Grok 4.5 trained on tens of thousands GB300 GPUs, tops coding benchmarks

The AI arms race for coding heats up as SpaceXAI pushes the frontier with Grok 4.5: trained on a cluster of tens of thousands GPU accelerators and reinforced with multi‑step engineering tasks, this model promises to shift the benchmark for agentic reasoning and knowledge work.

SpaceXAI has launched Grok 4.5, a new AI model specifically engineered for coding, agentic tasks, and knowledge work. Developed in close collaboration with Cursor, the model arrives as the company positions itself to challenge incumbents in the AI-assisted development space with a sharp focus on speed, efficiency, and real-world engineering performance.

“

Grok 4.5 is immediately available to developers through Cursor on all plans and via SpaceXAI’s own Grok Build environment and API console.

The model’s training pipeline underscores its ambition. SpaceXAI deployed tens of thousands of NVIDIA GB300 GPUs, feeding them curated datasets spanning coding, science, engineering, and mathematics. The data preparation involved rigorous filtering, deduplication, quality scoring, and domain-specific data selection, combined with stability techniques designed to maintain reliability across large-scale runs. Perhaps most distinctive is the reinforcement learning regimen: hundreds of thousands of multi-step software engineering and technical tasks were used, with automated and model-based grading to refine Grok 4.5’s reasoning on extended, agentic workflows. The company’s asynchronous training infrastructure allowed training to continue seamlessly across tens of thousands of GPUs while simultaneously rolling out long-horizon agentic tasks, a setup that could provide a durable moat in model capability.

On the surface, Grok 4.5’s technical claims are aggressive. The model reportedly delivers inference speeds of up to 80 tokens per second (TPS) and achieves roughly 2x better token efficiency than comparable leading models, meaning it completes tasks using far fewer generated tokens. This combination directly translates into lower per‑task costs and faster user experiences. SpaceXAI states that Grok 4.5 outperforms peers on real-world engineering tasks—though independent benchmarks have yet to verify these assertions. If they hold up, the model could reset expectations for AI coding tools.

The partnership with Cursor is central to the go‑to‑market strategy. Grok 4.5 is immediately available to developers through Cursor on all plans and via SpaceXAI’s own Grok Build environment and API console. Notably, the company is offering free usage for a limited time, a move likely aimed at rapidly building a user base and generating feedback. However, the model is not yet available in the European Union across any of these channels, with availability expected in mid‑July—a delay that hints at regulatory or compliance hurdles, possibly related to the EU AI Act or data localization rules.

Contextually, the launch intensifies competition in the coding model arena, where players like GitHub Copilot (backed by OpenAI), Google’s Codey, and Anthropic’s Claude already vie for developer mindshare. SpaceXAI’s focus on agentic task reasoning goes beyond simple autocomplete, aiming to enable models to undertake multi‑step engineering projects autonomously. For enterprises and startups alike, cutting token cost by half while boosting throughput is a compelling value proposition, potentially accelerating development cycles and lowering the barrier to building complex software.