SpaceXAI (the xAI unit that used to just be called xAI before Musk merged the branding) opened public access to Grok 4.5 on Thursday morning. The timing was tight. OpenAI’s broad GPT-5.6 rollout, which had been the anchor tech event of the week, was scheduled for that afternoon. Grok 4.5 shipped roughly four and a half hours earlier. That is not a coincidence and Musk is not pretending it is.
The model runs on a new foundation, internally called V9, sized at roughly 1.5 trillion parameters. That is about three times the v8-small architecture that had been powering Grok 4.3. Supplemental training incorporated data from Cursor, which SpaceXAI acquired earlier this year, and Musk said future monthly Grok releases will include Cursor coding data “from the start of pre-training rather than supplemental fine-tuning.” So the coding pitch is going to keep getting louder.
The pitch itself is that this is an Opus-class model at lower cost. Musk’s exact line: “It is an Opus-class model, but faster, more token-efficient and lower cost.” And: “roughly comparable to Opus 4.7, but much faster.” Musk also said xAI is “closing the loop on real-world usefulness, not benchmarks,” which is convenient because there are no benchmarks. No system card. No third-party evaluation. The only outside data point at launch is that SpaceX and Tesla engineering teams have been running the model since June 28 in beta and gave good feedback, which is roughly the level of proof any founder gives about their own product to their own employees.
Availability at launch: Grok Build, Cursor across all plans, and the SpaceXAI console. Not available in the EU yet, with a mid-July window promised. Pricing was not published in the announcement. If you are trying to actually evaluate this against Opus 4.7 or GPT-5.6, you are going to have to spin up your own eval harness, which is exactly how the next two weeks of frontier-model coverage are going to go.