SpaceXAI launches Grok 4.5, Musk compares it to Opus

SpaceXAI released Grok 4.5 on July 8. The company described the model as lower cost and faster than previous versions and positioned it for coders, engineers and what the company calls knowledge workers.

SpaceXAI set Grok 4.5 pricing at $2 per million input tokens and $6 per million output tokens. The model is available via API and on SpaceXAI’s Hermes and Grok Build platforms. SpaceXAI said access in the European Union is expected in mid-July. Grok 4.5 is the first public model the company has released since the SpaceX–xAI merger closed in February; SpaceX is also moving to complete an acquisition of Cursor.

SpaceXAI published four benchmark results at launch. On DeepSWE 1.1, which measures how reliably models can resolve real software bugs, Grok 4.5 scored 53 percent. SpaceXAI reported comparative scores of 59 percent for Opus 4.8, 67 percent for GPT 5.5 and 70 percent for Fable 5. On SWE Bench Pro, Grok 4.5 scored 64.7 percent, compared with 58.6 percent for GPT 5.5, 69.2 percent for Opus 4.8 and 80.4 percent for Fable 5. SpaceXAI noted the comparison uses GPT 5.5 because GPT 5.6 launched hours after Grok 4.5’s announcement.

The company highlighted efficiency and speed as key attributes. Grok 4.5 runs at about 80 tokens per second and supports roughly 500,000 tokens of context, which the company equated to a little under 400,000 words. On SWE Bench Pro tasks, Grok 4.5 used an average of 15,954 output tokens per job, while Opus 4.8 used 67,020 output tokens on the same tests.

SpaceXAI said Grok 4.5 was trained in collaboration with Cursor AI on tens of thousands of Nvidia GB300 GPUs inside Colossus, the company’s Memphis supercomputer infrastructure. SpaceXAI reported total Colossus capacity across more than 200,000 GPUs. Training data included developer session records from Cursor, such as debugging traces and code edits rather than only static code repositories.

Elon Musk posted on X that Grok 4.5 is “roughly comparable to Opus 4.7, but much faster.” Court filings referenced by the company show the founders have faced scrutiny over earlier training practices; SpaceXAI said the new training pipeline runs through a platform the company is acquiring.

SpaceXAI described Grok 4.5 as intended for teams that run large volumes of code generation and debugging tasks where lower token costs and faster responses affect operating expense. The company provided the published benchmarks, pricing and availability details at launch.