SpaceXAI Launches Grok 4.5, a Cheaper Coding Model Built With Cursor

SpaceXAI released Grok 4.5, a coding and agentic model trained jointly with Cursor that undercuts rivals on price and token use, though it trails Fable 5 on benchmarks.

By Daniel Mercer Edited by Maria Konash Published: Jul 9, 2026 at 1:36 pm UTC

SpaceXAI launched Grok 4.5, a coding and agentic model trained with Cursor, priced at $2 per million input tokens. Image: SpaceXAI

SpaceXAI, the company formerly known as xAI and now part of Elon Musk’s SpaceX, launched Grok 4.5, calling it its strongest model yet and one built for coding, agentic tasks and knowledge work. Unusually, the same model shipped under two brands the same day: it was trained jointly with Cursor, the AI coding tool whose parent SpaceX acquired in June, and Cursor released it as its most capable model.

Grok 4.5 was trained across tens of thousands of Nvidia GB300 GPUs on SpaceXAI’s Colossus infrastructure, using trillions of tokens of Cursor data, including real developer interactions with codebases, plus broader science, engineering and math material.

The model’s real pitch is economics, not raw capability. Grok 4.5 is priced at $2 per million input tokens and $6 per million output tokens, well below comparable frontier models, and runs at about 80 tokens per second.

Its standout feature is token efficiency: SpaceXAI says it solves tasks in under half the steps of rivals, using roughly 4.2 times fewer output tokens than Anthropic’s Opus 4.8 on the SWE-Bench Pro test, which sharply lowers the real cost of a completed task.

Musk described it as an “Opus-class model, but faster, more token-efficient and lower cost,” while adding that his internal assessment puts it roughly on par with the older Opus 4.7. The Cursor tie-up gives SpaceXAI a valuable feedback loop from how programmers actually write and debug code, and a ready distribution channel inside a tool developers already use.

An important correction is warranted on the benchmark claims. Some coverage framed Grok 4.5 as beating Anthropic’s Claude Opus 4.8 across the board, but SpaceXAI’s own published charts tell a more modest story.

Anthropic’s Fable 5 leads all four coding benchmarks SpaceXAI cited, and Opus 4.8 beats Grok 4.5 on a couple of them; Grok 4.5’s best showing is a near-tie with Fable 5 on Terminal-Bench 2.1 at 83.3%. Independent tracker Artificial Analysis places Grok 4.5 behind Fable 5, GPT-5.5 and Opus 4.8 on overall intelligence. In other words, it is competitive and efficient rather than the outright leader, and all these figures are self-reported.

The Efficiency Play

Grok 4.5 reframes the AI model race around cost per useful outcome rather than peak benchmark scores. For high-volume, repetitive coding and agentic work, a model that is cheaper and burns far fewer tokens can win business even if it is not the smartest, especially as enterprises grow wary of depending on a single provider and start routing tasks by cost and complexity.

Analysts note the likely pattern: high-stakes work may still go to models like Claude, while cheaper models handle bulk workflows. The Cursor integration is the strategic core here, giving SpaceXAI a direct on-ramp to developers and validating SpaceX’s rationale for buying the startup.

The Caveats

Several practical limits temper the launch. Grok 4.5 is not available in the European Union at launch, with access expected only in mid-July, so “available today” is not global.

Analysts caution that token efficiency on public benchmarks may not translate to messy corporate codebases, and that a cheaper model can cost more in practice if it needs repeated attempts to produce working code; the real test is cost per successful task on a company’s own repositories.

There are also transparency gaps: SpaceXAI has not disclosed the model’s size, and Musk’s separate 1.5-trillion-parameter framing is unconfirmed, while Cursor flagged possible training contamination on one of its own benchmarks. The launch also lands in a crowded week alongside OpenAI’s GPT-5.6 and voice releases, and it raises a quiet irony, since SpaceXAI trained Grok 4.5 on the same Colossus compute it leases to rivals Anthropic and Google.