Grok 4.5 Was Trained on Your Coding Sessions Before xAI Owned Them

SpaceXAI launched Grok 4.5 on July 8, and the headline claim is straightforward enough: Opus-class performance at a lower price, built for coding and agentic work. Elon Musk put it directly: "an Opus-class model, but faster, more token-efficient and lower cost." Fine. The interesting part is buried one level down.

Grok 4.5 was trained on trillions of tokens from Cursor developer sessions. Real sessions. The kind where you're debugging something at 11pm and trying three different approaches before one sticks. SpaceX agreed to acquire Cursor for $60 billion in June 2026. That deal has not closed yet. So the first model trained on Cursor's data shipped before the company is technically xAI's to own. SpaceXAI says that the legal and product relationship "will evolve after close," which is a careful way of phrasing something that probably involved some creative contract work.

I find this genuinely strange, and I think it matters more than the benchmark numbers. What Cursor had that xAI didn't was not just an IDE. It was a recording of how developers actually think through problems: the wrong turns, the refactors, the point where someone deletes a function and starts over. That's richer training signal than synthetic code tasks. Whether it's ethically tidy is a different question, and one I don't have the answer to. Users who worked in Cursor over the last year were not, as far as I can tell, told their sessions would train a model for a company that hadn't yet acquired their tool.

On the benchmarks: Grok 4.5 lands fourth on the Artificial Analysis Intelligence Index with a score of 54, ahead of every open-weight model and all Gemini models, but behind Fable 5 and Opus 4.8 in raw capability. xAI's own published chart confirms this, even as the press release framing works hard to obscure it. Musk said "Opus-class" in one post and then clarified "roughly comparable to Opus 4.7, but much faster" in a follow-up. That's not the same claim, and Opus 4.7 is not Opus 4.8.

The token-efficiency number is the one I keep coming back to. Grok 4.5 uses roughly 14,000 output tokens per Intelligence Index task. Opus 4.8 uses around 67,000. If that ratio holds in real engineering work, it changes the cost calculation completely. Priced at $2 per million input tokens and $6 per million output tokens, versus $5 and $25 for Opus 4.8, the per-task economics aren't close. A model that uses fewer tokens to get the same answer is, in practice, a cheaper model by a larger factor than the headline rates suggest.

Built on the 1.5-trillion-parameter V9 architecture, it also scores first on the Harvey Legal Agent Benchmark, which is an odd result for a model pitched primarily at engineers. Either the Cursor training mix generalizes further than expected, or Harvey's benchmark is easier to game than it looks. Probably worth finding out before drawing conclusions about legal use cases.

Musk also flagged that the current speeds are not the ceiling. xAI hasn't yet deployed its C/C++ inference stack optimized for GB300 hardware, and when it does, he expects latency to drop significantly. That's a promise about future infrastructure, not a current capability.

The model is live now in Cursor on all plans, in Grok Build, and through the API console. It is not yet available in the EU, with European access expected in mid-July.

The acquisition story and the training data story are the same story. xAI needed Cursor's signal badly enough to start using it before the deal closed. That's the read I keep landing on.