Grok 4.5: SpaceXAI bets on coding and price with Cursor

SpaceXAI made Grok 4.5 official on July 8, opening it to the public the following day. It's a Mixture-of-Experts architecture built on roughly 1.5 trillion parameters, trained on tens of thousands of NVIDIA GB300 GPUs. Elon Musk described the system as "an Opus-class model, but faster, more token-efficient and lower cost." Reuters reports a 500,000-token context window, though that spec may still be revised.

The strategic selling point is its integration with the Cursor development environment, which confirmed the training partnership using real session data. That move made it possible to set the price at 2 dollars per million input tokens and 6 for output, landing roughly halfway to Anthropic's Opus 4.8 model. The aim is clear: break into the coding and agentic-task market by offering a cheaper alternative.

Still, the enthusiasm over the numbers needs tempering. The 83.3% score on Terminal-Bench 2.1 comes partly from the vendor and hasn't yet been blessed by third-party verification. On top of that, using real sessions for training raises questions about potential data contamination. For now the model stays out of the European market, with a debut expected by the end of July 2026.

Price efficiency is a market signal, but the architecture of trust rests on data transparency. When training depends on users' workflows, the line between improving the product and affecting privacy grows thin — and that calls for vigilance, not just benchmarks. — Olya