI model trained on tens of thousands of Nvidia GPUs that achieves mixed results on coding and engineering benchmarks—trailing competitors like Fable 5 and GPT-5.5 on some tests but matching them on others. The model's real advantage is price: at $2 per million input tokens and $6 per million output tokens, it undercuts rivals by a significant margin while also claiming to use far fewer tokens per task, potentially offering the lowest total cost for software engineering workloads if its efficiency gains hold in real use.
Summaries like this, in your inbox every morning.
Sign up free →What happened
xAI released Grok 4.5, trained on tens of thousands of Nvidia GB300 GPUs. The model scores 83.3% on Terminal Bench 2.1 (nearly matching GPT-5.5's 83.4% and Fable 5's 84.3%), but trails on DeepSWE 1.1 at 53% compared with GPT-5.5's 67% and Fable 5's 70%, and achieves 64.7% on SWE Bench Pro versus Fable 5's 80.4%.
Why it matters
Grok 4.5 costs $2 per million input tokens and $6 per million output tokens—far below Opus 4.8 ($5 input / $25 output), GPT-5.5 ($5 input / $30 output), and Fable 5 ($10 input / $50 output). xAI also claims it uses 4.2 times fewer tokens than Opus 4.8 on software engineering tasks, making it by far the cheapest option in this tier if efficiency claims hold up.
What to watch
Grok 4.5 is available now via Grok Build, Cursor, and the xAI console; plugins are live for Word, PowerPoint, and Excel. The model is not yet available in the EU, with xAI targeting a mid-July launch.
xAI's pricing strategy mirrors the approach that Chinese vendors like Zhipu and DeepSeek have used: achieve performance close enough to market leaders, then win on cost. Grok 4.5's benchmark results reveal a performance profile that is competitive on some tasks—particularly on Terminal Bench 2.1, where it nearly matches the leading models—but lags meaningfully on others, especially on DeepSWE 1.1, a test of real-world GitHub issue resolution. However, the combination of lower per-token pricing and the claimed 4.2× token efficiency gain over Opus 4.8 on software engineering tasks suggests xAI is betting that total cost of ownership, not raw benchmark performance, will drive adoption among businesses focused on engineering and coding workloads.
The company also reports heavy investment in training methodology, including data filtering, deduplication, domain-specific selection, and hundreds of thousands of reinforcement-learning tasks mostly from software engineering. xAI built asynchronous learning infrastructure to let agentic (self-directed AI) runs continue over many hours while training progressed in parallel, a technical approach aimed at improving both performance and efficiency. The timing of this release follows xAI's acquisition of Cursor, the code editor, in mid-June via SpaceX stock—a move that signals vertical integration of development tools alongside model capability.
No comments yet. Be the first to share your thoughts!
Log in to join the discussion
Get curated AI news from 200+ sources delivered daily to your inbox. Free to use.
Get Started FreeFree · takes 30 seconds · unsubscribe anytime
1 minute a day. The AI essentials.
200+ sources · Email / LINE / Slack
