SpaceXAI on Wednesday released Grok 4.5, a coding and agentic model priced at $2 per million input tokens and $6 per million output tokens, well below Anthropic’s Claude Opus 4.8 at $5 in and $25 out, and cheaper on output than OpenAI’s GPT-5.6 Luna at $1 in and $6 out. It’s the first model shipped since SpaceX absorbed xAI in February and, last month, announced the all-stock acquisition of Anysphere, the startup behind Cursor, at a $60 billion valuation.

The model was co-trained with Cursor. “We’ve partnered with SpaceXAI to train Grok 4.5,” the company said in a statement, a line that quietly confirms the strategic logic of the Anysphere deal: distribution-layer coding tools now feed directly into the frontier lab that owns them. SpaceXAI says Grok 4.5 was trained across tens of thousands of Nvidia GB300 GPUs, with what it calls meticulous data filtering, deduplication and quality scoring.

Published benchmarks put Grok 4.5 at 29.0 percent on SWE Marathon pass@1 against Opus 4.8’s 26.0 percent, and 83.3 percent on Terminal Bench 2.1. Elon Musk, on X, was more direct. “It is an Opus-class model, but faster, more token-efficient and lower cost,” he posted, while noting internally that he places it roughly at Opus 4.7.

The pricing is the point. Since May, when Musk said xAI would cease to exist as a separate company and operate as SpaceXAI, the entity has been telegraphing a compute-and-scale posture underwritten by its parent’s balance sheet. Grok Build and the SpaceXAI developer console open access now; EU availability is expected in mid-July.

Undercutting Anthropic by roughly three-quarters on output tokens isn’t a coding-tools skirmish. It’s a bid to make Opus-class inference the commodity tier.

Sources