xAI dissolved into SpaceX to form SpaceXAI, then dropped Grok 4.5 — a 1.5 trillion parameter MoE model trained on real Cursor developer interaction data. It scores 83.3% on Terminal-Bench 2.1 at $2/$6 per million tokens, using roughly a quarter of the output tokens Opus 4.8 needs for equivalent coding tasks.
On July 8, 2026, xAI dissolved into SpaceX and became SpaceXAI. Two days later, the newly merged entity dropped Grok 4.5 — a 1.5 trillion parameter model trained on trillions of tokens of real Cursor agent-interaction data. The coding benchmarks are strong. The training story is stronger. And the whole thing raises a question nobody's asking out loud: what happens when AI companies merge with launch providers?
SpaceXAI is the result of folding xAI into SpaceX, a move Elon Musk announced with characteristic brevity — a few posts on X, no press release, no analyst call. The unified brand covers Grok, the xAI API, and the broader AI product suite. Grok 4.5 is the first release under the new structure, and it's positioned as an "Opus-class model" that's faster, cheaper, and more token-efficient than Anthropic's flagship.
The architecture is a 1.5 trillion parameter Mixture of Experts on a new V9 base. The headline differentiator is the training data. SpaceXAI disclosed that Grok 4.5 was trained on trillions of tokens of real Cursor agent-interaction data — actual developer sessions, actual coding workflows, actual debugging patterns. This isn't synthetic data generated by another model. This is behavioral data from human developers using an AI coding tool at scale.
The results are visible in the benchmarks. Grok 4.5 hits 83.3% on Terminal-Bench 2.1, a notoriously difficult shell-command and system-administration benchmark that tests real operational competence. SpaceXAI claims it uses about a quarter of the output tokens Opus 4.8 needs per solved SWE-Bench Pro task — meaning it gets to the answer with less back-and-forth, fewer corrections, more direct execution.
Pricing is aggressive. $2 per million input tokens, $6 per million output — roughly comparable to GPT-5.6 Terra and well below Opus 4.8. The model is available inside Cursor and through the SpaceXAI API, with Grok.com serving as the consumer interface. No open weights, no public training details beyond the broad strokes, and a self-disclosed data contamination issue that affected the CursorBench score.
That disclosure is worth noting. SpaceXAI admitted a Cursor codebase snapshot contaminated its training data and inflated the related benchmark. The company disclosed it voluntarily — which is more than most labs have done — but the incident raises the same question every training data disclosure raises: what else is in there that they haven't found yet?
The training-on-real-behavior approach is the most interesting piece of this release. Every frontier lab uses some form of RLHF or preference tuning. SpaceXAI's approach is different in scale and source — millions of developer sessions, not thousands of human raters. The model learned coding by watching how real developers actually interact with AI coding tools: what they accept, what they reject, what they modify, what they throw away. That's fundamentally different training signal from "human annotators rank these two completions."
The SpaceX merger angle is harder to evaluate. On paper, combining an AI lab with a launch provider makes limited business sense. In practice, Musk has been explicit about wanting to put AI data centers in orbit — reduced cooling costs, unlimited solar power, no terrestrial regulatory footprint. Whether that's a real near-term plan or a long-term narrative is unclear. What's clear is the merger gives SpaceXAI a capital structure and physical infrastructure position no other AI company can match.
Grok 4.5 is a credible coding model at aggressive pricing. The training methodology is genuinely novel. The corporate structure behind it is unprecedented. Whether those three things add up to a sustainable competitive position depends on execution — and on whether developers trust a model trained on their own behavior to serve their interests.
Get AI news in your inbox
Daily digest of what matters in AI.
Key Terms Explained #
Anthropic An AI safety company founded in 2021 by former OpenAI researchers, including Dario and Daniela Amodei.
Benchmark A standardized test used to measure and compare AI model performance.
GPT Generative Pre-trained Transformer.
Mixture of Experts An architecture where multiple specialized sub-networks (experts) share a model, but only a few activate for each input.