Labs

SpaceXAI ships Grok
4.5 with Cursor,
priced under Anthropic
and OpenAI

The first joint model from Elon Musk's $60 billion Cursor acquisition lands at $2 per million input tokens, targets legal and finance work alongside coding, and skips the EU until mid-July.

By Casey Vanderhoof Frontier labs correspondent · July 8, 2026

SpaceXAI released Grok 4.5 on Wednesday at $2 per million input tokens and $6 per million output, undercutting Anthropic’s Claude Opus 4.8 ($5/$25) and both of OpenAI’s flagship tiers, Luna ($1/$6) and Sol ($5/$30), on the output side where enterprise bills actually accumulate. It’s the first jointly-trained model from Musk’s $60 billion all-stock acquisition of Anysphere last month, and the first release since SpaceXAI itself went public several weeks ago.

Bloomberg reports the model targets legal and finance workflows alongside coding, which is the more interesting tell. Grok 4.5 shipped inside Cursor on day one, trained on trillions of tokens of what Anysphere describes as “how developers work and how agents interact with their environments.” That’s a data moat framed as a product feature. It’s also why Anysphere disclosed, in the same post, that Grok 4.5 has “an advantage on CursorBench because an earlier snapshot of the Cursor codebase was accidentally included in training.” A rare instance of a lab pre-empting its own benchmark critique.

Musk, posting on X, called it “an Opus-class model, but faster, more token-efficient and lower cost,” and separately “roughly comparable to Opus 4.7.” SpaceXAI’s own benchmark page lists four evaluations and tells a mixed story: 62.0% on DeepSWE 1.0 (behind Fable’s 66.1% and GPT 5.5’s 64.31%, ahead of Opus 4.8’s 55.75%), 53% on DeepSWE 1.1 (last of four), 83.3% on Terminal Bench 2.1, and 64.7% on SWE Bench Pro. The lab claims roughly 2x token efficiency at comparable tasks and serves the model at 80 tokens per second, trained across tens of thousands of Nvidia GB300 GPUs. A faster Cursor variant runs $4/$18.

The pricing is the strategy. Undercut Anthropic on the output side by 76%, undercut Sol by 80%, and let the coding integration handle distribution. Cursor is doubling model usage on individual and team plans for a week to seed the numbers.

The timing isn’t subtle. Reuters reports OpenAI’s GPT-5.6, delayed last month after U.S. government requests citing national-security concerns about powerful AI, is expected to ship Thursday, one day after Grok 4.5. Grok Build and the SpaceXAI console are live now everywhere except the EU, which gets access in mid-July.

A safety-throttled competitor releasing the day after an unthrottled one, at triple the output price, is the kind of structural asymmetry that defines this phase. The pricing sheet is the policy debate, rendered in dollars per million tokens.

Sources