Cursor Unleashes Grok 4.5 AI Model

Cursor introduces Grok 4.5, an advanced AI model co-developed with SpaceXAI, designed for complex tasks across multiple domains beyond just software engineering.

4 min read
Illustration representing the Grok 4.5 AI model interface.
Cursor announces the release of its advanced Grok 4.5 AI model.· Cursor Blog
Visual TL;DR
Moves beyond Composer 2.5Context
previous model focused solely on software engineering tasks
From the articleThis new iteration moves beyond specialized coding tasks, aiming to handle complex, long-running problems across diverse fields like data science, finance, and legal work.
Cursor launches Grok 4.5Core
new advanced AI model co-developed with SpaceXAI
From the article 2 mentionsCursor has launched Grok 4.5, its latest AI model developed in partnership with SpaceXAI.
Broader training dataDriver
trillions of tokens from user interactions and STEM research
From the article 2 mentionsUnlike its predecessor, Composer 2.5, which focused on software engineering, Grok 4.5 features a broader training dataset.
Mixture-of-experts architectureContext
learns from existing software and developer-agent interactions
From the articleThe model's architecture is a mixture-of-experts (MoE), trained on a wide array of data.
Reinforcement learningContext
trained in challenging realistic environments for complex problems
From the articleGrok 4.5 was trained using reinforcement learning in challenging, realistic environments.
Handles complex problemsEffect
From the articleThis new iteration moves beyond specialized coding tasks, aiming to handle complex, long-running problems across diverse fields like data science, finance, and legal work.
New capabilities unlockedOutcome
tackles long-running, multi-domain challenges effectively
From the articleThis represents a significant advancement in AI model release capabilities.

Cursor has launched Grok 4.5, its latest AI model developed in partnership with SpaceXAI. This new iteration moves beyond specialized coding tasks, aiming to handle complex, long-running problems across diverse fields like data science, finance, and legal work.

Unlike its predecessor, Composer 2.5, which focused on software engineering, Grok 4.5 features a broader training dataset. This includes trillions of tokens from Cursor's extensive user interactions with codebases and software tools, combined with high-quality STEM tasks and research papers.

A Foundation Built on Diverse Data

The model's architecture is a mixture-of-experts (MoE), trained on a wide array of data. This approach allows it to learn from existing software and developer-agent interactions, capturing real-world workflows.

Tackling Complex Problems with Reinforcement Learning

Grok 4.5 was trained using reinforcement learning in challenging, realistic environments. These simulations pushed the model to investigate issues, utilize tools, recover from errors, and verify outcomes, often designed to challenge even state-of-the-art models.

A distributed agent system was employed to create these demanding environments at scale, accelerating the development cycle. This system allowed engineers to specify problems and verification methods, with agents constructing and refining the scenarios.

This represents a significant advancement in AI model release capabilities.

Availability and Pricing

Grok 4.5 is accessible today through Cursor's desktop, web, iOS, and CLI interfaces. Individual and team plans offer substantial usage, with double capacity for the first week.

The base model is priced at $2 per million input tokens and $6 per million output tokens. A faster variant costs $4 per million input tokens and $18 per million output tokens.

Cursor continues to offer Composer 2.5, with plans for future models in that class. The company acknowledges that Grok 4.5 had an advantage in internal CursorBench testing due to an accidental inclusion of its codebase in the training data, which has since been removed for future iterations.

© 2026 StartupHub.ai. All rights reserved. Do not enter, scrape, copy, reproduce, or republish this article in whole or in part. Use as input to AI training, fine-tuning, retrieval-augmented generation, or any machine-learning system is prohibited without written license. Substantially-similar derivative works will be pursued to the fullest extent of applicable copyright, database, and computer-misuse laws. See our terms.