Cursor and SpaceXAI ship Grok 4.5, target legal and finance

The pricing is where I'd start. Cursor and SpaceXAI shipped Grok 4.5 at $2 per million input tokens and $6 per million output, with a faster variant at $4 and $18. Cursor pitches it as their most intelligent model and the first they've built for more than software engineering, targeting legal work, finance and general knowledge tasks alongside code.

Under the hood it is a mixture-of-experts model trained jointly with SpaceXAI, sitting in a different weight class from Composer 2.5 rather than replacing it. Independent write-ups from Axios and Forbes put the model in Opus-class territory on coding benchmarks, with a headline claim of roughly 2x the token efficiency of comparable leading models. On SWE-Bench Pro, xAI reports Grok 4.5 resolves tasks using an average of 15,954 output tokens against 67,020 for Opus 4.8 at max, a 4.2x gap.

That efficiency number is the one that would actually reshape procurement math, if it holds up in production. The sticker comparison is already flattering, with Anthropic's Opus 4.8 at $5 and $25 and OpenAI's GPT-5.6 Luna at $1 and $6. But once agents start running long tool loops, the line that dominates real bills is total output tokens per solved task, and the reported gap against Opus at max compounds fast.

The honest caveats are three. The benchmark story is mixed. Grok 4.5 reportedly beats Opus 4.8 on DeepSWE 1.0 and Terminal-Bench 2.1 but loses on DeepSWE 1.1 and SWE-Bench Pro, so treat Opus-class as a claim, not a settled result. The Cursor blog itself only shows charts, and a footnote flags that third-party benchmark scores are self-reported. And EU availability is reportedly delayed to mid-July, which is the wrinkle if the legal or finance team you're pitching is the one that sits inside the EU.

What the reporting doesn't give you is the strategic overhang. The reported SpaceX acquisition of Cursor at a $60bn valuation puts a lot of the agentic coding stack under one owner, and what that means for data-sharing terms with existing enterprise customers is not yet spelled out. For everyone else, the read is straightforward: a materially cheaper agentic model aimed squarely at the work Cursor used to route to hosted APIs from other labs. Watch how quickly the token-efficiency numbers get reproduced by independent evals.