Grok 4.5 Is Here Coding, AI Agents, and Real Work in One Model

Last Updated: 17 July 2026 Grok 4.5 has arrived as a new model for coding, digital tasks, and research. It was developed with Cursor and is available through Grok Build, Cursor, and an API. The release is aimed at developers and teams that need an AI system capable of completing several connected steps. The model can write software, fix bugs, use tools, prepare office files, and work through longer assignments. Pricing starts at $2 per million input tokens and $6 per million output tokens. Cursor users can access it across desktop, web, iOS, the command line, and its software development kit. This connects with the recent open source launch of Grok Build, which expands how developers can use Grok 4.5 in real projects. By making the agent framework available, SpaceXAI allows teams to customize workflows, integrate tools, and build their own systems on top of the model. Grok 4.5 is a new AI model built for coding, computer based tasks, and knowledge work. It can write and repair software, search for information, use digital tools, and complete work that may require several steps. The model is available through Grok Build, Cursor, and the SpaceXAI developer console for developers worldwide. Although coding is a central use, Grok 4.5 also handles tasks in data science, finance, legal work, research, and common office software. It can build Excel models, prepare PowerPoint slides with native tools, and write documents in Word. These abilities place the model inside practical work flows beyond chat files. Agentic tasks are jobs where an AI model plans several actions, uses tools, checks its progress, and corrects errors before finishing. Grok 4.5 was trained for this type of work. That makes it suitable for assignments that need more than a single answer, such as examining code or preparing a detailed project. Grok 4.5 was trained jointly with Cursor, the coding platform used by individual developers and software teams. Its training included examples from software projects, developer interactions, science, mathematics, engineering, and research. This wider mix helps the model move between programming work and other computer tasks without being limited to one narrow specialty. The training process also used reinforcement learning, which gives a model repeated practice on difficult problems and rewards useful results. In these exercises, Grok 4.5 learned to investigate a task, select tools, recover from mistakes, and check its own output. The problems covered software engineering and wider forms of professional work. SpaceXAI used tens of thousands of Nvidia GB300 graphics processors during training. The data preparation included duplicate removal, quality scoring, and selection for specific subject areas. These steps were intended to reduce weak material and give the model clearer examples across coding, science, engineering, mathematics, and general knowledge tasks for complex work. Cursor also used groups of AI agents to create, test, and refine difficult training environments, giving Grok 4.5 more practice with tasks that require several connected actions. Published tests suggest the model performs well on real-world coding tasks, including fixing bugs, working in the terminal, and handling longer projects. Overall, it appears to be a capable option for developers who need help across different types of software work. These numbers suggest solid performance across code repair, terminal work, and longer engineering tasks. Benchmark scores still need context because each test measures a limited set of problems. Results may also change with different prompts, tools, settings, or evaluation methods, so practical testing remains useful before teams adopt the model widely. Cursor disclosed one important issue with its own CursorBench test. An earlier copy of the Cursor codebase had entered the training data, giving Grok 4.5 an unclear advantage on that test. Cursor excluded the result and removed the data for later models, which makes that score unsuitable for comparison overall. Published benchmark scores include: The standard Grok 4.5 API costs $2 for every million input tokens and $6 for every million output tokens. Input tokens cover the text, code, or data sent to the model. Output tokens cover the response it creates. This pricing applies through the SpaceXAI developer service and the base version in Cursor. The model is served at about 80 tokens per second. Its official results also report lower token use on SWE Bench Pro, with an average of 15,954 output tokens per task. Fewer generated tokens can reduce cost and waiting time when a model completes long coding or computer based assignments. Cursor also offers a faster version priced at $4 per million input tokens and $18 per million output tokens. The added cost is intended for users who value quicker responses. Teams should compare both options with their own tasks because speed, answer quality, and token use can vary across projects. Limited free access is available through Grok Build and Cursor, giving users a way to test the model before choosing paid API use for regular work. Grok 4.5 is the default model in Grok Build, the coding and computer task agent released as an open source project. People can use it for code work, research, spreadsheets, presentations, and documents. Grok Build can also use web information and leave notes inside files for later review by SpaceXAI. In Cursor, the model works across desktop, web, iOS, the command line interface, and the software development kit. Individual and team subscriptions include model usage, and Cursor doubled the included amount during the first week. This gives developers several ways to use the same model across different devices and work settings. Developers can also connect through the SpaceXAI console with an API key. That option supports adding Grok 4.5 to applications, internal tools, and automated processes. Grok Build requires less setup. The API gives developers more control over how requests are sent, how responses are used, and how the model fits existing systems. Grok 4.5 brings coding, tool use, research, and office tasks into one model. Its joint training with Cursor gives it a strong focus on the way developers and software agents work. Availability across Grok Build, Cursor, and the API also gives users several entry points based on their technical needs. The published benchmark results are promising, though they should be tested against real projects and read with the disclosed CursorBench issue in mind. Pricing is competitive, the standard model is fast, and limited free access gives developers a practical chance to measure its value before paying for regular use. Grok 4.5 is an AI model designed for coding, tool use, research, and other computer based work. It can write and repair software, examine information, use digital tools, and complete tasks that involve several connected steps. The model was trained with Cursor and is available through Grok Build, Cursor, and the SpaceXAI API. It also supports work in areas such as data science, finance, legal tasks, spreadsheets, presentations, and document writing, depending on the tool used to access it today. Yes. Grok 4.5 is available in Cursor on desktop, web, iOS, the command line interface, and the software development kit. Individual and team plans include access as part of Cursor's model pool. Cursor also doubled the included usage during the first week of release. A faster version is available at a higher token price. Access may depend on the user's plan, location, and current product limits, so the available amount can differ between accounts and work settings over time and by region. The standard Grok 4.5 model costs $2 per million input tokens and $6 per million output tokens. Input tokens are the text, code, or data sent to the model. Output tokens are the response it creates. Cursor also lists a faster version at $4 per million input tokens and $18 per million output tokens. Limited free use is available through Grok Build and Cursor, which allows developers to test common tasks before choosing regular paid API access for future projects. Grok 4.5 can support research, data science, finance, legal work, and common office tasks. In Grok Build, it can create Excel models, use web information, work with formulas, and leave notes inside files. It can also prepare PowerPoint slides with native presentation tools and write documents in Word. These functions depend on the software and permissions available around the model. Users should still review results, check important facts, and inspect files before using them in professional work or sharing them publicly. Grok 4.5 and Grok Build are different products. Grok 4.5 is the AI model that processes instructions and produces answers, code, or files. Grok Build is an agent and work tool that uses the model to carry out tasks on a computer. Grok 4.5 is now the default model inside Grok Build. Developers can also use the same model through Cursor or connect it to their own applications through the SpaceXAI API with an API key for custom business systems.