xAI and Cursor Jointly Release Grok 4.5: Pricing and Training Data Become Key Variables

On July 9, 2026, xAI and Cursor released the Grok 4.5 model, claiming to reach the level of Opus 4.7 with a speed of 80 tokens/s. The pricing is $2 per million input tokens and $6 per million output tokens, with a cache cost of $0.50 and a context window of 500,000 tokens.

This release stems from a joint training initiative between the two parties. Cursor provided real interaction data from developers on its platform, including behavior records such as feature initiation, mid-process abandonment, comment writing, and context repair. xAI contributed computational infrastructure and training workflows, using this data for supplementary training rather than pre-training from scratch. As a result, the model was exposed to actual engineering workflows, not just finished code from public code repositories.

The difference in training data sources directly affects model behavior. Conventional coding models often scrape completed code from public repositories like GitHub for training, lacking the developer decision-making process. Grok 4.5, through Cursor data, obtains training signals for multi-file codebase navigation, intent understanding, and long-task consistency. xAI engineers noted that the supplementary training approach is not as thorough as embedding data in initial pre-training, and the next-generation model plans to incorporate Cursor data from the pre-training stage.

For developers, Grok 4.5 is directly integrated into the Cursor editor, supporting tab completion, inline editing, chat sidebar, and multi-file Composer mode. It is also available via the xAI API for external teams to call. The combination of pricing and speed makes it attractive for daily coding tasks.

The impact on the Cursor platform lies in the strengthening of the data flywheel. Users generate more interaction data with Grok 4.5, which can further optimize the model, forming a closed loop. Through the acquisition of Cursor, xAI gains control over the tooling layer, more tightly coupling model training with usage scenarios. Private testing has already been conducted within SpaceX and Tesla, demonstrating vertical integration extending from internal needs to external markets.

For xAI itself, Grok 4.5 is a coding-focused variant positioned between Grok 4 and subsequent general-purpose models. On June 28, Elon Musk announced it was based on the 1.5 trillion parameter V9 base model, with the actual release version being a mixture of experts system. The claimed performance approaching Opus 4.7 comes from evaluations by SpaceX and Tesla engineers, which are internal self-assessments, without independent benchmark results such as LMArena or SWE-bench.

For competitors such as Anthropic and OpenAI, this move constitutes direct pressure. xAI leverages Cursor data to fill gaps in the programming domain and expand into enterprise-level scenarios such as finance, law, and cybersecurity. This is the first joint outcome following the $60 billion acquisition deal, marking its push from general-purpose models into vertical domains.

For enterprise users, Grok 4.5 provides options for handling long-running tasks, with a 500,000 token context window supporting complex codebase operations. However, the model of relying on a single vendor controlling compute, model, and tools may increase switching costs.

Horizontal comparison shows that previous model releases mostly relied on public data or synthetic exercises, while Grok 4.5's supplementary training path is different. In historical cases, deep binding between tools and models often accelerates iteration, but may also limit the open ecosystem. If xAI's next-generation model incorporates data from the pre-training stage, the performance gap may narrow.