On July 8, SpaceXAI (formerly xAI) unveiled Grok 4.5 — and the model matters to our readers for one clear reason: it takes direct aim at Claude.
Trained with Cursor
According to SpaceXAI, Grok 4.5 is the first model it trained specifically for coding and autonomous agents. The twist: it was trained jointly with Cursor — the AI coding startup SpaceX acquired just weeks ago for around $60 billion. Grok 4.5 is the first tangible result of that deal. It’s available now in Grok Build, in Cursor (all plans) and via the SpaceXAI console — though not in the EU until mid-July.
‘Opus-class,’ half the price
Elon Musk put it this way: ‘Our internal assessment is that Grok 4.5 is roughly comparable to Opus 4.7, but much faster.’ The real lever, though, is price. Grok 4.5 costs $2 per million input tokens and $6 per million output tokens. Claude Opus 4.7 runs $5 and $25 respectively. On output, that’s about a quarter of the price.
On the published benchmarks, Grok 4.5 plays in the leading pack without consistently dominating: 62.0% on DeepSWE 1.0, 83.3% on Terminal-Bench 2.1, 64.7% on SWE-Bench Pro. More interesting is efficiency: on SWE-Bench Pro, SpaceXAI says the model uses about 4.2x fewer output tokens than Claude Opus 4.8 in max mode. Fewer tokens at similar quality means a smaller bill.
My take
VentureBeat’s headline said the model could ‘rattle Anthropic and OpenAI’ — and I think there’s something to that. The competition is shifting from pure benchmark rankings to a simpler question: how much does one solved task cost me? If Grok 4.5 really delivers near-Opus quality at a fraction of the token cost, that becomes a genuine argument for high-volume teams.
Two things temper my enthusiasm. First: these are SpaceXAI’s own numbers, and independent tests are still pending. Second: tying your coding to a model from Musk’s universe means accepting platform lock-in — and the debate about tracking and data leakage in coding agents is only just heating up. But you can’t ignore Grok 4.5.
Sources: