SpaceXAI and Cursor Launch Grok 4.5 – Not Yet in the EU
Elon Musk’s rebranded AI unit SpaceXAI released Grok 4.5 on Tuesday – the first model developed jointly with the coding startup Cursor. It is not yet available in the EU.
SpaceXAI, the AI division that emerged from the former xAI, unveiled its most capable model to date on 8 July 2026: Grok 4.5. It is the first AI model the company has trained and shipped together with the coding startup Cursor (Anysphere) – the firm SpaceX announced in mid-June it would acquire for 60 billion US dollars. According to the company, the model targets software engineering, agentic tasks and knowledge work, and explicitly extends to legal and finance-related use cases.
For European users, however, there is a decisive caveat: Grok 4.5 is initially not available in the EU – neither in SpaceXAI’s products nor via the API console. The company says EU availability is expected in mid-July, but has named no fixed date and given no reason for the delay.
Availability and Pricing
Outside the EU, Grok 4.5 is available immediately in Grok Build (where it is the default model), in Cursor across all plans, and via the SpaceXAI console. A broader public rollout is reported to be scheduled for Thursday.
On pricing, the model positions itself aggressively: the base model costs 2 US dollars per million input tokens and 6 US dollars per million output tokens; Cursor additionally lists a faster variant at 4 and 18 US dollars respectively. For comparison, several outlets point to Anthropic’s Opus 4.8 (5 and 25 US dollars) and OpenAI’s GPT-5.6 (1 and 6 US dollars). Cursor subscriptions for individuals and teams include the model in their own model pool and double the usage limit in the first week.
Performance: “Opus-Class” – According to the Vendor
Musk describes Grok 4.5 as an “Opus-class” model that is nonetheless faster, more token-efficient and cheaper. This framing is a vendor claim; independent confirmation of the benchmark figures was not available at launch.
The numbers SpaceXAI published itself paint a mixed picture: on DeepSWE 1.0 and Terminal-Bench 2.1, Grok 4.5 comes in ahead of Anthropic’s Opus 4.8; on DeepSWE 1.1 and SWE-Bench Pro, behind it. In several charts, Anthropic’s Fable 5 leads the field. On the legal-focused Harvey Legal Agent Benchmark, SpaceXAI claims the top spot. “Opus-class” as a category description is therefore plausible, but the company’s own data does not support consistently beating Opus 4.8.
The company highlights efficiency as a strength: on SWE-Bench Pro tasks, Grok 4.5 reportedly needs around 15,900 output tokens on average – roughly 4.2 times fewer than Opus 4.8. The model is served at about 80 tokens per second. Technically, according to Cursor, it is a Mixture-of-Experts model trained on trillions of tokens of real Cursor usage data as well as STEM tasks, research papers and further knowledge work; reports describe a 1.5-trillion-parameter “V9” foundation model.
The Difference to Composer: An In-House Coding Model on a Chinese Base
Alongside Grok 4.5, Cursor’s previous in-house model Composer 2.5 remains available. The two models belong to different size and weight classes, and Cursor says it will continue to release new models in the smaller Composer class in the future.
The key difference lies in purpose and origin. Composer is a deliberately lean, low-latency coding specialist for use directly inside Cursor. According to its technical report, Composer 2 was produced through continued pretraining on the open Kimi K2.5 base model from the Chinese provider Moonshot AI, supplemented by large-scale reinforcement learning in Cursor-like environments; Composer 2.5 built on that. Grok 4.5, by contrast, was trained jointly by SpaceXAI and Cursor as a broader, larger model, and covers general knowledge work in addition to coding.
Context: Consolidation in the AI Coding Market
The launch comes during a phase of rapid consolidation. SpaceX absorbed xAI in February 2026 and rebranded the AI unit as SpaceXAI; its IPO followed on 12 June, with the 60-billion-dollar Cursor acquisition announced a few days later. That deal has not yet formally closed – completion is expected in the third quarter of 2026. According to several reports, Grok 4.5 was trained partly on the same compute capacity that SpaceXAI simultaneously leases to competitors Anthropic and Google.
The timing is also notable: on the same day, OpenAI rolled out its new GPT-5.6 line more widely. Amid heightened regulatory attention to the cybersecurity risks of new models, Cursor additionally stresses that it has added further safeguards to detect and block misuse without impeding legitimate security research.
For the European market, it therefore remains open when – and on what terms – Grok 4.5 will become available, and whether the efficiency and price advantages claimed by the vendor will hold up in independent testing.
