SpaceXAI, the company formerly known as xAI, released Grok 4.5 on 8 July 2026, describing it as its smartest model and its strongest yet for coding, agentic tasks and knowledge work. The model is available immediately in Grok Build, in Cursor on all plans, and through the SpaceXAI console — but not in the EU, where the company says none of its products or the API console yet offer it, with availability indicated for mid-July.

The launch landed in the same week as OpenAI's GPT-5.6 general availability, compressing two flagship releases into a single news cycle and sharpening what is increasingly a pricing contest rather than a pure capability contest.

The economic argument

Grok 4.5 is priced at US$2 per million input tokens and US$6 per million output tokens. For comparison, OpenAI's GPT-5.6 tiers launched a day later at US$5/US$30 for the flagship Sol, US$2.50/US$15 for Terra and US$1/US$6 for Luna — putting Grok 4.5's rates near the bottom of the frontier band while the company positions its capability against rival flagships.

The deeper argument SpaceXAI is making is about tokens per task, not price per token. The company claims roughly twice the token efficiency of comparable leading models, solving tasks in under half the number of steps — which, if it holds in production, compounds the headline discount. Those efficiency figures are the company's own launch claims, and buyers should weigh cost per successful outcome rather than cost per token; a cheaper model that needs more retries can cost more in practice.

On capability, the company's own eval tables show a mixed picture rather than a clean sweep: Grok 4.5 leads some long-session software-engineering evals — including a 29.0 per cent pass@1 on SWE Marathon, a benchmark of extended real-repository work — while trailing rival flagships on others, posting 64.7 per cent on SWE-Bench Pro and 83.3 per cent on Terminal-Bench 2.1. Artificial Analysis, an independent benchmarking firm, placed the model fourth on its Intelligence Index with a score of 54. Elon Musk has said current serving speed of around 80 tokens per second should rise materially once a custom inference stack ships, and has signalled a further model release next month — forward-looking statements that remain the company's own.

The Cursor pipeline

What genuinely distinguishes this release is where the training signal came from. Grok 4.5 is the first Grok model trained alongside Cursor, the AI coding tool whose maker, Anysphere, SpaceX agreed to acquire in an all-stock deal reported at roughly US$60 billion in June. Cursor has confirmed the training partnership. Cursor and SpaceXAI say Grok 4.5 was trained alongside Cursor. In a separate statement, Cursor said the model was trained jointly with SpaceXAI and used trillions of tokens of Cursor data, including user interactions with codebases and software tools. The companies have not publicly detailed how that interaction data is handled — for instance, what consent or opt-out applies — which is the part enterprise buyers with their own confidentiality obligations will want answered.

That gives SpaceXAI two assets rivals will find hard to replicate quickly: access to a large coding-focused developer ecosystem, and a distribution channel sitting directly inside the daily workflow of paying engineering teams using Cursor. The strategy has an obvious test attached — Cursor's own users can run head-to-head comparisons immediately, so the market's answer on whether the model holds up in messy production repositories will arrive quickly. Independent analysts strike a similar note: Forrester's Biswajeet Mahapatra told InfoWorld that the pricing lowers the economics of agentic coding workloads, but that enterprise buyers should judge on cost per successful outcome, since a cheaper model that needs repeated attempts can cost more in practice.

Key Takeaways

  • SpaceXAI released Grok 4.5 on 8 July 2026 at US$2/US$6 per million input/output tokens, available in Grok Build, Cursor and the SpaceXAI console.

  • Not yet available in the EU; the company indicates mid-July availability.

  • First Grok trained alongside Cursor, using real developer-session data; Cursor confirmed the partnership.

  • Company-reported evals show leadership on some long-session coding benchmarks and gaps against rival flagships on others; Artificial Analysis ranks it fourth on its Intelligence Index.

  • SpaceXAI says the model was trained across tens of thousands of Nvidia GB300 GPUs; a faster custom inference stack and a further release are signalled by Musk but not yet shipped.