Grok 4.5 Launches in Cursor After Joint Training on User Data

Grok 4.5 reached general availability inside Cursor after Cursor and SpaceXAI trained the mixture-of-experts model on trillions of tokens drawn from Cursor user interactions.

The release covers software engineering, data science, finance and legal tasks through reinforced learning on domain-specific problems. Base pricing sits at two dollars per million input tokens and six dollars per million output tokens. The fast variant lists at four dollars per million input tokens and eighteen dollars per million output tokens.

Grok 4.5 now runs in Cursor desktop, web, iOS, CLI and SDK. Personal and team plan usage doubled during the first week. The model operates in a separate weight class from Composer 2.5, so both continue to receive support.

The training data came directly from Cursor sessions rather than generic web text. Engineers at both companies filtered the interaction logs, ran reinforcement learning loops on difficult user tasks and produced the final checkpoints now served through Cursor.

This approach gave the model concrete examples of multi-file edits, debugging sequences and data pipeline construction that typical pre-training rarely captures at scale.

Cursor and SpaceXAI chose this route because generic instruction tuning left gaps in real-world coding workflows. By starting from millions of actual sessions the teams shortened the path to reliable performance on the tasks Cursor users face daily.

The result is a model that already sits inside the same interface developers use for every edit and command.

Existing Cursor tools remain unchanged. Users switch between Grok 4.5 and Composer 2.5 without workflow changes because each checkpoint targets distinct strengths and memory footprints.

The pricing tiers apply immediately to all accounts. Free-tier users receive the same access limits they had before, while paid users see the new rates reflected in usage dashboards.

Early internal tests showed the largest gains on multi-step refactors and cross-language migrations. Teams that ran parallel evaluations noted shorter review cycles once the model handled repository-wide context.

The doubling of usage in the first week came mostly from existing subscribers rather than new sign-ups, suggesting the model filled an immediate need inside current customer bases.

No public claims have been made about benchmark superiority over other frontier models. The companies instead point to the volume of real session data and the reinforcement learning stage as the primary differentiators.

Independent verification of those claims remains limited because the training corpus stays private.

The next indicators to watch are weekly usage reports from Cursor, any adjustments to output caps on the fast tier, and the first third-party evaluations that compare Grok 4.5 against other coding models on held-out repositories.

Those three signals will show whether the training approach sustains adoption beyond the initial week or requires further iteration.